uva-dare (digital academic repository) commercial and ... · van dijk, d.w. link to publication...

UvA-DARE is a service provided by the library of the University of Amsterdam (https://dare.uva.nl)

UvA-DARE (Digital Academic Repository)

Commercial and residential real estate market liquidity

van Dijk, D.W.

Publication date2019Document VersionFinal published versionLicenseOther

Link to publication

Citation for published version (APA):van Dijk, D. W. (2019). Commercial and residential real estate market liquidity.

General rightsIt is not permitted to download or to forward/distribute the text or part of it without the consent of the author(s)and/or copyright holder(s), other than for strictly personal, individual use, unless the work is under an opencontent license (like Creative Commons).

Disclaimer/Complaints regulationsIf you believe that digital publication of certain material infringes any of your rights or (privacy) interests, pleaselet the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the materialinaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letterto: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. Youwill be contacted as soon as possible.

Download date:20 Aug 2021

https://dare.uva.nl/personal/pure/en/publications/commercial-and-residential-real-estate-market-liquidity(ff7c9fab-ca0d-42ea-94bc-feb49c36cc72).html

Comm

ercial and residential real estate market liquidity

Dorinth W

illem van D

ijk

Dorinth Willem van Dijk

Commercial and residential real estate market liquidity

Universiteit van Amsterdam

This dissertation explores several issues regarding the role of market liquidity in real estate assets, in particular for the Dutch residential and U.S. commercial real estate markets. Three main topics are addressed throughout all chapters. The first topic concerns the development of new ways to measure real estate market liquidity. The second topic addresses the link between changes in market liquidity and price developments. The third topic focusses on how market liquidity can be used by policymakers and market participants to get a better grasp on real estate markets.

DORINTH VAN DIJK received his master’s degree in Business Economics – Real Estate Finance & Finance from the University of Amsterdam in 2014 (cum laude). Dorinth won the “H.K. Nieuwenhuis Thesis Award” for his master’s thesis. He subsequently started his PhD in finance at the University of Amsterdam in collaboration with De Nederlandsche Bank. In 2015, he won the “Best Paper Award” for his research at the International Real Estate summit in Washington D.C. During his PhD, Dorinth also worked as an economist at De Nederlandsche Bank and was a visiting researcher at the MIT Center for Real Estate for which he received the “UvA385 Lustrumbeurs” travel grant.

ISBN 978-94-92679-73-4

Commercial and residential real estate market

liquidity

Universiteit van Amsterdam

Dorinth Willem van Dijk

COMMERCIAL AND RESIDENTIAL REAL ESTATE MARKETLIQUIDITY

ACADEMISCH PROEFSCHRIFTter verkrijging van de graad van doctor

aan de Universiteit van Amsterdamop gezag van de Rector Magnificus

prof. dr. ir. K.I.J. Maexten overstaan van een door het College voor Promoties ingestelde

commissie, in het openbaar te verdedigen in de Aula der Universiteitop vrijdag 1 februari 2019, te 13:00 uur

door Dorinth Willem van Dijkgeboren te Maastricht.

iii

Promotiecommissie:

Promotoren:prof. dr. M.K. Francke Universiteit van Amsterdamprof. dr. J. de Haan Rijksuniversiteit Groningen

Overige leden:prof. dr. J.M. Clapp University of Connecticutprof. dr. J.B.S. Conijn Universiteit van Amsterdamprof. dr. P.M.A. Eichholtz Universiteit van Maastrichtdr. E. Eiling Universiteit van Amsterdamprof. dr. D.M. Geltner Massachusetts Institute of Technologyprof. dr. P. van Gool Universiteit van Amsterdam

Faculteit Economie en BedrijfskundeUniversiteit van Amsterdam

iv

Contents

List of Tables ix

List of Figures xi

1 Introduction 1

1.1 Market liquidity in real estate markets . . . . . . . . . . . 2

1.2 How can real estate market liquidity be measured? . . . . 8

1.3 What is the relationship between prices and market liq-uidity in real estate markets? . . . . . . . . . . . . . . . . 9

1.4 How can market liquidity be used for a better understand-ing and monitoring of real estate markets? . . . . . . . . . 11

2 Regional constant-quality housing market liquidity in-dices 13

2.1 Introduction and motivation . . . . . . . . . . . . . . . . . 14

2.2 Model and data . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.2.2 Estimation . . . . . . . . . . . . . . . . . . . . . . 23

2.2.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3 Market liquidity and risk . . . . . . . . . . . . . . . . . . 28

2.3.1 Liquidity indices . . . . . . . . . . . . . . . . . . . 28

2.3.2 Commonality with transaction prices . . . . . . . . 31

2.3.3 Liquidity risk . . . . . . . . . . . . . . . . . . . . . 34

2.4 Determinants, decomposition, and robustness . . . . . . . 35

2.4.1 Determinants of the time on market . . . . . . . . 35

2.4.2 Time-varying effect of list price premium . . . . . 41

v

CONTENTS

2.4.3 Decomposition of effects: the effect of quality andwithdrawals . . . . . . . . . . . . . . . . . . . . . . 43

2.4.4 Decomposition of effects: the effect of list pricepremium . . . . . . . . . . . . . . . . . . . . . . . 47

2.4.5 Revisions . . . . . . . . . . . . . . . . . . . . . . . 50

2.4.6 Correlated unobserved heterogeneity . . . . . . . . 53

2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 55

3 Revisiting supply and demand indices in real estate 57

3.1 Introduction and Motivation . . . . . . . . . . . . . . . . . 58

3.2 Reservation Prices and Liquidity . . . . . . . . . . . . . . 62

3.2.1 Model . . . . . . . . . . . . . . . . . . . . . . . . . 62

3.2.2 Estimation . . . . . . . . . . . . . . . . . . . . . . 69

3.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

3.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.4.1 Probability of sale . . . . . . . . . . . . . . . . . . 73

3.4.2 Repeat sales indices . . . . . . . . . . . . . . . . . 75

3.4.3 Supply and demand indices . . . . . . . . . . . . . 77

3.4.4 A recent anomaly? . . . . . . . . . . . . . . . . . . 79

3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 80

4 Internet search behavior, liquidity, and prices in thehousing market 83

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . 87

4.2.1 Market liquidity . . . . . . . . . . . . . . . . . . . 87

4.2.2 Search and matching models . . . . . . . . . . . . 87

4.2.3 Downpayment constraints . . . . . . . . . . . . . . 88

4.2.4 Behavioral explanations . . . . . . . . . . . . . . . 88

4.3 Theoretical framework . . . . . . . . . . . . . . . . . . . . 89

4.3.1 Market tightness and prices . . . . . . . . . . . . . 89

4.3.2 Market tightness and liquidity . . . . . . . . . . . 90

4.3.3 Temporal differences . . . . . . . . . . . . . . . . . 90

4.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4.4.1 Transaction data . . . . . . . . . . . . . . . . . . . 92

4.4.2 Internet search data . . . . . . . . . . . . . . . . . 96

4.5 Empirical model . . . . . . . . . . . . . . . . . . . . . . . 98

4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

4.6.1 Estimation results . . . . . . . . . . . . . . . . . . 99

vi

CONTENTS

4.6.2 Impulse responses . . . . . . . . . . . . . . . . . . 1014.7 Robustness checks . . . . . . . . . . . . . . . . . . . . . . 104

4.7.1 Earthquake area . . . . . . . . . . . . . . . . . . . 1044.7.2 Geographical variation in responsiveness . . . . . . 106

4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 110

5 Conclusion and summary “Commercial and residentialreal estate market liquidity” 1135.1 How can real estate market liquidity be measured? . . . . 1145.2 What is the relationship between prices and market liq-

uidity in real estate markets? . . . . . . . . . . . . . . . . 1155.3 How can market liquidity be used for a better understand-

ing and monitoring of real estate market? . . . . . . . . . 117

Appendices 121

A Appendix for Chapter 2 123A.1 Comparison with Carillo and Pope (2012) . . . . . . . . . 123A.2 Estimated transaction price at time of entry . . . . . . . . 127A.3 Transaction prices indices . . . . . . . . . . . . . . . . . . 128

B Appendix for Chapter 4 131B.1 Theoretical framework . . . . . . . . . . . . . . . . . . . . 131B.2 House price index and rate of sale estimation . . . . . . . 134

Bibliography 141

Nederlandse samenvatting (Summary in Dutch) 149

List of Co-Authors 155

Acknowledgments 157

vii

CONTENTS

viii

List of Tables

2.1 Description of the three markets between 2005 and 2016. . 27

2.2 Estimates of the control variables for three different mar-kets in two different models. . . . . . . . . . . . . . . . . . 37

2.3 Summary statistics of revisions. . . . . . . . . . . . . . . . 52

2.4 Posterior means and 95% HPD intervals for the factorloadings and standard deviations of the unobserved het-erogeneity components. . . . . . . . . . . . . . . . . . . . 53

3.1 Descriptive statistics of the two markets over the sample. 72

3.2 Coefficient estimates including 95% confidence intervalsof the probit equation for the two markets over 2005Q1and 2017Q2. . . . . . . . . . . . . . . . . . . . . . . . . . . 74

3.3 Coefficient estimates on the inverse Mills ratio (σε,η) inthe repeat sales equations including the 95% HPD inter-vals and model statistics. . . . . . . . . . . . . . . . . . . 77

4.1 Overview of variables. . . . . . . . . . . . . . . . . . . . . 92

4.2 Descriptive statistics 2011 - 2013 per year. . . . . . . . . . 94

4.3 Results of the combining p-value tests to test for a unitroot in the specified variables. . . . . . . . . . . . . . . . . 96

4.4 Panel VAR regression results. . . . . . . . . . . . . . . . . 100

4.5 Panel VAR regression without earthquake areas. . . . . . 105

4.6 Robustness check: Panel VAR regression results of urbanareas and rural areas. . . . . . . . . . . . . . . . . . . . . 107

ix

LIST OF TABLES

4.7 Robustness check: Panel VAR regression results with thesample split in two according to the average annualizedreturns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

A.1 Hedonic estimation of the coefficients on log transactionprice. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

A.2 Hedonic (HTM) estimation of the coefficients on log trans-action price. . . . . . . . . . . . . . . . . . . . . . . . . . . 130

B.1 Summary statistics of HTM estimations of the price index.137B.2 Estimates of the coefficients of house characteristics on

the log of transaction price in COROP region 23 (Ams-terdam region). . . . . . . . . . . . . . . . . . . . . . . . . 138

x

List of Figures

1.1 Buyers’ and sellers’ reservation price distributions at dif-ferent points in the cycle consistent with pro-cyclical liq-uidity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 Cumulative distributions of buyers’ and sellers’ reserva-tion prices. . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Pro-cyclical liquidity in a classical supply and demanddiagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.4 Pro-cyclical liquidity in a Robinson Crusoe economy. . . . 7

2.1 Fraction of houses withdrawn over the cycle in the Ams-terdam region, 2005-2016. . . . . . . . . . . . . . . . . . . 16

2.2 Two constant-quality liquidity indices and an index of themean TOM of sold properties, 2005-2016. . . . . . . . . . 30

2.3 Comparison of market liquidity (random walk indices) be-tween the three markets, 2005-2016. . . . . . . . . . . . . 31

2.4 Illiquidity (constant-quality and random walk, left axis)and transaction price (right axis) indices, 2005-2017. . . . 33

2.5 Comparison of market liquidity risk between the threemarkets, 2005-2016. . . . . . . . . . . . . . . . . . . . . . 34

2.6 Time-varying effect of list price premium and 95% HPDintervals (top panel) and the marginal effect (bottom panel)in the large market (Amsterdam), 2005-2016. . . . . . . . 42

2.7 Decomposition of effects. . . . . . . . . . . . . . . . . . . . 45

2.8 Difference in illiquidity due to quality. . . . . . . . . . . . 46

2.9 Illiquidity indices with and without list price premium,2005-2016. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

xi

LIST OF FIGURES

2.10 Difference in illiquidity due to list price premium, 2005-2016. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.11 Illiquidity indices with and without correlated unobservedheterogeneity, 2005-2016. . . . . . . . . . . . . . . . . . . . 54

3.1 Buyers’ and sellers’ reservation price distributions at dif-ferent points in the cycle consistent with pro-cyclical liq-uidity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

3.2 Raw and smoothed/seasonally adjusted estimates of thetime fixed effects in the probit equation (γt) for the twomarkets over 2005Q1 and 2017Q2. . . . . . . . . . . . . . 75

3.3 Repeat sales indices of commercial real estate in New Yorkand Phoenix between 2005Q1 and 2017Q2. . . . . . . . . 76

3.4 Supply and demand indices in the New York City area. . 783.5 Supply and demand indices in the Phoenix area. . . . . . 793.6 Buyers’and sellers’ cumulative distribution of reservation

prices in case of pro-cyclical liquidity. . . . . . . . . . . . . 80

4.1 Response of prices (B), matching probability (C) to achange in market tightness (A) in week 11. . . . . . . . . 91

4.2 Map that depicts the value of a standardized home withineach municipality between 2011 and 2013. . . . . . . . . . 95

4.3 Map that depicts the average number of clicks per houseof each municipality between 2011 and 2013. . . . . . . . 95

4.4 Impulse-response functions, impulse variable → responsevariable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

4.5 Cumulative impulse-response functions, impulse variable→ response variable. . . . . . . . . . . . . . . . . . . . . . 103

A.1 Comparison between constant-quality random walk in-dices and Carillo and Pope (2012). . . . . . . . . . . . . . 126

B.1 Price indices (A), rate of sale (B) and clicks per house(C) of six municipalities within COROP region 23 (Am-sterdam region). . . . . . . . . . . . . . . . . . . . . . . . 139

xii

Chapter 1Introduction

What have the Romans everdone for us?

Reg, Monty Python’s Life ofBrian

In Ancient Rome, wealthy Romans bought and rented out apartments–Insulae– as an investment. This makes real estate one of the oldestasset classes. Although much has changed in the functioning of real es-tate markets and financial markets in a broader sense, people still investin real estate. Professional investors trade office buildings, apartmentcomplexes or industrial buildings on a, compared to the Romans, rela-tively frequent basis. In addition, a lot of “ordinary” people buy andsell houses to live in, which makes real estate the most important assetclass for many households.

An important aspect for professional investors and households is howprices change over time. Positive returns can boost the real economy,but negative returns can have a serious dampening effect. Related tothese price changes is the liquidity of real estate markets. In the financialeconomics literature, market liquidity is usually defined as the ease atwhich an asset can be traded (Brunnermeier and Pedersen, 2009). Theimportance of a liquid market became apparent as early as the 17thcentury, when the formation of a liquid market to trade shares was cru-cial for the success of the first corporation, the Vereenigde OostindischeCompagnie or the Dutch East India Company (Dari-Mattiacci et al.,2017).

1

CHAPTER 1. INTRODUCTION

Real estate is an inherent illiquid asset class compared to, for ex-ample, stocks and bonds. Especially during the Global Financial Crisis(GFC) the importance of (the lack of) liquidity became clear. Prices felltremendously, but, maybe even more importantly, investors and house-holds were not able to sell their assets as quickly as desired. For in-vestors, lower liquidity of their investment portfolio means that it is moredifficult to rebalance their portfolio. For households, lower liquidity im-plies that they are not able to move if desired, which has consequencesfor labor mobility as well.

This thesis explores several aspects of market liquidity of real estateassets. The main question is “What is the role of market liquidity inreal estate markets?”. This question is divided into three sub-questions:(i) “How can real estate market liquidity be measured?”, (ii) “Whatis the relationship between prices and market liquidity in real estatemarkets?”, and (iii) “How can market liquidity be used for a betterunderstanding and monitoring of real estate markets?”.

This introduction will start with a view on the concept of real es-tate market liquidity and the relationship with prices based on basiceconomic theory. This will be followed by short discussion of the threesub-questions, how the chapters relate to these, and the main findings.

1.1 Market liquidity in real estate markets

Central to the illiquidity of real estate is that real estate markets canbe characterized as search markets, in which buyers and sellers need tobe matched. Both properties and agents are heterogeneous. The formerimplies that it is difficult to assess the true value of an asset and thelatter implies that agents can have different reservation prices for thesame property. A match –that is, a transaction– can only occur if thereservation price of the buyer is at least equal to the reservation priceof the seller. If this is not the case, the seller keeps the property on themarket and waits until another buyer arrives with a reservation pricethat is high enough. Additionally, the seller can also choose to lower thereservation price or choose to withdraw the property from the market.For now, assume that lowering reservation prices and withdrawals donot occur.

Generally, the frequency distribution of buyers and sellers will looklike the top panel of Figure 1.1. Buyers have, on average, a lower reser-

2


vation price than sellers, but there are some buyers with a higher reser-vation price than some sellers. If these are matched, a transaction willoccur. The shaded area depicts such situations and indicates the numberof transactions. The average transaction price is indicated by P0.

When the economic situation changes, buyers and sellers may ad-just their reservation prices. Obviously, this has consequences for theobserved average transaction price. It will also have an impact on thenumber of transactions if the reservation prices of buyers and sellersmove differently over the cycle. If they would move in parallel, pricesmay change but the number of transactions would be equal across time.Empirically, there is ample evidence that prices and transaction vol-ume are co-moving. In other words, the reservation prices of buyersand sellers have to move in a different manner over time. In a boom,both transaction volume and prices are high, while in a bust transac-tion volume and prices are low. The only way this can occur, is whenthe reservation prices of buyers and sellers in an up market are furtherapart and in a down market are closer together.1 Combined with thefact that reservation prices tend to move up (down) in a boom (bust),this implies that buyers change their reservation prices quicker than sell-ers. These situations are presented in the middle and bottom panels inFigure 1.1. There can be multiple reasons why buyers and sellers movetheir reservation prices differently, these will be discussed in Section 1.3.

It might be more informative to look at the cumulative distributionsof buyers’ and sellers’ reservation prices, which are shown in Figure 1.2.When looking at the cumulative distributions, the x-axis shows the equi-librium transaction price and the y-axis the number of transactions. Ifthe axes are flipped, a classical diagram with downward-sloping demandand upward-sloping supply curves is obtained (Figure 1.3).

In the initial situation, the equilibrium price and quantity are indi-cated by P1 and Q1. In case of a boom, the demand curve will shiftmore to the right than the supply curve, as buyers respond quicker thansellers. The result is that the new equilibrium price and quantity, P2

and Q2, will both be higher than the old price and quantity.Related to the number of transactions is the time on market (TOM).

The expected TOM of an asset is related to the number of buyers with

1Note that potential sellers can also choose to keep their properties of the market,which is equivalent to having very high reservation prices. This would also imply thatthe reservation prices of buyers and sellers are further apart, which results in fewertransactions.

3


Figure 1.1: Buyers’ and sellers’ reservation price distributions at different points inthe cycle consistent with pro-cyclical liquidity.

P0 Reservation Prices

Normalmark

et

Demand

Supply

P0P1 Reservation Prices

Boom

P0P1P2 Reservation Prices

Bust

4


Figure 1.2: Cumulative distributions of buyers’ and sellers’ reservation prices.

P1

Q1

Transactions

Reservation Prices

Buyers

Sellers

Figure 1.3: Pro-cyclical liquidity in a classical supply and demand diagram.

Reservation Prices

Transactions

S1

D1

S2

D2

P1

Q1

P2

Q2

5


an “appropriate” reservation price (i.e. buyers with a sufficiently highreservation price). In case the reservation prices of buyers in Figure1.1 moves more to the right, there are simply more buyers with anappropriate reservation price and the probability that a successful matchoccurs increases. Intuitively, when the proportion of suitable buyersincreases, the time to find a buyer and the TOM decreases. As such,there is an inverse relationship between the probability of sale and theTOM (Fisher et al., 2003).

A different perspective on the pro-cyclicality between liquidity –measured by the TOM– and prices can be obtained when looking atthe real estate market as a simplified “Robinson Crusoe” productioneconomy.2,3 Suppose that asset owners want to sell their assets. In do-ing so, they face a trade-off between the TOM and the price. Settinga lower (higher) reservation price will result in a lower (higher) trans-action price, but also in a lower (higher) TOM. Production curves intwo market situations (boom and bust) together with the indifferencecurves are presented in Figure 1.4. In a bust, the production functionis less efficient as sellers will hold properties off the market, which issimilar to keeping their reservation prices too high. In a boom, moreproperties are on the market and more transactions occur, resulting ina more efficient, steeper, production function and quicker price discov-ery.4 Because sellers prefer a higher price and a lower TOM, the sellers’isoquants are increasing towards the northwest. As such, for sellers theutility maximizing TOMs and prices are pro-cyclical.

2In the original example, Robinson Crusoe can choose to spend 24 hours per dayby either gathering coconuts or to enjoy leisure time.

3Note that pro-cyclicality in this case implies a high (low) TOM and a low (high)price.

4A potential analogue with the original Robinson Crusoe example would be atechnological shock in which it becomes easier to gather coconuts, e.g. the introduc-tion of a ladder.

6


Figure 1.4: Pro-cyclical liquidity in a Robinson Crusoe economy.

Price

T ime on Market

Bust

BoomU1

U2

TOM1TOM2

P1

P2

7


1.2 How can real estate market liquidity be mea-sured?

Ametefe et al. (2016) identify five dimensions of real estate market liq-uidity: tightness, depth, resilience, breadth, and immediacy. This thesiswill discuss most of these characteristics, how they differ over the cycle,and how these cyclical movements relate to price movements. Markettightness refers to the costs related to taking a “round-trip” (i.e. si-multaneously buy and sell or sell and buy). Market depth measuresthe extent to which trading can occur without affecting prices. After awhile, more trading will affect prices more, the magnitude by which thishappens is called resilience. The breadth refers to the overall size of alltrades. Finally, immediacy relates to the discount or premium relatedto selling or buying quickly.

Chapter 2 focuses on the TOM of the Dutch housing market. Thismeasure is mostly related to the immediacy characteristic of market liq-uidity. From an investors’ perspective, a lower expected TOM is relatedto more immediacy (lower costs of selling quickly). Practitioners andpolicymakers frequently use the average TOM of sold properties as amarket liquidity indicator. This chapter shows that the average TOMcan be misleading, mainly due to two reasons. Firstly, in calculatingthe average TOM, only properties that are sold are considered. A sellermight also choose to withdraw the property. If many sellers choose todo so, this is also an indication of an illiquid market. If, for example,the probability of a withdrawal increases during some periods, the av-erage TOM might give a wrong signal about liquidity. Secondly, housesare heterogeneous assets: no house is exactly the same. Some houses,usually more homogeneous properties like apartments, transact quicker.The main aim of this chapter is to construct a measure for market liq-uidity that corrects for these features. Novel features of the presentedmethod include that the liquidity indices can be created reliably up tothe end of the sample (until the most recent data comes in) and thatindices can be constructed in markets where transactions or withdrawalsoccur infrequently.

Chapter 3 develops a model of reservation prices of buyers andsellers in the US commercial real estate market to obtain a measurefor market tightness. In this model, reservation price dynamics are theroot of price and liquidity changes in the market. The model builds

8


on the empirical fact that liquidity and prices are highly pro-cyclical inreal estate. The model in this chapter is an extension of the model ofFisher et al. (2003). Their model is extended in a repeat-sales structuraltime series framework. This makes it possible to estimate reliable, robustinvestor supply and demand indices for granular markets. The differencebetween the central tendencies of these reservation prices can be usedas a measure for market tightness –the first aspect of market liquidity.This measure be viewed as an analogue to the bid-ask spread, which iscommonly used as a market liquidity measure in the stock market.

Chapter 4 employs a volume-based measure for market liquidityand relates it to the breadth of the Dutch housing market. More specif-ically, the rate of sale –i.e. the number of transactions divided by thenumber of houses for sale– is used as a measure for liquidity. The rateof sale can be viewed as an ex post sale probability and is very much re-lated to the TOM and the difference between average buyers’ and sellers’reservation prices. The chapter further proposes a leading measure formarket tightness based on internet search behavior. More specifically,the measure is the ratio of the number of clicks of all listed propertiesin this region divided by the number of listed properties. Conceptually,clicks are related to demand in the market and the number of listedproperties to supply. The ratio is then defined as market tightness (i.e.demand / supply). Intuitively, houses in more popular regions shouldreceive more clicks than houses in less popular regions and markets inthe more popular regions should be tighter.

1.3 What is the relationship between prices andmarket liquidity in real estate markets?

In the literature, there are roughly three theories that focus on the com-monality between real estate liquidity and transaction prices (De Witet al., 2013). The first theory is related to search models with asym-metric information, the second theory to anchoring and other behavioralexplanations, and the third theory relates liquidity to mortgage markets.The theories are not mutually exclusive and can be very much overlap-ping. All chapters in this thesis will discuss the commonality betweenprice and liquidity movements in real estate markets. Central to this isthe lead-lag relationship that tends to exist between these.

The first theory (asymmetric information models) focuses on the fact

9


that real estate assets are highly heterogeneous, for which the marketvalues are difficult to assess. As such, “news” regarding the asset valuesis not priced instantaneously. Due to the institutional structure of realestate markets, information regarding the supply of houses (e.g. list-ing prices, sold properties, and recent transaction values) is known tothe public. The number of buyers or other demand-side characteristics,such as income, are not known (instantaneously) to sellers. Because ofthis, sellers adapt more gradually, pro-cyclical liquidity occurs, and con-summated transaction prices will change gradually. The second theoryexplains the relationship with behavioral arguments. The most impor-tant argument is that sellers anchor their reservation prices to the pricethey paid when buying the property. When markets are going up, thisimplies that sellers keep their reservation prices too low. Conversely,when market are going down, the reservation prices are kept too high(i.e. sellers are loss-averse). The third theory relates to downpaymentand negative equity constraints. In an up-market, a homeowner is morelikely to have made a profit on the current home. Therefore, if thishomeowner would like to “move up the housing ladder”, he would haveenough money left for the downpayment for the next, more expensive,house. Therefore, current homeowners are more likely to move on toa more expensive home in a booming market. This initiates a “chain”of transactions and the result is that there will be more transactionsin booms. Conversely, in a bust, current homeowners will have less (oreven negative) equity to make a new downpayment on a more expensivehouse. In this case, the chain is less likely to be initiated, and fewertransactions will occur.

In Chapter 2 the empirical commonality between price changesand changes in the TOM is discussed. The results suggest a strongcommonality between transaction prices and the TOM: when prices arehigh / increasing, the TOM is low / decreasing and vice versa. Grangercausality tests show that changes in the TOM lead changes in transactionprices. Additionally, the relationship between list prices and the TOMis examined: setting a higher list price results in a higher TOM or slowersale.

Chapter 3 shows that time-varying liquidity has profound implica-tions for price measurements such as price indices that market partici-pants tend to look at. Chapter 3 constructs buyer and seller reservationprice indices. In tracking the relative changes of these, there is a par-ticularly interesting interpretation of the demand side index. Consider

10


the case where sellers match the buyers’ reservation prices at all times.Here the liquidity of the market would be constant over time. As such,the demand reservation price index has a constant-liquidity price indexinterpretation. Consequently, the difference between the demand-sideindex and the observed transaction price price index is the effect thatliquidity has on price measurements in real estate.

Chapter 4 provides insights in the relationship between homebuy-ers’ internet search behavior, housing prices, and liquidity. The theoret-ical and empirical model in this chapter allows for a three-way interac-tion between these variables. This simultaneous causality is importantto account for, since higher price levels might also affect liquidity. Fur-thermore, the two measures for liquidity (one based on transaction dataand one on internet search behavior) are, for obvious reasons, also re-lated. This chapter specifically looks at the intertemporal connectionsbetween these variables and regional differences.

1.4 How can market liquidity be used for a bet-ter understanding and monitoring of realestate markets?

Real estate markets are of vital importance for the economy, financialmarkets, and financial stability. Movements in real estate asset pricesare closely followed to ensure that financial institutions remain resilientto future shocks. Moreover, big swings in asset prices can pose substan-tial risks for households, which heavily invested –usually with debt– inthe real estate asset class. Commercial real estate is an important assetclass for public and non-public investment institutions such as pensionfunds, large real estate funds, and small private investors. Hence, under-standing the real estate market is important to assess the stability of thefinancial system and contributes to sustainable economic growth. Realestate liquidity plays an important role in the understanding and moni-toring of the real estate market. Therefore, it is essential to understandand track real estate liquidity. The three chapters in this thesis helpto better understand and monitor liquidity and real estate markets ingeneral. Additionally, the proposed indicators can be used as forecastinginstruments.

Chapter 2 provides a novel way to measure the TOM for regionalmarkets. A frequently used measure for liquidity used by policymakers

11


is the average (seller) TOM: how long has a house been on the marketbefore before it is sold? This chapter shows that it is important to correctthe TOM for features such as censoring and housing quality. To monitorthe market situation, it is of vital importance to correctly measure thesituation. The chapter shows that the usually employed average TOMunderestimates market liquidity in good times and overestimates marketliquidity in bad times. Furthermore, the constructed constant-qualityindices lead both price indices and the simple average TOM, and cantherefore be useful for both monitoring and forecasting purposes.

Chapter 3 shows that a more comprehensive view of the marketcan be obtained by tracking demand and supply reservation prices sep-arately. Particularly, the demand-side index provides a joint metric ofprice and liquidity changes and is therefore interesting for monitoringpurposes. Also, since liquidity is an important part of these demand(constant-liquidity) indices, they tend to lead observed transaction priceindices, which can make them useful for forecasting purposes as well.

Chapter 4 presents a model where liquidity and prices are linkedto internet search behavior. As shown in the empirical literature andchapters 2 and 3, liquidity itself is a leading indicator of prices. Thischapter shows how to “lead the leading indicator”. Since internet searchpopularity measures preparatory steps that potential home buyers takebefore the actual purchase, the indicator based on this behavior is shownto be useful in explaining changes in liquidity and prices. Furthermore,the chapter shows that it is likely that sellers respond more slowly tochanges in the market. This is an important feature of the housingmarket and real estate markets in general. Behavioral patterns like theseare essential for a thorough comprehension of the real estate market.

12

Chapter 2Regional constant-quality housingmarket liquidity indices

Rhymes grow and flow sosmooth like a fluid. Or liquid,with high liquidity.

Def Jef, Droppin’ Rhymes onDrums

The average time on market (TOM) of sold properties is frequently usedby practitioners and policymakers as a market liquidity indicator. Thisfigure might be misleading as the average TOM only considers propertiesthat have been sold. Furthermore, traded properties are heterogeneous.Since these features differ over the cycle, the average TOM could providewrong signals about market liquidity. These problems are more severe inmarkets where properties trade infrequently. In this chapter, a methodol-ogy is provided that allows for the construction of constant-quality hous-ing market liquidity indices in thin markets that can be estimated up tothe end of the sample. The latter is particularly important since marketwatchers are generally interested in the most recent information regard-ing market liquidity and less in historical information. Using individualtransaction data on three different types of Dutch municipalities (small,medium, and large) it is shown that the average TOM overestimatesmarket liquidity in bad times and underestimates market liquidity in goodtimes. The option to withdraw is the most important reason why the av-

13

CHAPTER 2. REGIONAL CONSTANT-QUALITY HOUSINGMARKET LIQUIDITY INDICES

erage TOM is misleading. Furthermore, constant-quality liquidity leadsthe average TOM and price changes. The indices not only show thatilliquidity is higher during busts, but also that liquidity risk is higher.Additional results suggest that setting a high list price relative to the es-timated value results in a higher TOM, but this effect differs over time.Both the list price premium and the effect on sale probability are higherduring busts. Differences in housing quality over the cycle, however, alsoplay a significant role. Finally, the method allows for the construction ofindices that are more robust to revisions, especially in thinner markets.1

2.1 Introduction and motivation

Market liquidity is frequently used by researchers, policymakers, andpractitioners to assess the current market conditions. In research, forexample, there is ample evidence that developments of market liquidityforeshadow price developments (De Wit et al., 2013; Carrillo et al., 2015;Van Dijk and Francke, 2018). Policymakers look at market liquidity toidentify hot markets (DNB, 2016; Hekwolter of Hekhuis et al., 2017) andbrokers use liquidity to assess the market situation (NVM, 2016).

The definition of market liquidity stems from the financial economicsliterature and refers to the ease at which assets can be traded (Brun-nermeier and Pedersen, 2009). Frequently, the average time on market(TOM) of sold houses is used as market liquidity indicator of the prob-ability of sale. The TOM is very much related to market liquidity sincea longer TOM is more costly to the holder of the asset. A quicker salewould imply that it is easier to trade the asset, at least from a hold-ers’ perspective. There are many examples of TOM analyses in thehousing market literature, including Belkin et al. (1976); Haurin (1988),and more recently Carrillo and Pope (2012). The average TOM of soldhouses, however, might provide a misleading view regarding market liq-uidity.

To illustrate, consider two houses with equal listing prices. House(1) is put on the market in January and sold in March, while house (2)

1Acknowledgments: I thank seminar participants at the AREUEA 2017 Interna-tional conference, the Real Capital Analytics Index Seminar 2017, and the Universityof Amsterdam brownbag seminar. In particular thanks to John Clapp, Martijn Droes,Peter van Els, Jakob de Haan, Larisa Fleishman, Marc Francke, Alex van de Minne,and Hans van Ophem for insightful comments and feedback. Finally, thanks to theNVM for supplying the data.

14


enters the market in June and is sold in July. Hence, the TOM for house(1) is 2 months and the TOM for house (2) is 1 month. What can weinfer from this? Did market liquidity improve between the sales of thetwo houses? It might as well be the case that house (1) was a very well-maintained property traded in a homogeneous market, whereas house(2) was a very badly maintained property traded in a heterogeneousmarket. Furthermore, the listing prices of one of the properties mightbe set “strategically” to ensure either a higher selling price or lowerTOM.

This example is a drastically simplified view of reality and in cal-culating the mean some of these problems might cancel out. However,some of these may still exist when considering multiple properties indetermining the market situation, especially in thin markets (marketswith few transactions). But perhaps even more importantly, there is acensoring problem as only actually sold listings are included when usingthe mean TOM of sold properties as measure for the probability of sale.There could be a third house that is withdrawn and is not consideredin the mean TOM of sold properties. Disregarding withdrawn proper-ties is problematic, since not all information regarding the ease at whichassets can be traded is used. Additionally, the time that a house waslisted before it is withdrawn, i.e. the TOM of withdrawn properties, canprovide useful information

If these features differ over the cycle, these problems will be ampli-fied. The number of withdrawals with respect to the number of listingsin the Amsterdam region between 2005 and 2016 is presented in Figure2.1. Before the bust, starting in 2007, the percentage of houses that arewithdrawn was about 10 – 15 %. During the the bust (the through ofthe Dutch housing market was between 2011 and 2013) the percentageof properties that are withdrawn was much higher, about 30%. In recentyears, starting in the second half of 2013, the market started recovering.This was accompanied by a significant decrease in the number of with-drawals. The fact that the probability of withdrawal is not constant overthe cycle, has implications for estimating liquidity indices based on theprobability of sale. Another feature that might differ over the cycle isquality. This is examined in earlier work by Clapp et al. (2017). Theseissues, among others, will be discussed in this chapter.

Analogous to existing regional constant-quality house price indices,this chapter proposes to create regional constant-quality liquidity in-dices. The method creates a measure for the probability of sale that

15


Figure 2.1: Fraction of houses withdrawn over the cycle in the Amsterdam region,2005-2016.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

0

0.1

0.2

0.3

0.4W

ithdra

wals

/Numberoflistin

gs

takes withdrawals and quality differences into account. There are otherpapers that propose such corrections (Carrillo and Pope, 2012; De Witand Van der Klaauw, 2013). I propose a method that allows for the con-struction of regional constant-quality market liquidity indices and addto the literature in three ways. Firstly, the indices can be constructedreliably up to the end of the sample (i.e. until the most recent datacomes in). This is quite an important aspect since policymakers andother market watchers are interested in current market liquidity (andnot only in the historical situation). Secondly, a novel feature of themethod is that it allows to construct indices for markets in which trans-actions and withdrawals occur infrequently. Previous papers that createconstant-quality liquidity indices do not attempt to tackle the data spar-sity issue explicitly. By replacing fixed effects with a stochastic trend,it is possible to generate liquidity indices for small markets. This inno-vation has a positive side effect, namely that the indices become morerobust to revisions (i.e. the change of the index in the past due to theaddition of new data). Thirdly, the method allows for the examinationof calendar time-varying effects of housing characteristics on sale proba-bility. Examining these time-varying effects is interesting, as these couldalso give an indication about the market situation (besides the indicesthemselves).2

2This should not be confused with time-varying effects related to the durationof market time (i.e. duration dependence). This is examined for the Dutch housing

16


The article closest to the current research is Carrillo and Pope (2012),in which for a large suburb of Washington D.C. annual and quality-adjusted TOM distributions and hazard functions are created and an-alyzed. This has been subsequently extended by Carrillo (2013) withother heat measures for the housing market. Although the methodol-ogy of Carrillo and Pope (2012) theoretically allows for the creation ofquarterly or monthly indices at a local scale, it is not possible to createindices at the end of the sample. The main reason is that Carrillo andPope (2012) look at the ex ante distribution of the TOM, hence at theexpected market time when the house is listed. Close to the end of thesample, only houses that are sold quickly are included in the sample andthis will result in biased index estimates. Therefore, the proposition hereis to create an index based on the realized TOM of sold and withdrawnhouses, rather than on the expected TOM.

The downside is that the presented method is not able to correct forright-censoring (i.e. correct for the number of properties that are still onthe market at the end of the sample). In stable times, right-censoringis exogenous and should not bias the indices. In some periods, however,this could result in a loss of information.3

Another issue that is tackled in this chapter is related to the sparsityof data. Issues regarding data sparsity arise when creating quarterlyindices for somewhat larger markets or when creating annual indices forthe smallest markets. Another paper close to the current research is thatof Carrillo and Williams (2015), who create quarterly “Repeat-Time-On-The-Market” indices for 15 MSAs in the US. The “smallest” region inthis paper (Medford, OR) still includes around 707 listings per quarteron average. The smallest market in the current research contains 37listings per quarter on average. As the repeat listings method discardssingle sales, this also poses difficulties for estimating the indices for smallmarkets, as there is less information available to estimate the indices.

This chapter also looks briefly at the determinants of the TOM (Hau-rin, 1988), the relationship between liquidity and list prices (Knight,2002), and the relationship between liquidity and transaction prices(Fisher et al., 2003; Goetzmann and Peng, 2006; Dube and Legros, 2016;Van Dijk et al., 2018).

The results show that the mean TOM of sold properties overesti-

market in De Wit and Van der Klaauw (2013).3For example, in a crisis the market time of properties still on the market might

also increase, and this could also give a useful indication regarding liquidity.

17


mates market liquidity in bad times and underestimates market liquid-ity in good times with respect to the constant-quality indices. Perhapseven more importantly, the mean TOM lags behind the constant-qualityindices. This indicates that it is better to use constant-quality marketliquidity as leading indicator compared to the mean of the TOM of soldproperties. While it is not the main subject of study of this chapter,market liquidity is also shown to have a large commonality with trans-action prices. When liquidity increases, prices increase as well and viceversa. Consistent with existing literature, changes in market liquidityGranger cause price changes. Furthermore, it is shown that not onlyilliquidity increases in down markets, but that the uncertainty regard-ing liquidity, i.e. liquidity risk as measured by the standard deviation,increases as well. This notion is consistent with the evidence from othermarkets such as the bond and stock markets.

The examination of the determinants of the probability of sale sug-gests that setting a relatively high list price compared to the estimatedvalue results in a higher TOM. However, the effect is not constant overtime. Not only is the list price premium higher during busts, but thetotal effect is also larger. In very hot markets, the average list pricepremium becomes a list price discount and the effect becomes positive.Most likely, the reason is that sellers change their behavior in this mar-ket, a phenomenon documented for the markets that are examined inthis study. The most important factor explaining the difference betweenthe average TOM and constant-quality liquidity indices is the possibil-ity to withdraw. Quality differences, however, also play a significantrole. Finally, the results show that the magnitude of revisions sub-stantially decreases when replacing fixed effects with a stochastic trend.The added value of the stochastic trend becomes larger as the marketbecomes thinner.

The remainder of the chapter is structured as follows. The next sec-tion presents the model and the data discussion. Section 2.3 presentsthe indices and a discussion on the commonality with transaction pricesand liquidity risk. Section 2.4 offers additional analyses regarding thedeterminants of the TOM, a decomposition analysis, and the robust-ness with respect to revisions and correlated unobserved heterogeneity.Finally, section 2.5 concludes.

18


2.2 Model and data

When a house is on the market, it can either be sold or withdrawn. Thedecision to sell or withdraw can therefore be characterized as competingrisks. The hazard function is then defined as the probability of a sale orwithdrawal, conditional on survival up to that moment. The dependentvariable is the time it takes for a house to be sold or withdrawn (TOM).The TOM is modeled in a hazard framework.

Besides conditioning on survival time, it is also desirable to conditionon other covariates, in this case housing characteristics. This is usuallydone in a proportional hazard framework. Part of these covariates are,for example, calendar-time dummy variables that indicate in which pe-riod (i.e. annual, quarterly, or monthly dummies) a sale or withdrawaltook place. These dummy-variables account for (time) fixed effects.

Intuitively, the coefficient on the dummy variable indicates the shiftin the hazard rate. The size of the shift in the hazard rate indicatesthe magnitude of change in the probability of sale in this period. Thedummy coefficients subsequently form an index of how the probabilityof sale has evolved of time. Note that these coefficients are conditionedon housing characteristics. By modeling the TOM in a competing riskshazard framework, the hazard rate of sale and the dummy coefficientsalso take withdrawals and the time that the house has already been onthe market into account.

2.2.1 Model

The conditional (on current TOM and property characteristics) proba-bility of sale or withdrawal are given by their respective hazard functions.Let t be the time the house is on the market, the proportional hazardsof sales (s) and withdrawals (w) are given by:

λj(ti,xi, zi, νj) = λ0,j(ti) exp(µi,j), (2.1)

µi,j = exp(xiβj + ziαj + νj). (2.2)

Here, subscript i for i = 1, ..., N denotes the property and subscript jfor j = s, w denote the competing risks, sales and withdrawals. Further,λ0,(ti) are the baseline hazard functions, xi is a row vector of size K ofobserved housing and other characteristics including a constant, and βis a vector of corresponding coefficients of size K. Next, zi is a (T −

19


1) row vector of calendar time dummy variables in which the housewas sold or withdrawn, and α is the corresponding (T − 1) coefficientvector. Note that the coefficient vectors of the time dummy variablesand covariates are risk-specific. Finally, νj denotes the hazard-specificunobserved heterogeneity term that is allowed to be correlated acrosshazards.

Theoretically, besides a sale or withdrawal, a third option option ispossible: The house can still be on the market at the end of the sample(i.e. right-censoring). Practically, however, it is not feasible to includethese observations as right-censored observations (see the discussion atthe end of this section). For now, however, assume that right-censoringis also a possibility. In this case, the likelihood contribution of thecompeting risk model consists of three types (Jenkins, 2005):

Ls = fsSw : exit the market through a sale,

Lw = fwSs : exit the market through a withdrawal,

Lc = SsSw : still on the market at the end of the sample,

where f and S are the density and survivor functions, respectively. Next,let δ be a variable that indicates whether the observed property is sold(δs = 1), withdrawn (δw = 1) or right-censored (δs = δw = 0). Theindividual likelihood contribution can in this case be written as:

L = (Ls)δs(Lw)δw(Lc)1−δs−δw , (2.3)

which can be written as (see Jenkins, 2005):

lnL = ` = [δs lnλs + lnSs] + [δw lnλw + lnSw], (2.4)

= `A + `B. (2.5)

Assuming that the correlation between the unobserved heterogeneity ofthe two competing risks is zero (i.e. Cov(νs, νw) = 0), the likelihoodfactorizes into two parts (A and B) and can be maximized separately.This is done by maximizing the partial likelihood of each competing riskand treating the other risk as censored (Cameron and Trivedi, 2005).Note that part A only depends on parameters from the sale hazardrate and survivor function and part B only on parameters from the

20


withdrawal functions. Assuming the relatively flexible (2-parameter)Weibull distribution as baseline hazard function with shape parametersρj ∈ (0,∞), scale parameters µj ∈ (0,∞), the log of the hazard functionis given by:

ln[λj(ti,xi, zi, νj)] = ln[µi,jρjtρj−1i ]. (2.6)

The log of the survivor function is given by:

ln[Sj(ti,xi, zi, νj)] = ln[exp(−µi,jtρj )] = −µi,jtρji . (2.7)

The loglikelihood is given by:

`(ti,xi, zi, νs, νw) =

N∑i

{δs ln(µi,sρst

ρs−1i ) + ln(−µi,stρsi )

}+

N∑i

{δw ln(µi,wρwt

ρw−1i ) + ln(−µi,wtρwi )

}= À + `B.

(2.8)

In the case of uncorrelated unobserved heterogeneity, the heterogene-ity gets absorbed in the risk-specific constants, which are included inβs and βw (part of µi,s and µi,w). The assumption of no correlatedunobserved heterogeneity is made since the focus in the present studyis on smaller markets and the problem of correlated unobserved het-erogeneity is expected not to play a major role. Also, with correlatedunobserved heterogeneity, convergence is much more difficult to achieveand the estimation takes substantially longer. Section 2.4.6 includesa robustness check and provides indices corrected for correlated unob-served heterogeneity. The unobserved heterogeneity is assumed to benormally distributed.

Note that even though À and `B are maximized separately, the sale(withdraw) likelihood does take withdrawals (sales) into account. Forexample, when withdrawals would not be taken into account in À, thelast part of the likelihood, ln(−µi,stρsi ) would be different as this does notonly depend on observations where δs = 1. This impacts the estimatedcoefficients which are included in µi,s.

21


In order for the methodology to work in thin markets, the calendartime-fixed effects are replaced by a stochastic structure. More specifi-cally, these are modeled as a random walk. The latter is similar to realestate price applications in Francke and De Vos (2000), Francke (2010),and Geltner et al. (2017) who allow for a local linear trend, which alsoincludes the random walk specification. The random walk specificationis given by:

ατ,j = ατ−1,j + ετ,j , (2.9)

where ετ,j ∼ N(0, σ2ε,j) for τ = 1, ..., T for j = s, w, and with α1 = 0.

Note that t represents the duration to sale/withdraw and τ representsthe calendar time period. At this point, it might be useful to pointout that when the random walk structure on α is left out (equivalentto setting σ2

ε,j to a very large number), the model is a regular survivalmodel with calendar time fixed effects.

The coefficients that need to be estimated in this procedure are αj

(calendar time effects), βj (coefficients on property characteristics), ρj(shape parameters), and σε,j (signals of αj).

Taking into account withdrawals partly corrects for a censoring prob-lem that occurs. However, right-censored observations (houses that arestill for sale at the end of the sample period) are removed in the currentsetup. The reason is that the period in which the sale of withdrawalwill take place is not known yet. An alternative would be to includedummy variables for the period the house has been listed (a setup moresimilar to Carrillo and Pope, 2012). This, however, causes a downwardbias in the coefficients on the dummy variables of the final time periods.The reason is that only houses that are sold or withdrawn quickly areincluded in the sample which drives the estimated TOM for that perioddown (see Appendix A.1). Another alternative is to follow De Wit andVan der Klaauw (2013) and assume that the time of exit is equal tothe time the house was at the market at the end of the sample. Thedownside of this alternative is that the “observed” TOMs of these obser-vations will be artificially low, so that this will result in an upward biasin the calendar-time effects of sold properties at the end of the sample.4

If the problem at hand is to merely control for these fixed effects, thiswill not pose an issue. However, in the present study, the interest is on

4The average TOM of the withdrawn observations will be lower at the end of thesample, resulting in a higher index value of sold observations.

22


the estimated values of these coefficients (αt,s) as these form the liquidityindex. The downside, however, is that some information is disregarded.For example, in periods of crisis, an increasing number of houses thatremain on the market also might give a useful indication regarding themarket situation. In stable markets, however, right-censoring should belargely exogenous and therefore should not play a large role.

2.2.2 Estimation

In order to estimate the model, the parametric Bayesian ProportionalHazard Model of Dellaportas and Smith (1993) is extended with a stochas-tic calendar time trend.5 In the estimation procedure uninformative pri-ors are used: βj ∼ N(0, 10), ρj ∼ Log-Normal(0, 1), and σε,j ∼ InverseGamma (3, 1).

Markov Chain Monte Carlo (MCMC) techniques are used to evaluatethe posterior density. More specifically, the RStan package is employedthat uses the No-U-Turn-Sampler (NUTS) (Hoffman and Gelman, 2014;Carpenter et al., 2016).6 To determine the mixing and convergenceof the models the R and Neff statistics are used (Francke et al., 2017).Additionally, the Monte Carlo error, the Heidelberger-Welch stationarityand half-width statistics are examined. To compare the model fit, theWatanabe Akaike information criterion (WAIC; Watanabe 2010) andleave-one-out cross-validation statistic (LOO; Vehtari et al. 2017) areconsidered.

The variables are non-centered reparameterized for estimation pur-poses (i.e. Matts’ trick; Betancourt and Girolami, 2015). This implies

5Although there are non-parametric methods to estimate the hazard functioncorrected for covariates, it is not clear how to apply a random walk structure on thecoefficients in this case. Moreover, the thinness of the market poses additional difficul-ties to estimate the model non-parametrically. Therefore, in this chapter parametricmethods are applied. Carrillo and Pope (2012) have looked at the ex ante distribu-tion of the sale probability, conditional on housing characteristics, in a non-parametricfashion. The analogy in this case would be to include calendar-time dummies thatindicate in which period the house was listed. A more detailed empirical compari-son between the presented methodology and the methodology by Carrillo and Pope(2012) is offered in Appendix A.1.

6Four parallel chains with different initial values and 2,000 (including 1000 warm-up) iterations per chain are used. Therefore, the maximum effective sample size is4,000. In the small market, convergence was more difficult to achieve, therefore thetotal number of samples per chain is set to 5,000 (including 2,500 warm-up). Thechains are not thinned.

23


that instead of sampling αj directly, the innovations in αj are sampledwith the prior N(0, 1), these innovations are then multiplied with σε,j toobtain αj . This drastically improves the computational time and con-vergence is more easily achieved. The estimation time depends on thenumber of observations. On a modern computer with 4 cores and 32GB RAM-memory, the estimation time (without correlated unobservedheterogeneity) ranges from 2 hours (small market, 2,175 observations)to 24 hours (large market, 113,005 observations).

2.2.3 Data

The main source of data originates from the Dutch Association of RealEstate Brokers and Real Estate Experts (NVM). The data contain alarge share of housing transactions within the Netherlands from 2005–2016 including many property characteristics. However, not all housestransacted or withdrawn are included in this database. Earlier arti-cles that use the NVM database report a market share of around 75%,(De Wit et al., 2013; Van Dijk and Francke, 2018).

The focus in this chapter is on three different types of markets: small,medium and large markets. The small market is Aalsmeer, represent-ing a situation in which the problem of data sparsity is large. Themedium market is Amstelveen, in which there might be data sparsityproblems during some periods (e.g. in busts). The large market is Am-sterdam, in which there are no data sparsity problems. Although thesemarkets are relatively close in terms of geographical distance, they canbe characterized as different markets. Amsterdam is the largest cityin the Netherlands. Amstelveen is a medium-sized suburban munic-ipality at the southern border of Amsterdam, whereas Aalsmeer is asmall suburban municipality at the southern border of Amsterdam. Ta-ble 2.1 includes descriptive statistics of the markets. The small marketcontains, on average per quarter, around 37 sales, the medium mar-ket around 187 sales, and the large market around 1,883 sales. Thesmall market contains approximately 13,000 houses in 2016, of whicharound 8,100 (62%) are owner-occupied. The medium market contains43,000 houses, of which 19,500 (45%) are owner-occupied. Finally, thelarge market contains 425,000 houses of which 126,000 (30%) are owner-occupied (Statistics Netherlands, 2017).

The data base includes the date the house was put on the marketand the date the house was sold or withdrawn, hence it is possible to

24


infer the TOM. Some houses are withdrawn and almost immediatelyrelisted. These observations can cause a bias in the indices as the TOMseems shorter than the true one. Therefore, if a house is relisted within90 days, the original date of when the house was put on the market isused to calculate the TOM. When a house is still on the market at theend of the sample, the property is removed. As mentioned before, thisthrows away some information and could give some problems at the endof the sample. For example, in the beginning of the bust the amount ofnot yet sold or withdrawn properties might increase. However, at theend of the sample there are only 16 (small market), 57 (medium market)and 534 (large market) properties not yet sold or withdrawn in the threemarkets. This is less than 1% of the total sample size in each marketand therefore should not pose a big problem.

The property characteristics for which the liquidity indices are con-trolled for are log size, squared log size, dummies for gardens, park-ing places, landleases, maintenance (bad, normal and well-maintained),construction period (before 1905, 1906-1944, 1945-1990, 1991-2000, after2001), and property type (terraced, back-to-back, corner, semi-detached,detached, ground floor split level apartment, upper floor split level, otherapartment).7

Besides these property characteristics, the list price premium is ex-pected to influence the TOM. Following Genesove and Mayer (2001),Bokhari and Geltner (2011), and Clapp et al. (2017), the list price pre-mium is defined as the difference between the list price and the estimatedmarket value of the property at the time of entry.8 The estimated valueat the time of entry is determined by a hedonic price model, which isincluded in Appendix A.2. In case a house is relisted and the list pricehas been revised, the first known list price is used.9

The average percentage of houses that are withdrawn from the mar-ket is fairly similar in the three markets, around 18%. Over time, how-

7Some properties are freehold whereas others and leasehold, this is corrected forby including a dummy-variable in case the property is leasehold.

8Clapp et al. (2017) actually use the residual of list price on estimated transactionvalue, controlling for anchoring. However, this requires a repeat sales framework,which is problematic in the context of thin markets. Furthermore, the residual wouldget rid of cyclical variation in the list price premium, which is also a topic of interestin the case of this study.

9List price reductions might also influence the TOM, see, for example, De Witand Van der Klaauw (2013). This issue is outside the scope of this chapter and willnot be pursued here.

25


ever, there are substantial differences in the fraction of houses with-drawn, ranging from 32% in 2012 to 10% in 2016. The average TOM ofboth sold and withdrawn houses is shorter in larger markets. This alsoholds for the standard deviation. Interestingly, the standard deviationis also lower when the TOM is shorter for both sold and withdrawnproperties. This indicates that the observed TOMs are more dispersedand possibly less informative in times of crisis.

Finally, the liquidity indices will be compared to transaction priceindices. The transaction price indices are constructed by estimating aHierarchical Trend Model (Francke and De Vos, 2000; Francke and Vos,2004). This method allows to construct price indices in thin markets.Appendix A.3 includes a more elaborate discussion on the price indexestimation procedure.

26


Table

2.1

:D

escr

ipti

on

of

the

thre

em

ark

ets

bet

wee

n2005

and

2016.

Yea

rN

SW

TO

MS

SD

TO

MS

TO

MW

SD

TO

MW

2005

9624

8109

1515

120.9

130.1

213.7

200.6

2006

10143

8824

1319

99.1

117.9

197.2

187.2

2007

10234

9090

1144

78.5

101.6

161.2

173.3

2008

10287

8533

1754

76.0

94.4

138.6

147.4

2009

10425

7499

2926

112.9

113.3

185.3

144.4

2010

10311

7431

2880

141.4

155.8

248.0

200.7

2011

9929

6926

3003

144.3

157.9

271.6

219.6

2012

10041

6838

3203

171.4

174.9

304.8

228.4

2013

9589

6671

2918

189.7

211.5

368.8

264.6

2014

11532

9737

1795

158.1

219.5

349.9

295.6

2015

12358

10991

1367

96.2

159.0

262.0

282.2

2016

11610

10457

1153

58.6

103.3

169.7

222.5

Mark

etN

SW

TO

MS

SD

TO

MS

TO

MW

SD

TO

MW

Sm

all

2175

1758

417

201.7

208.2

331.4

257.6

Med

ium

10903

8961

1942

144.1

175.0

292.1

245.9

Larg

e113005

90387

22618

111.9

149.7

248.1

227.5

Note

s:N

(Num

ber

of

obse

rvati

ons)

=S(S

old

)+W

(Wit

hdra

wn),

TO

M=

mea

nT

OM

,SD

=Sta

ndard

Dev

iati

on.

27


2.3 Market liquidity and risk

2.3.1 Liquidity indices

Three different quarterly liquidity indices are presented for these mar-kets between 2005 and 2016: (i) an index based on the mean TOM ofsold properties, (ii) a constant-quality index that does not contain therandom walk structure (i.e. a regular parametric survival model), and(iii) a constant-quality index that contains the random walk structure.The indices for the three markets are shown in Figure 2.2. Notice that ahigher index indicates a higher TOM/lower probability of sale and thusa more illiquid market.

First of all, note the large differences between the indices based onthe mean of sold properties and the two constant-quality indices. Gener-ally, in times when the TOM is increasing, the constant-quality indicesare above the mean indices. This indicates that the mean of sold prop-erties underestimates market illiquidity during crises. Conversely, whenmarket liquidity is increasing, the constant-quality indices are below themean of sold properties. In other words, not only does it take longerfor houses to be sold, the quality of the houses that are sold is differentand/or the withdrawal probability is different. This shows the impor-tance to correct for property characteristics and withdrawals.

Moreover, the mean TOM indices lag behind the constant-qualityindices. The start of the bust is visible earlier in the constant-qualityindices than in the mean TOM of sold properties. Furthermore, therecovery is earlier visible. This phenomenon is most clear in the mediumand large market, but it also holds for the small market. More formally,a Granger causality analysis shows that the constant-quality randomwalk index Granger causes the mean TOM of sold properties at the1% level.10 This suggests that the leading indicator role that the TOMgenerally has for policymakers and brokers may even be stronger if theTOM is corrected for quality and withdrawals.

There are differences between the index without and with random

10The Granger causality analysis is based on a Panel VAR with 2 lags, but theresults are robust for lag lengths of 1 to 4 quarters. The Panel VAR is estimated byOLS with heteroscedastic robust standard errors. As the time-dimension is sufficientlylarge (48 quarters), the Nickel bias is expected to be negligible and a GMM approachis not necessary.

28


walk structure. Especially in the small market, the random walk indexcontains significantly less noise. Comparing between the different mar-kets, the results suggest that the difference between the indices becomessmaller as the markets become larger. Also, when the data contain lessnoise and are more informative about market liquidity (i.e. during calmtimes and in larger markets) the presented methodology offers similarresults to more conventional methods. However, when there is morenoise and the data are less informative (i.e. during crises and in smallermarkets), the differences are much smaller. However, when there is morenoise and the data are less informative (i.e. during crises and in smallermarkets), the random walk structure adds substantial value.

Note that since the indices are indexed at 100 in the first time pe-riod, this makes it difficult to compare between the municipalities. Byindexing the indices at the unconditional mean (i.e. the average TOMof sold properties over the whole sample, see Table 2.1) in the first timeperiod, the comparison becomes easier. In every quarter, liquidity washighest in the large market, and, for most of the time, liquidity waslowest in the small market. In all markets, there is an upward trendin the TOM during the bust, starting in 2008 and lasting until 2013(Figure 2.3). The Dutch housing market started recovering in late 2013,resulting in a higher sale probability and lower TOM. At the end of thesample, market liquidity surpassed pre-crisis levels. The decrease in thesale probability during the crisis (2007Q4–2013Q1) was the largest in thelarge market: the sale probability was roughly 5 times smaller 2013Q1than in 2007Q4. The recovery (2013Q1–2016Q4), however, was alsostrongest: the sale probability was more than 6 times larger in 2016Q4than in 2013Q1.

29


Figure 2.2: Two constant-quality liquidity indices and an index of the mean TOM ofsold properties, 2005-2016.

(A) Small market (Aalsmeer)2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

Mark

etilliquid

ity

(2005Q1=100)

(i) Mean TOM (sold only) (ii) Constant-quality

(iii) Constant-quality, RW

(B) Medium market (Amstelveen)

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

Mark

etilliquid

ity

(2005Q1=100)



(C) Large market (Amsterdam)

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

Mark

etilliquid

ity

(2005Q1=100)



Note: a higher index indicates a higher TOM/lower probability of sale and a moreilliquid market.

30


Figure 2.3: Comparison of market liquidity (random walk indices) between the threemarkets, 2005-2016.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

Mark

etilliquid

ity

(2005Q1=unconditionalm

ean) Small Market (Aalsmeer) Medium Market (Amstelveen)

Large Market (Amsterdam)

2.3.2 Commonality with transaction prices

The link between transaction prices and market liquidity is omnipresentin the literature. For example, Fisher et al. (2003, 2007) provide amethodology for controlling for varying liquidity in price indices andit is generally accepted that market liquidity leads transaction pricechanges (De Wit et al., 2013; Carrillo et al., 2015; Van Dijk and Francke,2018). Whereas it is outside the scope of this chapter to create constant-liquidity price indices, this section explores the commonality of liquidityand transaction prices (see Chapter 3 for a method to create constant-liquidity price indices). First, quarterly constant-quality transactionprice indices for the three markets are estimated using a HierarchicalTrend Model (Francke and De Vos, 2000; Francke and Vos, 2004). Theseprice indices are also constant-quality and are controlled for by the sameproperty characteristics as the liquidity indices. A more detailed discus-sion on the transaction price index estimation is included in AppendixA.3.

Figure 2.4 presents the estimated transaction price and liquidityindices for the three markets. Note the similarity between the devel-opment of transaction prices and liquidity. As expected, illiquidity ishigher when prices are low. The contemporaneous correlation betweenthe level of illiquidity and prices is -0.41. The contemporaneous correla-tion between the first differences is -0.55. Furthermore, developments inliquidity foreshadow price developments. The turning point from boom

31


to bust is around 3 quarters earlier in the liquidity indices compared tothe price indices.11 Moreover, a Granger causality analysis shows thatchanges in market liquidity Granger cause changes in prices at the 5%level.12

11The turning point from boom to bust is defined as the first quarter with negativeprice growth in all three municipalities (2008Q3) and the turning point from bust toboom is defined as the first quarter with positive price growth the three municipalities(2013Q2).

12The Granger causality analysis is based on a Panel VAR with 3 lags, but theresults are robust for lag lengths 2 to 4 quarters. The Panel VAR is estimated byOLS with heteroscedastic robust standard errors. As the time-dimension is sufficientlylarge (48 quarters), the Nickel bias is expected to be negligible and a GMM approachis not necessary.

32


Figure 2.4: Illiquidity (constant-quality and random walk, left axis) and transactionprice (right axis) indices, 2005-2017.

(A) Small market (Aalsmeer)

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

60

80

100

120

140

160

180

200

←− Bust −→

Mark

etilliquid

ity

(2005Q1=100)

60

80

100

120

140

160

180

Tra

nsa

ction

prices(2

005Q1=100)

Market illiquidity

Transaction prices


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

300

350

←− Bust −→

Mark

etilliquid

ity

(2005Q1=100)

80

100

120

140

160

180

200

Tra

nsa

ction

prices(2

005Q1=100)

Market illiquidity

Transaction prices


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

←− Bust −→

Mark

etilliquid

ity

(2005Q1=100)

50

100

150

200

Tra

nsa

ction

prices(2

005Q1=100)

Market illiquidity

Transaction prices


33


2.3.3 Liquidity risk

Another salient feature of the indices is that there is more variability inmarket liquidity in times of crisis. This is also visible in the summarystatistics as the standard deviation of the TOM of all properties is higherduring the bust (Table 2.1). Figure 2.5 shows indices based on the9-quarters centered (-4 and +4 quarters) rolling standard deviationsof the returns of the constant-quality random walk indices. These areindexed at the unconditional standard deviation of the TOM of soldproperties over the whole sample (Table 2.1). The figure indicates thatthe standard deviation of the liquidity indices becomes higher in timesof crisis as markets become thinner. In the small market, the increase inrisk is visible somewhat earlier than in the other two markets and seemsto recover somewhat more slowly after the crisis. The main picture,however, is that liquidty risk increases up to the through of 2013Q1 inall three markets. In other words, not only illiquidity is higher in badtimes, but liquidity risk is also higher. These findings are consistentwith the general asset pricing literature in which it is well documentedthat illiquid stocks and bonds also entail more liquidity risk (Acharyaand Pedersen, 2005; Acharya et al., 2013).

Figure 2.5: Comparison of market liquidity risk between the three markets, 2005-2016.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

600

700

Mark

etliquid

ity

risk

(Std

ev

ofTOM

indays)

Small Market (Aalsmeer) Medium Market (Amstelveen)

Large Market (Amsterdam)

Note: liquidity risk indices are based on 9-quarters centered rolling window standarddeviation of growth rates of the random walk indices.

34


2.4 Determinants, decomposition, and robust-ness

This section looks at the (time-varying) determinants of the TOM. Theeffect on the indices of withdrawals and quality is disentangled to ex-amine which is most important. Finally, the magnitude of revisions isexamined. To what extent do the indices change when new data comesin? And are there differences in revisions between models that correctfor quality and/or withdrawals and models that do not correct for these?Finally, some robustness checks are performed with respect to correlatedunobserved heterogeneity between withdrawals and sales.

2.4.1 Determinants of the time on market

The estimated coefficients for the control variables βs and shape pa-rameter ρs are presented in Table 2.2. The coefficient estimates are al-most equivalent across the estimations (with and without random walk).There is a substantial decrease in the WAIC and LOO statistics of therandom walk models in the small and medium market, which indicatesthat the fit of the random walk models is better than that of the mod-els without the random walk structure. In the large market, the indexwithout the structure has a slightly better fit, but the difference is verysmall. Note that the coefficients contain the effects on the sale proba-bility, so a positive (negative) coefficient indicates a positive (negative)effect on sale probability and market liquidity, and a negative (positive)effect on the TOM. There are substantial differences in the estimatesacross markets. It is difficult to attach a causal interpretation to thecoefficients. Also, the credible intervals are quite large, especially in thesmaller markets. Nevertheless, some general patterns, consistent withthe literature, seem to arise.

For example, apartments, which are usually more homogeneous thanother house types, sell quicker in the medium and large market. There-fore, this notion is consistent with the finding that the TOM of homoge-neous houses is generally shorter than that of heterogeneous house typesthat are more atypical (Haurin, 1988).13

A higher list price premium (i.e. the list price is relatively highcompared to the estimated market value at the time of listing), results

13Although atypicality is not explicitly controlled for, controlling for the separatecharacteristics per submarket has a similar effect.

35


in a lower probability of sale, hence a higher TOM. The effect of listprice premia will be discussed in more detail in section 2.4.4.

The estimates of the shape parameter ρ are very similar across themodels. For the small market, the estimate is not significantly differentfrom 1, indicating that the TOM follows a (negative) exponential distri-bution in this market. In the medium and large markets, the estimatedshape parameters are 0.96 and 0.90, respectively. This indicates thatthe probability of sale decreases as the property remains longer on themarket. This finding is consistent with the results of from De Wit andVan der Klaauw (2013).

The model statistics indicate that the mixing went satisfactorily, theMCMC-error is close to 0 and the the R (not to be confused with theR2) is close to 1. The effective sample size relative to the number oftotal samples (4000 in the medium and large market) is close to 1 in thetwo largest markets. It is smaller for the small market, but by setting alarger sample size (of 10,000) the remaining number of samples is largeenough to produce reliable results. Both Heidelberger tests are passedin almost all cases, the value in Table 2.2 indicates the mean of all testson all parameters and every tests gets the value 1 if the test is passed.

36

CH

AP

TE

R2.

RE

GIO

NA

LC

ON

ST

AN

T-Q

UA

LIT

YH

OU

SIN

GM

AR

KE

TL

IQU

IDIT

YIN

DIC

ES

Table 2.2: Estimates of the control variables for three different markets in two different models.

Variable βcq p2.5,cq p97.5,cq βrwcq p2.5,rwcq p97.5,rwcq

Small Market

Constant 0.755 -9.262 10.364 0.727 -9.898 10.430Bad Maint. (Omitted) (Omitted)Normal Maint. -0.359 -0.545 -0.168 -0.358 -0.544 -0.166Good Maint. -0.420 -0.661 -0.185 -0.419 -0.644 -0.171< 1905 (Omitted) (Omitted)1906− 1944 -0.442 -1.096 0.228 -0.448 -1.102 0.2841945− 1990 -0.550 -1.201 0.117 -0.555 -1.186 0.1501991− 2000 -0.669 -1.340 -0.004 -0.678 -1.333 0.025> 2001 -0.665 -1.304 0.019 -0.652 -1.341 0.016HT Terraced (Omitted) (Omitted)HT Back-to-Back -0.457 -0.844 -0.079 -0.450 -0.838 -0.082HT Corner 0.064 -0.070 0.202 0.062 -0.073 0.201HT Semi-Detached -0.184 -0.356 -0.019 -0.191 -0.362 -0.030HT Detached -0.387 -0.605 -0.178 -0.401 -0.608 -0.187AT Split-Level (Ground or multiple) 0.000 -0.283 0.297 0.015 -0.265 0.305AT Split-Level (Upper floor) -0.180 -0.472 0.113 -0.191 -0.469 0.093AT Other -0.049 -0.303 0.191 -0.039 -0.274 0.207log(size) -1.144 -4.579 2.151 -1.061 -4.353 2.624log(size)2 0.040 -0.237 0.341 0.030 -0.279 0.317

37

CH

AP

TE

R2.

RE

GIO

NA

LC

ON

ST

AN

T-Q

UA

LIT

YH

OU

SIN

GM

AR

KE

TL

IQU

IDIT

YIN

DIC

ES

Estimates of the control variables for three different markets in two different models (continued)


Garden 0.205 0.006 0.415 0.218 0.021 0.406Parking -0.103 -0.223 0.016 -0.094 -0.211 0.020Landlease 0.857 -0.137 1.837 0.982 0.000 1.927List Price Premium -1.875 -2.217 -1.534 -1.850 -2.195 -1.523ρ 1.001 0.963 1.036 1.005 0.968 1.041

Loglike -11446.8 -11440.8MCMC-error 0.0027 0.0020Neff 3310.2 4094.2Heid. stationarity 1.0000 1.0000Heid. Halfwidth 0.9886 0.9864LOO 22952.8 22920.1WAIC 22952.2 22919.8

R 1.0008 1.0005

Medium Market

Constant -11.601 -13.237 -10.076 -11.310 -12.929 -9.904Bad Maint. (Omitted) (Omitted)Normal Maint. -0.352 -0.414 -0.289 -0.352 -0.413 -0.291Good Maint. -0.352 -0.434 -0.273 -0.355 -0.440 -0.276< 1905 (Omitted) (Omitted)1906− 1944 0.404 -0.022 0.817 0.384 -0.036 0.8281945− 1990 0.370 -0.052 0.790 0.348 -0.073 0.7931991− 2000 -0.081 -0.497 0.351 -0.102 -0.508 0.364> 2001 -0.152 -0.580 0.300 -0.174 -0.609 0.279HT Terraced (Omitted) (Omitted)HT Back-to-Back -0.739 -0.942 -0.542 -0.741 -0.943 -0.544HT Corner -0.025 -0.091 0.045 -0.026 -0.096 0.036HT Semi-Detached -0.565 -0.680 -0.453 -0.560 -0.682 -0.450HT Detached -1.016 -1.183 -0.829 -1.006 -1.180 -0.828

38

CH

AP

TE

R2.

RE

GIO

NA

LC

ON

ST

AN

T-Q

UA

LIT

YH

OU

SIN

GM

AR

KE

TL

IQU

IDIT

YIN

DIC

ES



AT Split-Level (Ground or multiple) 0.137 -0.003 0.282 0.126 -0.016 0.268AT Split-Level (Upper floor) 0.134 -0.017 0.283 0.126 -0.014 0.279AT Other 0.222 0.080 0.360 0.217 0.075 0.351log(size) 1.811 1.387 2.277 1.779 1.366 2.216log(size)2 -0.108 -0.141 -0.077 -0.106 -0.138 -0.076Garden 0.112 -0.015 0.233 0.111 -0.010 0.238Parking -0.273 -0.332 -0.210 -0.270 -0.330 -0.210Landlease 1.397 0.813 2.074 1.377 0.729 1.962List Price Premium -1.793 -1.915 -1.656 -1.785 -1.911 -1.650ρ 0.960 0.944 0.975 0.960 0.945 0.975


R 1.0001 0.9999

Large Market

Constant -1.953 -2.167 -1.744 -1.971 -2.193 -1.750Bad Maint. (Omitted) (Omitted)Normal Maint. -0.302 -0.326 -0.278 -0.302 -0.325 -0.278Good Maint. -0.284 -0.311 -0.258 -0.284 -0.310 -0.258< 1905 (Omitted) (Omitted)1906− 1944 0.102 0.083 0.122 0.102 0.082 0.1231945− 1990 -0.163 -0.187 -0.141 -0.163 -0.186 -0.1371991− 2000 -0.213 -0.242 -0.185 -0.212 -0.242 -0.183> 2001 -0.320 -0.355 -0.288 -0.320 -0.355 -0.285

39

CH

AP

TE

R2.

RE

GIO

NA

LC

ON

ST

AN

T-Q

UA

LIT

YH

OU

SIN

GM

AR

KE

TL

IQU

IDIT

YIN

DIC

ES



HT Terraced (Omitted) (Omitted)HT Back-to-Back -0.105 -0.241 0.038 -0.106 -0.242 0.025HT Corner -0.038 -0.091 0.013 -0.040 -0.090 0.012HT Semi-Detached -0.117 -0.206 -0.030 -0.118 -0.203 -0.030HT Detached -0.382 -0.473 -0.293 -0.383 -0.477 -0.292AT Split-Level (Ground or multiple) 0.154 0.124 0.185 0.154 0.124 0.186AT Split-Level (Upper floor) 0.199 0.167 0.232 0.198 0.165 0.232AT Other 0.138 0.101 0.171 0.137 0.102 0.172log(size) -0.656 -0.714 -0.595 -0.656 -0.719 -0.596log(size)2 0.035 0.031 0.040 0.035 0.031 0.040Garden 0.126 0.102 0.148 0.126 0.104 0.151Parking -0.277 -0.303 -0.251 -0.277 -0.301 -0.252Landlease 0.100 0.084 0.115 0.100 0.084 0.115List Price Premium -1.571 -1.611 -1.531 -1.571 -1.610 -1.532ρ 0.901 0.897 0.906 0.901 0.897 0.906


R 0.9999 0.9999

Dependent variable is the probability of sale, a positive coefficient indicates a positive effect on this proba-bility. βcq and βcq,rw are the coefficient estimates of the constant-quality model and constant-quality modelwith random walk, respectively. The 95% HPD intervals are given by P2.5 and P97.5. HT = House type, AT= Apartment type. Loglike is the log-likelihood, MC-error the mean of Monte Carlo standard error for allparameters, Neff the mean effective sample size (total number of samples = 4000 and 4000 warm-up) of allparameters, Heidelberger-Welch stationarity/Halfwidth is the mean of the test of all parameters (a parame-ter gets the value 1 (0) when the test is passed (failed) at the 5% level). The WAIC is the Watanabe Akaikeinformation criterion and the LOO is leave-one-out cross-validation statistic. Finally, R is the mean Rhatstatistic for all parameters.

40


2.4.2 Time-varying effect of list price premium

Theoretically, every coefficient on property characteristics can be made(calendar) time-varying. Practically, model identification becomes muchmore complicated and simulation and sampling becomes more time-consuming. Additionally, having more observations makes it easier toidentify a time-varying effect. For illustration purposes, the coefficienton the list price premium, i.e. the premium of the original list priceover the predicted list price (based on a hedonic regression), has beenmade time-varying in the large market (Amsterdam). A more detaileddescription of the definition of this variable is included in Appendix A.2.Similar to the time-fixed effects, the time-varying coefficient on the listprice premium is structured to follow a random walk.

The likelihood takes a similar form as that of Equation 2.8. Themain difference is that the location parameters are now equal to: µi,j =exp(xiβj+ziαj+liγj+νj) ∈ (0,∞), for j = s, w. Besides the parametersas discussed in section 2.2.1, li is a (T −1) row-vector containing the listprice premium (scalar) multiplied with a (T − 1) selection row-vector(which is 1 at the quarter of withdrawal / sale and 0 otherwise). Next,γ is a (T − 1) vector containing the time-varying effects of the list pricepremium. The effect of the list price premium is also assumed to followa random walk:

γτ,j = γτ−1,j + ητ , η ∼ N(0, σ2η,j). (2.10)

The additional coefficients that need to be estimated are the signalsof γj (ση,j) and the innovations in γj . For the signal and innovationsuninformative priors are used: ση,j ∼ Inverse Gamma(3, 1) and N(0, 1),respectively. The initial value (γ1,j) is equal to the estimated effect of themodel without time-variation in the parameters (i.e. the unconditionalmean of the effect).

The estimated time-varying effect of the list price premium in thelarge market is shown in the top panel of Figure 2.6. Similar to Table2.2, the list price premium has a negative effect in all time periods. Thisindicates that a higher list price premium leads to a lower probability ofsale. The effect is somewhat smaller during the bust and starts increas-ing close to the end of the bust after which it starts increasing again.However, since the list price premium itself is also likely to exhibit cycli-cal behavior (see section 2.4.4), it is more informative to look at the

41


Figure 2.6: Time-varying effect of list price premium and 95% HPD intervals (toppanel) and the marginal effect (bottom panel) in the large market (Amsterdam),2005-2016.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−3

−2.5

−2

−1.5

−1

−0.5

0

←− Bust −→

Eff

ect

on

pro

babil

ity

Effect of list price premium

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.2

−0.1

0

0.1

0.2

←− Bust −→

Eff

ect

on

sale

pro

babil

ity

(i) Average effect on sale probability (ii) Average list price premium

Note: a higher positive (negative) value indicates a larger positive (negative) effecton the probability of sale.

42


effect multiplied with the average list price premium of sold propertiesper time period (bottom panel in Figure 2.6). This shows that the effecton sales probability becomes more negative during the bust since thelist price premia are higher during this time.

Starting in 2014, the strong recovery of the Amsterdam market isclearly visible. Especially close to the end of the sample an interest-ing phenomenon is visible. The average list price premium decreases,whereas the effect on sales probability also decreases. The average listprice premium actually turns into an average list price discount startingin 2015. The reason is the huge demand relative to supply. This resultsin a different behavior of sellers than usual. Koster and Rouwendal(2017) state that a better strategy in these times is to set a relativelylow list price, as this results in both a quicker sale and higher transactionprice. This strategy, however, only works in an extremely booming mar-ket. According to the Dutch Central Bank, the Amsterdam market wasshowing signs of overheating in 2015 and 2016 (Hekwolter of Hekhuiset al., 2017).

The result is an increase in the total (multiplied) effect (i.e. a lessnegative or even positive effect). Therefore, this time-varying coefficienton the list price premium, in combination with the average list pricepremium, could potentially be used as an indicator to spot overheatingmarkets.

2.4.3 Decomposition of effects: the effect of quality andwithdrawals

Apart from the fact that the mean TOM is noisy in thin markets, theproposed methodology corrects the mean TOM for two features: (i)quality and (ii) withdrawals. The aim of this section is to disentanglethese effects and to determine their respective importance. To examinethe effect of quality, an index (with random walk structure) is estimatedwithout the control variables. To examine the effect of withdrawals, anindex (with random walk structure) is estimated for sold properties only.In this case, `A of the likelihood in Equation 2.8 without the last part,ln(−µi,stρsi ), is maximized. This index is subsequently compared to anindex that controls for both withdrawals and housing quality.14

The results for the small, medium, and large market are shown inFigure 2.7. In all three markets the constant-quality indices are more

14But not for the list price premium to isolate the effect of quality.

43


similar to the indices corrected for withdrawals only than the indicescorrected for quality only. Hence, withdrawals seem to be the majordriver of differences. This is in line with the findings of Carrillo andPope (2012).

Although the effect of quality is smaller than the effect of with-drawals, the effect of quality is still substantial. Furthermore, the in-fluence of quality differs over the cycle. The percentage differences be-tween the constant-quality indices and withdrawal-only controlled in-dices in levels are plotted in Figure 2.8. The largest difference is 17%(in 2016Q4 in the small market). The results further indicate that theindices controlled for withdrawals and quality predict a more illiquidmarket during busts than indices controlled for withdrawals only. Moreformally, a regression of the dummies for the three depicted periods inFigure 2.8 on the difference between the indices yields statistically sig-nificant (at 1%) coefficients on the bust dummy compared to the twoother periods. The average difference between boom (either the first orsecond) and bust ranges from 3.5% to 7.1%. This indicates that illiquid-ity according to the constant-quality indicator was actually worse thanilliquidity according to the indicator not corrected for quality. This inturn implies that housing quality differs over the cycle: in busts differentquality properties sell than during booms. The conjecture that qualityis different in busts and booms has also been put forward in Clapp et al.(2017).

Although differences in quality play a role, a very large part of thedifference between the mean TOM and the constant-quality indices canbe accounted for by only correcting for withdrawals. A big advantageof not controlling for property characteristics is that it simplifies theestimation and that the the computational time decreases by about 60%.For some applications this might be an interesting alternative.

44


Figure 2.7: Decomposition of effects.


2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

300

Mark

etilliquid

ity

(2005Q1=100)

(i) Corrected for quality and withdrawals (ii) Corrected for withdrawals

(iii) Mean TOM (sold only)


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)

(i) Constant-quality, RW (ii) RW



2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

300

Mark

etilliquid

ity

(2005Q1=100)

(i) Constant-quality, RW (ii) RW



45


Figure 2.8: Difference in illiquidity due to quality.


2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.2

−0.1

0

0.1

0.2

←− Bust −→More

illi

quid

ity

inin

dex

corr

ecte

dfo

rquali

tyDifference in illiquidity due to quality Average


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.15

−0.1

−0.05

0

0.05

0.1


illi

quid

ity

inin

dex

corr

ecte

dfo

rquali

ty

Difference in illiquidity due to quality Average


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.1

−0.05

0

0.05

0.1


illi

quid

ity

inin

dex

corr

ecte

dfo

rquali

ty


Note: a bigger (positive) difference indicates a higher TOM / more illiquidity in theindex corrected for quality compared the the index not controlled for quality.

46


2.4.4 Decomposition of effects: the effect of list price pre-mium

Although the list price premium is based on the estimated market valueat the time of entry, there might be concerns that the list price itselfis endogenous to the TOM. Therefore, this subsection is devoted to in-dices that do not contain the list price premium as independent variable.Indices controlled for and not controlled for the list price premium areshown in Figure 2.9, while the differences between the indices are de-picted in Figure 2.10.

In the small market, the differences between the indices are not struc-tural over the cycle (Figure 2.10). A regression of the difference on theboom dummy variables and bust dummy yields insignificant coefficients.However, in both the medium and large market the differences are signif-icant and follow a similar pattern over the cycle. In busts, the illiquidityis lower in the indices that are corrected for the list price premium,than in booms. This indicates that indices corrected for the list pricepremium are less cyclical than the uncorrected indices. This reflectsthat list price behavior of sellers is cyclical as well (see bottom panel ofFigure 2.6). Obviously, for the large market, there is a strong similaritybetween the index differences and the time-varying effect of the list pricepremium as shown in Figure 2.6 in section 2.4.2.

Setting the list price too high during busts is more likely to occurif homeowners are loss averse (Genesove and Mayer, 2001; Clapp et al.,2017). This has also been documented for the Dutch market (Van derCruijsen et al., 2018). In other words, by not controlling for the listprice premium, the indices are more cyclical as these will pick up someof the cyclicality of the list price behavior. The question of whetherto control for the list price premium depends on the problem at hand.For example, a policymaker who wants to identify boom and bust cyclesmight want to use the indices not controlled for the list price premiumas the identification of these cycles becomes easier.

47


Figure 2.9: Illiquidity indices with and without list price premium, 2005-2016.


2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)

(i) Index with list price premium (ii) Index without list price premium


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)



2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)



48


Figure 2.10: Difference in illiquidity due to list price premium, 2005-2016.


2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.05

0

0.05

0.1

0.15

0.2


illi

quid

ity

inin

dex

corr

ecte

dfo

rquali

ty



2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.15

−0.1

−0.05

0

0.05

0.1

←− Bust −→

More

illi

quid

ity

inin

dex

corr

ecte

dfo

rli

stpri

ce

pre

miu

m

Difference in illiquidity due to list price premium Average


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

−0.05

0

0.05

0.1

0.15

←− Bust −→

More

illi

quid

ity

inin

dex

corr

ecte

dfo

rli

stpri

ce

pre

miu

m

Difference in illiquidity due to list price premium Average

Note: a bigger (positive) difference indicates a higher TOM / more illiquidity in theindex corrected for the list price premium compared the the index not controlled forthe list price premium.

49


2.4.5 Revisions

Revisions are changes that occur to previous values of an index whennew data comes in. The magnitude of revisions is a useful measure forthe quality of an index. This reflects both the precision of an indexas well as the practical usefulness for businesses and policy purposes(Francke et al., 2017).15 The expectation is that the magnitudes of therevisions are larger for thinner markets, as new data will be relativelymore influential in these markets. By introducing the random walkstructure, the expectation is that the magnitude of the revisions will besmaller.

I use a similar measure for the maginitude of revisions as Franckeet al. (2017). More specifically, indices without data on the last year(2016) are estimated and these are compared to the baseline indicesthat are estimated on all data. Next, the absolute difference betweenboth the levels (in percentage difference) and returns (in percentagepoints) are examined for the whole sample. Statistics for the magnitudeof revisions is shown in Table 2.3.

In general, the absolute average revisions (|Mean|) are substantiallysmaller for the indices estimated with a random walk structure. In levels,the magnitude of revisions is almost 4 times smaller for the RWCQmodel compared to the model without random walk. This holds for allthree markets. In returns, the magnitude is 3 times smaller in the smallmarket, 2 times smaller in the medium market and roughly the same inthe large market. Also, the maximum size of the revision is much smallerin the random walk indices than in the indices without the random walkassumption. Another feature is that the size of the revisions becomeslarger as the market becomes smaller. This holds for both levels andreturns. The reason is most likely that new data will be relatively moreinfluential in these markets.16

Finally, in the two largest markets, there are no substantial differ-ences in the degree of revisions between the three models that contain

15Francke et al. (2017) examine revisions in a repeat sales framework, in whichrevisions play a larger role than in a hedonic framework. The presented liquidityindices in the current study are estimated in a hedonic-like framework, but there stillmight be substantial revisions due to the thinness of some markets.

16In fact, even if the number of new observations is small in a thin market, thesenew observations will affect the signal more than the noise, so that the indices aremore prone to change. See Francke et al. (2017) for a more detailed discussion andsimulations of revisions with respect to the signal-to-noise ratio.

50


a random walk structure. The random walk indices not corrected forquality (but for withdrawals) seem to exhibit somewhat fewer revisionsin the small and medium markets, but the difference is very small. Inother words, controlling for quality or withdrawals does not result inextra revisions.

51


Table 2.3: Summary statistics of revisions.

Small MarketCQ RWCQ RW RWCQNC

LevelsMean -0.055 0.008 0.011 -0.080Std. Dev. 0.045 0.027 0.021 0.057|Mean| 0.061 0.016 0.014 0.084|Max| 0.141 0.133 0.106 0.183ReturnsMean -0.006 0.002 0.002 0.001Std. Dev. 0.045 0.014 0.010 0.034|Mean| 0.034 0.011 0.007 0.028|Max| 0.113 0.045 0.044 0.095

Medium MarketLevelsMean -0.038 0.010 0.005 0.004Std. Dev. 0.009 0.008 0.005 0.007|Mean| 0.038 0.010 0.005 0.006|Max| 0.052 0.024 0.019 0.027ReturnsMean -0.001 0.000 0.000 0.000Std. Dev. 0.011 0.005 0.004 0.004|Mean| 0.007 0.004 0.003 0.003|Max| 0.037 0.012 0.012 0.011

Large MarketLevelsMean -0.009 -0.001 -0.002 -0.001Std. Dev. 0.002 0.002 0.002 0.002|Mean| 0.009 0.002 0.002 0.002|Max| 0.013 0.005 0.006 0.005ReturnsMean 0.000 0.000 0.000 0.000Std. Dev. 0.002 0.003 0.003 0.002|Mean| 0.002 0.002 0.002 0.002|Max| 0.008 0.009 0.011 0.005

Revisions over 2005Q1–2015Q4 w.r.t. an extra year of data(2016Q1–2016Q4). A revision in levels is defined as the percentagedifference between the two index levels, in returns it is the percent-age point difference in returns. CQ constant-quality model withoutrandom walk structure, RWCQ a constant-quality model with ran-dom walk structure, RW is a model with random walk structure andnot controlled for quality differences, and RWCQNW is constant-quality model with random walk structure but not controlled forwithdrawals.

52


2.4.6 Correlated unobserved heterogeneity

A parametric specification is used for the unobserved heterogeneity term(νj , which is part of µj in equation 2.8). More specifically, a similarspecification as in Cameron and Trivedi (2005) is used:

ν1 = ε1 + ω1,2ε2, (2.11)

ν2 = ω2,1ε1 + ε2, (2.12)

εj ∼ N(0, σ2j ), j = s, w. (2.13)

Here subscript j denotes sale (j = 1) and withdrawal (j = 2) . Thefactor loadings ω1,2 and ω2,1 and the variances σ2

1 and σ22 are estimated.

The factor loadings are sampled with prior N(0,1) and the standarddeviations with Inverse Gamma(3, 1) priors. If the factor loadings ω1,2

and ω2,1 are non-zero, the unobserved heterogeneity terms are correlated.The posterior estimates of the factor loadings and standard deviationsare included in Table 2.4.

Since the credible intervals are quite large, there is little evidenceof correlated unobserved heterogeneity. Consequently, the indices es-timated with correlated unobserved heterogeneity are very similar tothose estimated without correlated unobserved heterogeneity (Figure2.11). The reason is probably that the markets are relatively small andhomogeneous (even the large market is quite homogeneous), thus theunobserved heterogeneity component should be small. Therefore, prac-tically, it is more convenient to estimate the model without correlatedunobserved heterogeneity (i.e assume ω1,2=ω2,1=0) as only the partiallikelihood has to be evaluated. As discussed in section section 2.2.2, thissubstantially reduces computation time.

Table 2.4: Posterior means and 95% HPD intervals for the factor loadings and stan-dard deviations of the unobserved heterogeneity components.

Small market Medium market Large marketMean p2.5 p97.5 Mean p2.5 p97.5 Mean p2.5 p97.5

ω1,2 0.013 -3.894 3.868 0.013 -3.878 4.068 0.028 -3.831 3.780ω2,1 -0.063 -3.834 3.816 0.012 -3.863 3.804 -0.117 -3.999 3.895σ1 0.481 0.139 1.488 0.489 0.137 1.513 0.483 0.137 1.522σ2 0.515 0.138 1.864 0.475 0.136 1.389 0.466 0.138 1.327

Mean denotes the posterior mean. p2.5 and p97.5 denote the lower (2.5%) and upperbound (97%) of the 95% credible intervals, respectively.

53


Figure 2.11: Illiquidity indices with and without correlated unobserved heterogeneity,2005-2016.


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)

(i) Constant-quality, RW (ii) Constant-quality, RW, with unobserved heterogeniety


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)



2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

Mark

etilliquid

ity

(2005Q1=100)


Note: a higher index indicates a higher TOM/lower probability of sale and a moreilliquid market. 54


2.5 Conclusion

In this chapter a methodology is presented that allows for the construc-tion of constant-quality liquidity indices when transaction data is sparse.These indices can be useful for policymakers, brokers, and other marketparticipants. The presented methodology addresses the problem thatheterogeneous properties are traded in different periods. Furthermore,in some periods more properties are withdrawn than in others. The lat-ter issue is treated as a censoring problem and is explicitly taken intoaccount in the methodology.

The results show that simply taking the mean of the TOM of soldproperties underestimates (overestimates) market liquidity in good (bad)times. In other words, the quality of the properties that are sold ishigher and/or the probability of withdrawal is different. Moreover, theconstant-quality indices lead the mean TOM of sold properties. This in-dicates that the leading indicator properties of constant-quality liquidityindices are better than those of indicators currently used.

One of the main advantages of the presented method is that it canalso be used in thin markets or in times of high uncertainty. In moredense markets or during “normal” times, the method is similar to moreconventional methods. In times of high uncertainty, indices producedby conventional methods are unreliable. The proposed methodologyinduces a structure that allows to also create reliable indices in thesetimes. Also, the method provides indices that are less sensitive for revi-sions than more conventional techniques. The magnitude of revisions issubstantially smaller when a random walk structure is introduced. Revi-sions are larger for thinner markets, but the added value of the randomwalk structure is also larger.

It is shown that during busts the TOM is high and market liquidityis low. Furthermore, it is shown that the constructed liquidity indicesand transaction price indices move similarly over the cycle. Consistentwith the literature, liquidity changes lead price changes. A novel finding,consistent with the general asset pricing literature, is that liquidity riskis also higher in busts. A suggestion for future research is to delve furtherin this issue. Liquidity risk, for example, can potentially be modeled ina more sophisticated way. The volatility of the dummy variables can bemodeled as a stochastic process itself (i.e. stochastic volatility). Thiscould give more accurate insights in liquidity risk. However, this mightresult in additional difficulties in the identification of the model.

55


The methodology also allows for an examination of the determinantsof market liquidity. The effects of housing characteristics are in generalnot equal across different regions. More homogeneous housing typeslike apartments generally sell quicker. A higher list price premium (i.e.a higher list price compared to the predicted list price) is related toa lower sale probability. The effect is shown to be varying over time;in busts both the average list price premium and the total effect onsale probability increase. Since 2015, the list price premium turns, onaverage, into a list price discount. The reason is that sellers change theirbehavior due to the extreme tightness of the market.

The presented methodology corrects for both quality and withdrawals.The results suggest that withdrawals are the main driver of the differ-ence between the average TOM of sold properties. Quality, however,also plays a significant role. The results suggest that the quality of soldproperties is different over the cycle. Nevertheless, correcting for with-drawals is the most important issue. By not controlling for quality, theestimation procedure is much quicker. Therefore, for some applicationscorrecting only for the number of withdrawals might be enough.

56

Chapter 3Revisiting supply and demandindices in real estate

Oh, you don’t have a penanymore. Supply and demand,bro.

Brad, The Wolf of Wall Street

In this chapter we disentangle reservation prices of buyers and sellers forcommercial real estate at the city level. To do so, we extend the Fisheret al. (2003, 2007) methodology to a repeat sales indexing framework.This has the advantage that it takes care of all unobserved heterogene-ity, which is an important consideration in commercial real estate. Fur-thermore, it allows for the construction of supply and demand indiceswithout the need for many property characteristics or assessed values.A key innovation in our methodology, which also enables granular indexproduction, is our use of a Bayesian, structural time series model to es-timate the index. By introducing these features, we are able to estimatereliable, robust supply and demand indices for all major metropolitan ar-eas in the US. Here we focus on two very different urban markets: NewYork and Phoenix. Consistent with the notion of pro-cyclical liquidity,we find that buyers’ reservation prices move much more extremely andearlier than sellers’ reservation prices. Our results show that the demandindices in both New York and Phoenix went down a full year earlier than

57

CHAPTER 3. REVISITING SUPPLY AND DEMAND INDICES INREAL ESTATE

the supply indices during the crisis.1

3.1 Introduction and Motivation

Two of the most salient characteristics of private property investmentmarkets in comparison with and distinct from public securities markets(such as the stock market) is the pronounced cyclicality of the privateproperty markets and the degree to which their trading volume is pro-cyclical. Price and volume move together in real estate.

This has profound implications for the nature and meaning of assetprice indices in real estate, and these implications are often overlooked.Price indices are essentially means for comparing the relative differencesin prices across time. For example, based on a price index, we might saysomething like, “values crashed by over 30 percent during the financialcrisis.” But what if over the same period trading volume crashed byover 60 percent. Then the “30 percent” drop in the price index does notfully capture the degree to which the asset market collapsed. Nor, byitself does the “60 percent” statistic on trading volume. The observedtrading prices that were 30 percent lower than their previous peak asindicated by the price index during the crisis also reflected the fact thatproperty owners chose to sell far fewer properties. The market was muchless “liquid” (as this term is used in real estate). Comparing price levelsbetween the peak and trough of the price index is comparing “applesversus oranges” to a large extent.

Fundamentally, observed prices in consummated transactions, onwhich price indices are based, reflect the reservation prices on the twosides of the asset market, potential buyers on the demand side, andproperty owners, i.e. potential sellers, on the supply side. The volumeof trading also reflects these reservation prices. A much more completepicture of the state of the asset market can be obtained if we can trackthe central tendencies of these two reservation prices separately overtime. The two reservation prices depict more directly the underlying

1This chapter is based on Van Dijk et al. (2018). Acknowledgments: Many thanksgo to seminar participants at the De Nederlandsche Bank, ERES 2018, MIT RealEstate Price Dynamics Platform, Ortec Finance, the RCA Index Seminar 2017, andWeimar School 2018 conferences and seminars. In particular thanks to John Clapp,Martijn Droes, Peter van Els, Edward Glaeser, Marc Francke, and Jakob de Haan,Rober Hill, and Miriam Steurer for providing detailed comments. Finally, thanks toReal Capital Analytics for supplying the data.

58


valuation decisions of the two sides of the market. Furthermore, track-ing of reservation prices collapses the price and volume dimensions into asingle metric. We can make a statement like, “if potential buyers wouldraise their reservation prices by 20 percent, and sellers would reducetheir reservation prices by the same amount, relative to current averagetransaction prices, then liquidity as indicated by turnover volume wouldrevert to its long-term historical average.”

There is a particularly interesting interpretation of the demand sideindex, the potential buyers’ reservation prices. The price changes indi-cated by this index are those which would, in principle, provide the assetmarket with “constant-liquidity” over time. If potential sellers would,in their own reservation prices, match the changes of potential buyers’reservation prices, then an equal volume of trading (equal turnover ra-tio) would transpire over time, and this would occur at price changestracked by the buyers’ reservation price index. Thus, the demand sidereservation price index may be taken as a “constant-liquidity” price in-dex for the asset market. This indicator expresses prices and liquidityin a single dimension, rather than separately.2

The demand side reservation price index will also provide a leadingindicator of the consummated transaction price index. This is becausein real estate markets, volume tends to slightly lead consummated trans-action prices in time. The reason for this may be that it is difficult toknow the precise implication of relevant “news” for the value of anyspecific property asset. Market participants can only “digest” the im-pact of news gradually by observing the prices at which “comparable”properties have been traded. Because sellers digest the news slower thanbuyers, their reservation prices (hence listing prices) tend to lag behindbuyers’ reservation prices (Genesove and Han, 2012; Carrillo et al., 2015;Van Dijk and Francke, 2018). The result is that transactions tend to goup (down) when markets go up (down) and that prices go up (down)more gradually. Hence market participants may first gain an indica-tion of the “heat” of the market, the relative movement in supply anddemand, by witnessing changes in trading volume, i.e. the number ofdeals completed. Therefore, demand indices can potentially be used as

2The model of buyers’ matching their reservation prices to the sellers’ as if sellerswere in “the drivers seat” would not be consistent with the pro-cyclical relationshipbetween price and volume that is so characteristic of the real estate market. SeeFisher et al. (2003). From a conceptual perspective, “liquidity” is created by buyers,not sellers, as buyers “have the money”, and sellers “want the money”.

59


leading indicators of the real estate market.The supply side reservation price index will also be of interest in its

own right, because it reveals sentiment among property owners. Dur-ing periods of uncertainty, both rational search theory and behavioraleconomic theory predict that property owners would tend to react con-servatively, holding back from selling into a down market. The supplyside index would reveal this phenomenon.

While it would be nice to have separate indices of the reservationprices of potential buyers and sellers in the property asset market, theseprices are private information. There is probably no single perfect way totrack reservation prices. But the fact that both consummated prices andtrading volume (turnover ratio) indirectly reflect the underlying reser-vation prices allows reasonable and useful inferences to be made aboutreservation prices. By modeling both price and volume as functions ofunderlying reservation prices, it is possible to “back out” a reasonablemodel of the demand and supply side reservation prices, in terms ofrelative movements or price changes over time.

The first and most widely employed technique for constructing sep-arate demand and supply reservation price indices in real estate waspioneered by Fisher et al. (2003, 2007).3 Their approach was basedon hedonic price modeling of individual sales observations, and it wasapplied to the National Council of Real Estate Investment Fiduciaries(NCREIF) population of properties. Hedonic price modeling is ham-pered in commercial real estate applications, because of the scarcity ofgood data on the relevant hedonic attributes of the properties, whichtend to be more heterogeneous than houses. Furthermore, the NCREIFpopulation of properties is rather narrow and specialized, consisting ofrelatively large, prime properties owned by cash-rich and tax-exemptedinstitutions.

In the present study we extend the Fisher et al. (2003, 2007) method-ology to a repeat sales indexing framework. This enables us to developdemand and supply reservation price indices for the much larger andbroader population of properties represented by the Real Capital Analyt-ics Inc (RCA) database. This database captures approximately 90 per-cent of all commercial property transactions in the US over $2,500,000

3Goetzmann and Peng (2006) present a slightly different approach that providessimilar results. The Fisher et al. (2003, 2007) methodology was employed by the MITCenter for Real Estate to produce and publish a quarterly-updated set of demandand supply indices during 2006-2011.

60


in value.The main benefit of using the repeat sales methodology is that it

takes care of all unobserved heterogeneity (Bailey et al., 1963), which asnoted is a particularly important consideration in commercial real es-tate. In order to apply the Fisher et al. (2003, 2007) method in a repeatsales framework, we adapt the sample selection methodology for repeatsales models developed by Gatzlaff and Haurin (1997). Perhaps evenmore importantly, our methodology allows us to estimate supply anddemand indices at a much more granular level, in part because it allowsus to use the much larger RCA database. The Fisher et al. (2003, 2007)methodology was limited by the NCREIF database coverage to just es-timating national indices, even though property markets are inherentlylocal in nature. A key innovation in our methodology, which also en-ables granular index production, is our use of a Bayesian, structuraltime series model for index estimation, building on the work of Francke(2010) and Francke et al. (2017). By introducing these features, we areable to estimate reliable, robust supply and demand indices for granularmarkets. In principle, the method can be applied for every repeat salesdata-set with a minimal set of (or even no) property characteristics.

In this study we construct quarterly supply and demand indicesfor every major metropolitan area in the United States from 2005Q1–2017Q2. (The index results are available upon request.) Our historicalperiod includes the Global Financial Crisis (GFC), which had not yetoccurred at the time when Fisher et al. (2003, 2007) published theirwork. The GFC is a particularly interesting period for observing thebehavior of the two sides of the market, how the demand collapsed first,while property owners behaved much more conservative and held on tohigher reservation prices.

In this chapter we focus on the results for the New York City andPhoenix Metro areas (as defined by RCA). These two provide a fascinat-ing “tale of two cities”, as they represent very different underlying urbanform and space markets, and displayed interestingly different responsesduring the GFC. New York is a typical land supply constrained marketwith high historical price growth, whereas Phoenix is characterized bysprawl and high supply elasticity in the space market.

Our results show that the demand indices in both New York andPhoenix went down a full year earlier than the supply indices duringthe crisis. In Phoenix, the demand index also went up first after thetrough of the crisis. In New York, supply and demand bounced back

61


up together. In New York, the supply index, i.e. property owners’reservation prices, hardly dropped even during the worst of the crisis,suggesting an impressive confidence in the NYC real estate market onthe part of property owners in view of a financial crisis that hit NewYork particularly hard. The gap between demand and supply pricemovements was much greater in Phoenix than in New York, resulting ina much greater collapse in trading volume (greater loss of liquidity) inthe Phoenix market.

The remainder of this chapter is organized as follows. Section 3.2introduces our methodology. Section 3.3 gives the data and some de-scriptive statistics. Section 3.4 presents the results. Finally, Section 3.5concludes.

3.2 Reservation Prices and Liquidity

First, we discuss the theory of demand and supply in real estate inSection 3.2.1. Our methodology also follows from this. Section 3.2.2gives the estimation procedures.

3.2.1 Model

The setup of the model is similar to that in Fisher et al. (2003) (hence-forth FGGH): Heterogeneous properties are traded among heterogeneousagents in a search market. Both buyers and sellers set their reservationprices based on property characteristics and the market situation:

RP bi,t = βbt +Xiαb + εbi,t, (3.1)

RP si,t = βst +Xiαs + εsi,t. (3.2)

Here RP is the reservation price and subscripts i and t denote the prop-erty and time period, respectively. Superscripts b and s represent buyersand sellers. X is a property-specific 1 × K vector of property charac-teristics and α is the corresponding K × 1 coefficient vector, and εb, εs,are independent normally distributed error terms with mean zero. βbtand βst are common trends across the reservation prices of all buyers andsellers, respectively. These reflect the market-wide movements of buyersand sellers that we are interested in.

In a search market, we observe a transaction if RP bi,t > RP si,t. Thedistributions of reservation prices of buyers and sellers are displayed in

62


Figure 3.1: Buyers’ and sellers’ reservation price distributions at different points inthe cycle consistent with pro-cyclical liquidity.

P0 Reservation Prices

Normalmark

et

Demand

Supply

P0P1 Reservation Prices

Boom

P0P1P2 Reservation Prices

Bust

63


Figure 3.1. The shaded area is the intersection between the distributionsand depicts transaction volume (a larger area reflects more transactions).The outcome of the transaction price Pi,t depends on both the sellers’and buyers’ bargaining power. We follow Wheaton (1990) and FGGH byassuming that the transaction price is the midpoint between the buyer’and sellers’ reservation price.4 The average midpoint price we observein the market is denoted as P0 in Figure 3.1. In a booming period, thedistribution of the reservation price of buyers moves to the right.5 Thismovement results in more transactions and a higher observed averagemidpoint price, denoted by P1. Finally, in a bust buyers move theirreservation prices to the left. The intersection area will be smaller whichreflects a lower transaction volume and the observed midpoint price willmove down to P2.

By estimating a normal hedonic regression model we are able toestimate the average midpoint price per time period βt = 1

2(βbt + βst ),other coefficients α = 1

2(αb + αs), and residuals εi,t = 12(εbi,t + εsi,t):

E(Pi,t) =1

2(βbt + βst ) +

1

2Xi(α

b + αs)

+1

2E((εbi,t + εsi,t)|RP bi,t ≥ RP si,t),

E(Pi,t) = βt +Xiα+ E(εi,t|RP bi,t ≥ RP si,t).

(3.3)

Let S∗i,t = RP bi,t − RP si,t and substitute Equations (3.1) and (3.2) toobtain:

S∗i,t = RP bi,t −RP si,t = (βbt − βst ) +Xi(αb − αs) + (εbi,t − εsi,t).

(3.4)

Note that S∗i,t is latent, instead we observe Si,t = 1 if a transaction is

consummated. Let γt = βbt−βst , ωj = (αb−αs), and ηi,t = εbi,t−εsi,t. The

4We will not pursue the relationship between bargaining power and liquidity heresince this would result in difficulties of identification of the reservation prices, seeCarrillo (2013) for a discussion on this topic.

5The distribution of sellers might also move, but it is generally accepted thatbuyers respond more quickly than sellers, see Genesove and Han (2012), Carrilloet al. (2015), and Van Dijk and Francke (2018). So in our case the buyer reservationprices move more to the right than seller reservation prices.

64


probability of sale can be estimated by estimating the following probitmodel:

S∗i,t = γt +Xiω + ηi,t, ηi,t ∼ N(0, 1). (3.5)

= Pr(Si,t = 1|Xi) = Φ(γt +Xiω), (3.6)

here Φ is the cumulative density function (CDF) of the normal distri-bution. Note that S includes both a subscript i (property) and t (timeperiod): A property is “tracked” over time. Si,t takes the value 1 whenthe property is sold (once, twice etc.), and 0 when it is not sold in a givenperiod. The γt reflects the shift in the probability of sale in a given timeperiod. Following FGGH, the inverse Mills ratio (λt) is calculated fromthe probit results. The probit estimation only yields the coefficients upto an estimated scale factor, σ: 6

γ = γ/σ = (βbt − βst )/σ, (3.7)

ω = ω/σ = (αb − αs)/σ. (3.8)

Deviating from FGGH, we estimate the transaction price model onlyon repeat sales. This has the advantage that we can take care of allunobserved heterogeneity, which is of importance in the heterogeneouscommercial real estate market. This, however, does require some ad-justments to the FGGH method. In the next steps, we will discuss theseadjustments. We start with the selection corrected repeat sales modelof Gatzlaff and Haurin (1997, henceforth GH). The main goal of GH isto correct for the selection bias of second sales versus first sales. Themain aim of our set-up is to correct for the bias of sales (either first ormultiple) versus no sales. However, with some adjustments we can stillapply the GH-model. The two equations for the first and repeat saleconditional on the sale being observed are:

E(Pi,fir|Si,fir = 1) = βfir +Xiα+ E(εi,fir|Si,fir = 1), (3.9)

= βfir +Xiα+ σ1,3λ1 + σ2,3λ2, (3.10)

6The conditions for the identifiability of this “probit σ” in our context are dis-cussed later on.

65


E(Pi,sec|Si,sec = 1) = βsec +Xiα+ E(εi,sec|Si,sec = 1), (3.11)

= βsec +Xiα+ σ1,4λ1 + σ2,4λ2. (3.12)

Here fir and sec denote the time of first and second sale, respectively.P denotes the (log) transaction price. Note that the equation includestwo selection correction variables: λ1 and λ2. These are the inverse Millsratios for the first and second sale, respectively. In GH these are derivedby estimating a bivariate probit on the probability of a single sale on theone hand and the probability (conditional on the first sale) of a repeatsale on the other hand. The σ coefficients are covariances between theerrors terms of the selection and sale equations. More specifically, thecovariance matrix of the four error terms (ηi,fir, ηi,sec, εi,fir, εi,sec) isdefined as (GH):

Σ =

1 σ1,2 σ1,3 σ1,4

σ1,2 1 σ2,3 σ2,4

σ1,3 σ2,3 σ3,3 σ3,4

σ1,4 σ2,4 σ3,4 σ4,4

. (3.13)

Subtracting (3.10) from (3.12) and adding disturbance terms results inthe repeat sales equation of GH:

Pi,sec−Pi,fir = βsec−βfir+(σ1,4−σ1,3)λ1 +(σ2,4−σ2,3)λ2 +υi. (3.14)

By taking the exponent of β, we get a price index. The correlations thatappear in Equation 3.14 are σ1,3, which is the covariance between the er-ror terms of the first selection equation (ηi,fir) and the first sale equation(εi,fir), σ1,4, which is is the covariance between the error terms of thefirst selection equation (ηi,fir) and the second sale equation (εi,sec), σ2,3,which is the covariance between the error terms of the second selectionequation (ηi,sec) and the first sale equation (εi,fir), and σ2,4, which is,the covariance between the error terms of the second selection equation(ηi,sec) and the second sale equation (εi,sec).

In order to estimate constant-liquidity indices, it is necessary to in-clude a time trend in the probit (e.g. by including time fixed effectsin the probit). It is not clear whether this should be included in theprobit on first sales, on repeat sales, or on both (the original GH model

66


does not include any time fixed effects). Furthermore, it is not straight-forward which λ (of the first or repeat sale) should be used to derivethe supply and demand indices in a later stage (see Equations 3.22 and3.23). Therefore, our proposition is to estimate the following repeatsales equation:

Pi,sec − Pi,fir = βfir − βsec + σε,η(λ2 − λ1) + υi, υi ∼ N(0, σ2υ).(3.15)

This does impose restrictions on the coefficients of the selection cor-rection variables λ. More specifically, (i) σ2,3 = σ1,4 = 0 and (ii)σ2,4 = σ1,3 = σε,η. (i) implies that there is no correlation between theerror terms of the first selection (sale) equation and the second sale (se-lection) equation. This means that the reason to buy cannot be relatedto the realized gain or loss in the future. Therefore, it is important to re-move properties built for redevelopment from the data. (ii) implies thatthe correlation between the first sale and the first selection is equal tothe correlation of the second sale and the second selection. This meansthat the decision to buy or sell must be “normal”. For example, if a thesecond sale was in distress, and the first one is not, this assumption isviolated. Therefore, properties sold in distress need to be removed.

In GH, the selection equations were estimated in a bivariate probitand these assumptions were not necessary. Since the selectivity in arepeat sales model (i.e. the selection effects of second sales versus singlesales) was the main goal of GH, these correlations needed to be explicitlymodeled. In our set-up, we are merely interested in the selection effectof no sale versus a sale, either a first or multiple. Hence, these assump-tions are not very restrictive in our case. Moreover, these assumptionsare also implicitly made for observations that are sold more than oncewhen deriving supply and demand indices in the widely applied hedonicframework of FGGH. Finally, as mentioned, one can account partly forthese factors by applying the appropriate data filters.

In order to apply the method in thin markets, the time trend in therepeat sales equation is modeled as a random walk with an additionalautoregressive parameter (ρ):

∆βt = ρ∆βt−1 + ξt, ξt ∼ N(0,σ2ξ

1− ρ2). (3.16)

67


The identifiability of the “probit σ” for a hedonic framework is discussedin the Appendix of FGGH and will also be briefly discussed here. Letσ2s = Var(εsi,t), σ

2b = Var(εbi,t), and σs,b = Cov(εsi,t, ε

bi,t). Note that these

are the (co)variances of equations (3.1) and (3.2). The scale parameter,the “probit σ”, is equal to σ2 = σ2

s + σ2b − 2σs,b, which is what we need

to solve for. Following FGGH, our model assumes random matchingbetween buyers and sellers and we can assume Cov(εbi,t, ε

si,t) = 0. This

simplifies the expression: σ2 = σ2s + σ2

b . The conditional expected vari-ance of the pricing errors (ε2

i,t) in a hedonic model follows from the knownrelations for the moments of the truncated bivariate normal distribution(Johnson and Kotz, 1972):

E(ε2i,t|Si,t = 1) = σ2

ε − σ2ε,η(γt +Xiω)λi,t, (3.17)

where σ2ε = Var((εbi,t + εsi,t)/2) = (σ2

b + σ2s)/4 = σ2/4. (3.18)

As expectation for the squared errors we plug in the squared residualsof the repeat sales model ε2

i,t. Note that FGGH use the sum of squaredresiduals (SSR) from a hedonic model. However, the SSR from a hedonicmodel with pair fixed effects is equivalent to the SSR of our repeat salesmodel.7 σ2

ε,η is the square of the estimated coefficient on the inverseMills ratio, γt +Xiω is the linear prediction from the probit model, andλi,t is the estimated inverse Mills ratio.8 To solve for σ2

ε we further plug

in the estimates of the probit (γt, ω, and λi,t) and repeat sales estimates(ε2i,t and σ2

ε,η):

σ2ε = (1/N)

N∑i=1

[ε2i,t + σ2

ε,η(γt +Xiω)λi,t

], (3.19)

σ = 2σε. (3.20)

7The SSR and MSE are equivalent, but the individual squared errors are differ-ent. However, since we are eventually interested in the MSE, we can safely assumeE(ε2i,t|Si,t = 1) = 1

2E(ε2i,fir|Si,t = 1) + 1

2E(ε2i,sec|Si,t = 1), where fir is the first sale

and sec is the second sale.8Note that we estimate the coefficient on the difference of the inverse Mills ratio.

However, the restriction σ24 = σ13 = σε,η implies that the coefficient on the difference(σε,η) is the same as the coefficient on the level of the inverse Mills ratio of the firstand second sale.

68


Here N is the number of observations from a hedonic model, so in ourcase that would be the number of repeat sales times 2.9 Also note thatwe stack the repeat sales data: every pair gets two observations (buyand sell). In this case we also use the level of the inverse Mills ratio ofthe first and second sale instead of the difference.

In order to derive the buyers’ and sellers’ reservation price indices,we combine the probit and repeat sales model results to obtain thedemand/supply indices (FGGH). As estimated values we have: γ =

(βbt − βst )/σ (from equation 3.7) and βt = 12(βbt + βst ) (from equation

3.3) → βst = 2βt = βbt . Substituting the latter in the former, we get thebuyers’ reservation (constant-liquidity) price index:

γ = (βbt − 2βt − βbt )σ, (3.21)

βbt = βt +1

2σγt. (3.22)

Further substitution yields the sellers’ reservation price index:

βst = βt −1

2σγt. (3.23)

3.2.2 Estimation

We estimate the model in a two-step approach (Heckman, 1979). Wefirst estimate probit Equation (3.6) by maximum likelihood. We subse-quently calculate the inverse Mills ratios of the first and second sales.We then plug in the difference between the inverse Mills ratios in therepeat sales model as given by Equation (3.15). We estimate the repeatsales model in a Bayesian framework similar to the Commercial PropertyPrice Indices published by RCA (Francke et al., 2017). The differenceis that we use normally distributed errors instead of t-distributed errorsin order to use the multivariate normality with the probit. Also, we donot allow the signal and noise to vary over time for the sake of simplicityand consistency with the theory.

As explained in Francke and van de Minne (2017), estimating repeatsales models in a structural time series framework, especially with AR

9When a property has more than 2 sales, for example 3 sales, this would result in4 observations in the hedonic model with 2 pair fixed effects. Hence the second saleenters twice.

69


components, is very difficult using the Kalman filter. We therefore esti-mate the model with Markov Chained Monte Carlo simulations. Morespecifically, we use the No-U-Turn-Sampler (NUTS) to estimate the in-dices (Hoffman and Gelman, 2014; Carpenter et al., 2016). The repeatsales indices are estimated over 4 parallel chains with different initialvalues. We use 4,000 iterations per chain of which 2,000 are warm-upiterations that we discard. Hence, the total sample size is 8,000. Fol-lowing Francke et al. (2017), we use the R combined with the effectivesample size to determine the convergence of the model. We additionallyinvestigate the Monte Carlo error and the Heidelberger-Welch station-arity and halfwidth statistics (Koehler et al., 2009; Heidelberger andWelch, 1981).

3.3 Data

We use transaction data from Real Capital Analytics (RCA) to esti-mate the indices between 2005Q1 and 2017Q2. RCA has captures morethan 90% of the properties with a value larger than $2,500,000. A highcapture rate is fundamentally important for our methodology since theprobit results should reflect the probability of sale with respect to thewhole population. In order to capture the whole population of prop-erties each property should be sold at least once to be included in thedatabase. We are confident that the high capture rate combined withthe fact that RCA has been recording transactions since 2000 ensuresthat almost all properties in the investment universe are in our data.Besides, properties that are not in the data after 17 years might neverbecome part of the investment universe that investors are interested in(i.e. properties that trade on a regular basis).

The descriptive statistics for the two metropolitan markets (NewYork City, NY and Phoenix, AZ) are presented in Table 3.1. The num-ber of properties is more or less constant in the data, which providesconfidence that the capture rate of the data is satisfactory. We are notable to identify whether a property gets demolished. We do observe ifa property is bought for the purpose of redevelopment. These proper-ties are removed for the reasons outlined in section 3.2.1. For the samereasons, properties sold in distress are removed.

In the probit equation, we control for property size (in log squarefeet), property type (Apartment, Hotel, Industrial, Office, Retail, and

70


Other), whether the property is in the Central Business District (CBD),and construction period (< 1920, 1920 − 1945, 1946 − 1989, and ≥1990). We also use the construction year to filter out properties thatwhere not yet built. Since properties sometimes transact before theconstruction has been finished, we include the property in the probitdata two years before completion. If the model is estimated withoutproperty characteristics, this provides comparable results (not shown).

In both New York and Phoenix, the crisis is clearly visible in boththe number of sales and average transaction price. In 2009 the numberof transactions is about 50% lower than in 2008 and 70-80% lower thanin 2007. The average property size is more or less equal over the sam-ple. Most properties sold in Phoenix are outside of the CBD, whereasin New York most are within the CBD. Properties in New York are, onaverage, older than in Phoenix and the average construction year in-creases somewhat over the sample as newly constructed properties enterthe data-set.

71


Table 3.1: Descriptive statistics of the two markets over the sample.

Properties Price SizeYear Sales in data MLN $ 1000 Sqft CBD Year built

New York City, NY

2005 1,948 17,867 18.7 35.8 0.73 19412006 2,215 17,993 21.5 32.4 0.74 19412007 2,248 18,062 23.8 30.2 0.75 19402008 1,313 18,127 18.8 28.1 0.71 19422009 690 18,174 9.9 25.1 0.72 19422010 852 18,227 18.3 34.1 0.73 19432011 1,090 18,265 23.9 30.9 0.75 19442012 1,790 18,298 18.9 24.8 0.83 19382013 2,218 18,334 22.1 25.1 0.78 19422014 2,423 18,358 24.2 25.7 0.79 19412015 2,715 18,366 25.0 23.4 0.77 19422016 2,253 18,370 27.0 25.3 0.74 19442017* 951 18,371 20.2 23.7 0.67 1946

Phoenix, AZ2005 884 4,548 11.5 61.2 0.01 19872006 801 4,720 12.6 57.4 0.01 19892007 696 4,777 13.6 53.1 0.03 19902008 288 4,794 10.4 41.3 0.01 19902009 130 4,801 9.1 51.0 0.02 19932010 166 4,823 10.4 54.7 0.00 19932011 255 4,847 12.3 70.3 0.01 19922012 433 4,876 11.6 66.8 0.02 19912013 449 4,908 10.9 58.5 0.00 19922014 610 4,942 11.3 54.9 0.03 19912015 708 4,948 13.4 55.1 0.02 19912016 725 4,948 13.6 53.7 0.02 19912017* 317 4,948 12.3 52.8 0.03 1990

*2017 only consists of data for the first two quarters. Sale price, size, CBD,and year built are sample means per year.

3.4 Results

As noted before, our estimation procedure consists of multiple steps.First, we estimate the probability of sale by equation (3.6) and calcu-late the inverse Mills ratio. Second, we estimate a repeat sales modelincluding the difference in the inverse Mills ratio by estimating equation(3.15). Finally, we derive the supply and demand reservation price in-dices by combining the probit and repeat sales results using equations(3.22) and (3.23).

72


3.4.1 Probability of sale

The estimates for the probability of sale for the two markets are shownin Table 3.2. The estimates of the time fixed effects (γt in equation 3.6)are shown in Figure 3.2. The magnitude of the coefficients cannot beinterpreted directly, but the sign can be interpreted. Furthermore, therelative effects of the categorical variables can be interpreted.10

In general, larger properties have a somewhat higher probability ofsale, although the effect is only marginally significant. In New York,properties located in the CBD sell quicker. With respect to propertytypes, apartments sell the quickest in both markets and “other” (i.e.elderly homes, nursing cares, parking facilities etc.) sell the slowest.Hotel, industrial and office buildings sell roughly at the same speed. Thenewest buildings (built after 1990) sell the quickest in both markets.

The estimates of the time fixed effects are somewhat noisy (Figure3.2). Also, the estimates show a seasonal pattern. Therefore we followFisher et al. (2007) and smooth and seasonally adjust these coefficients.More specifically, we use a local linear trend model with a seasonal factorto distinguish between signal and noise. The probability of sale in NewYork is approximately 50% lower mid 2009 than at the end of 2007 andhas recovered gradually to the pre-crisis level between 2010 and 2015.Since 2015, the probability of sale seems to be somewhat decreasing.

The probability of sale in Phoenix decreased by more than 80% dur-ing the crisis. The probability of sale increased between 2010 and 2016,but has never recovered fully to the pre-crisis level. Also, starting in2016 there is a small decrease in the probability of sale in Phoenix. Fi-nally, notice the very similar development of the rate of sales in the twomarkets. This indicates that there are small differences in regional dy-namics in the probability of sale.11 The development of the probabilityof sale actually shows a very similar pattern across different markets inthe US (not shown here, but available upon request). This does, how-ever, not imply that the probability of sale is similar across differentmarkets. The baseline level of sales, for example, could be substantiallydifferent.

10Another option would be to present marginal effects. The supply and demandindices, however, require the “raw” coefficients. Therefore, we choose to present thecoefficients instead.

11The correlation between the probability of sale in New York and Phoenix is 0.92in (index)levels and 0.66 in first differences.

73

CH

AP

TE

R3.

RE

VIS

ITIN

GS

UP

PLY

AN

DD

EM

AN

DIN

DIC

ES

INR

EA

LE

ST

AT

E

Table 3.2: Coefficient estimates including 95% confidence intervals of the probit equation for the two markets over 2005Q1 and2017Q2.

New York City, NY Phoenix, AZVariable ω P2.5 P97.5 ω P2.5 P97.5

Constant -1.971 -2.039 -1.903 -1.858 -2.227 -1.488Log(Size) 0.004 -0.001 0.009 0.006 -0.004 0.017CBD 0.036 0.020 0.052 0.022 -0.063 0.107Apartment (Omitted) (Omitted)Hotel -0.025 -0.077 0.027 -0.041 -0.105 0.024Industrial -0.075 -0.093 -0.057 -0.082 -0.115 -0.050Office -0.057 -0.074 -0.041 -0.072 -0.105 -0.039Retail -0.055 -0.071 -0.039 -0.083 -0.118 -0.049Other -0.079 -0.127 -0.032 -0.145 -0.210 -0.079Built <1920 (Omitted) (Omitted)Built 1920-1945 0.001 -0.014 0.016 0.137 -0.233 0.507Built 1946-1989 -0.009 -0.027 0.010 0.145 -0.198 0.489Built ≥ 1990 0.035 0.013 0.057 0.168 -0.175 0.512

Time fixed effects Yes (γt) Yes (γt)N 922450 244607Loglike -104607.6 -28847.5

74


Figure 3.2: Raw and smoothed/seasonally adjusted estimates of the time fixed effectsin the probit equation (γt) for the two markets over 2005Q1 and 2017Q2.

2005

200

6

2007

2008

2009

2010

2011

2012

2013

2014

2015

201

6

201

7

−1

−0.8

−0.6

−0.4

−0.2

0

0.2

0.4

γt(log

scale)

NYC = NYC (Seasonally adjusted and smoothed)

PHX PHX (Seasonally adjusted and smoothed)

3.4.2 Repeat sales indices

The repeat sales indices for New York and Phoenix are shown in Figure3.3. These are the “midpoint-indices” that follow from the estimationof equation (3.15). The indices are, by construction, on the midpointbetween the demand and supply indices. The GFC is clearly visiblein the indices. New York started recovering in 2010 and the pricessurpassed the pre-crisis level in 2013. In contrast, Phoenix has beenrecovering much more slowly, starting in 2011. Substantial recovery inPhoenix took off not before 2016 when prices increased by almost 25%.The price levels at the end of the sample, however, are still lower thanthe pre-crisis ones.

The model statistics and the estimate of the coefficient on the dif-ference of the inverse Mills ratio are presented in Table 3.3. All modelstatistics are satisfactory, the Monte Carlo error (MC error) is very closeto 0 and the effective sample size (Neff) relative to the number of sam-ples is close to 1 for all models (note that the number of samples is8,000 for every model). The Heidelberger stationary statistic is equal to1, indicating that the test is passed by every parameter in the model (aparameter gets the value 1 if the test is passed). The same holds for the

75


Heidelberger halfwidth statistic, which is close to 1. This indicates thatthe number of chains is high enough. Finally, the R values close to 1indicate that the models have converged well.

Figure 3.3: Repeat sales indices of commercial real estate in New York and Phoenixbetween 2005Q1 and 2017Q2.

2005

2006

2007

200

8

2009

2010

2011

2012

2013

201

4

201

5

201

6

2017

100

150

200

250

300

Index

(2005Q1=100)

(i) NYC (ii) PHX

76


Table 3.3: Coefficient estimates on the inverse Mills ratio (σε,η) in the repeat salesequations including the 95% HPD intervals and model statistics.

Statistic New York City, NY Phoenix, AZσε,η -0.1617 -0.3095σε,η,p2.5 -0.3843 -0.5533σε,η,p97.5 0.0549 -0.0735Repeat sales 4106 1452Loglike -2394.1 -777.2MC-error 0.0011 0.0024Neff 7986 7993Heidelberg stationarity 1.0000 1.0000Heidelberg halfwidth 0.9889 0.9783

R 1.0000 0.9999

Repeat sales denotes the number of repeat sales pairs, Loglike thelog-likelihood, MC-error the mean of Monte Carlo standard er-ror for all parameters, Neff the mean effective sample size (totalnumber of samples = 8000) of all parameters, Heidelberger-Welchstationarity and the Heidelberger halfwidth tests are the mean ofthe tests of all parameters (a parameter gets the value 1 (0) whenthe test is passed (failed) at the 5% level). Finally, R is the meanRhat statistic for all parameters.

3.4.3 Supply and demand indices

The indices for demand and supply reservation prices that follow fromequations (3.22) and (3.23) are shown in Figures 3.4 and 3.5 (mind thedifference in scale). Following Fisher et al. (2007), the indices are pre-sented such that the mean of the supply and demand indices are equal.The indices can also be used as a liquidity indicator; if the difference be-tween demand and supply becomes smaller (larger) there is less (more)liquidity in the market.

In both markets, the demand-side (or constant-liquidity) indicesseem to move quicker and more extreme than the supply side indices.The demand side indices are decreasing a year before the supply side in-dices during the crisis. Especially in Phoenix, the recovery also happensmuch earlier. Comparing the demand side indices from Figures 3.4 and3.5 and the midpoint indices from Figure 3.3, we observe that the de-mand side constant-liquidity indices also lead the midpoint indices. Thelead is smaller than the full year in the supply and demand indices, butit is still visible. A Granger causality analysis confirms this: Changes inthe demand side indices Granger cause changes in the midpoint indicesin both markets. The same holds for the relationship between the de-

77


mand side and supply side indices: Changes in supply reservation pricesare Granger caused by changes in demand reservation prices at the 1%level.12

In New York, suppliers’ reservation prices remained surprisingly con-stant which suggests that property owners remained very confident aboutthe New York commercial real estate market. The difference betweenpeak (2008Q3) and through (2010Q2) was only 17%, compared to 29%in the demand side index (2007Q3–2010Q2). After the crisis, supplyand demand show a joint recovery in New York. In Phoenix, suppliers’reservation remained sluggish through as late as 2015Q4. The differ-ence between peak and through is also much larger in Phoenix: Sellers’reservation prices went down by more than 40% (2008Q4–2012Q4) andbuyers’ reservation went down by 46% (2008Q1–2010Q2).

The more extreme movents of the demand side indices is also doc-umented in Fisher et al. (2007) and is consistent with the notion ofpro-cyclical liquidity. Liquidity, as indicated by the difference betweendemand and supply reservation prices, substantially decreased duringthe GFC. This is also clearly visible in the probability of sale (Figure3.2).

Figure 3.4: Supply and demand indices in the New York City area.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

300

Index

(Demand

2005Q1=100)

(i) Demand (ii) Supply

12Both Granger causality analyses are significant at the 1% level and are based ona VAR model with 4 lags estimated separately for each market.

78


Figure 3.5: Supply and demand indices in the Phoenix area.

2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

201

6

2017

60

80

100

120

140

160

Index

(Demand

2005Q1=100)

(i) Demand (ii) Supply

3.4.4 A recent anomaly?

Recently, a rare phenonemon is visible in the commercial real estatemarkets, as prices keep steady or increase, while deal volume is goingdown. Usually, liquidity is pro-cyclical: liquidity goes up when prices goup and vice versa. RCA suggests that a difference in buyer and sellerreservation prices is the cause: “The current impasse suggests a widen-ing gap between buyer and seller pricing expectations” (Real CapitalAnalytics, 2017). This implies that the distribution of sellers’ reser-vation prices moves more to the right than the distribution of buyers’reservation prices. This results in fewer transactions and a decrease inliquidity. Prices, however, increase. This is shown in Figure 3.6, whichdepicts the cumulative distributions in the reservation prices. Our em-pirical findings for both Phoenix and New York are consistent with thisview. The decrease in liquidity is visible in Figure 3.2 and the increasein (midpoint) prices is visible in Figure 3.3. These dynamics are alsovisible in the supply and demand indices, especially in New York, wherethe reservation prices of sellers are increasing more than the reservationprices of buyers in the final year (Figure 3.4).

79


Figure 3.6: Buyers’and sellers’ cumulative distribution of reservation prices in case ofpro-cyclical liquidity.

P0

Q0

Number of Agents

Reservation Prices

Demand

Supply

P0 P1

Q0

Q1

Number of Agents

Reservation Prices

3.5 Conclusions

In this chapter we extent the constant-liquidity index methodology in-troduced by Fisher et al. (2003). We introduce two new components: (i)We cast the method in a repeat sales framework, and (ii) we estimate themodel in a structural time series format. As a result, we can disentan-gle reservation prices of buyers and sellers for commercial real estate atthe city level without needing a substantial set of property characteris-

80


tics. We apply our model using data provided by Real Capital Analytics(RCA). RCA tracks properties over $2,500,000, or the typical institu-tional / large private investor space. In this chapter we focus on twocities in a very different urban setting: New York and Phoenix.

Tracking demand and supply separately does not only give us moreinsight into the real estate market, it can also be used as a predictivemodel. Indeed, supply tends to move slower than demand, due to an-choring and loss aversion (Bokhari and Geltner, 2011) and issues relatedto mortgage debt (Genesove and Mayer, 2001). The “midpoint” indexthat is usually estimated in a repeat sales framework also lags behindreservation prices of buyers. We find that in both New York and Phoenixdemand dropped a full year prior to supply during the crisis. We fur-ther find that buyers’ reservation prices went down by 29% and 46% inNew York and Phoenix, respectively. In New York, reservation prices ofsellers only dropped moderately by 17% compared to the large drop inPhoenix of 40%.

Finally, we document a rare phenomenon. In all markets and formost of the time, liquidity is pro-cyclical. In recent quarters, however,prices seem to be increasing and liquidity is decreasing. This is alsovisible in our demand and supply indices: Reservation prices of sellersare increasing more than reservation prices of buyers.

Although we only analyzed the indices for New York and Phoenix,indices for all major metropolitan areas are available upon request. Weencourage future research to examine more markets in different urbansettings and to answer additional questions related to reservation pricedynamics. Also, our method is general and can be applied to every dataset in which repeat sales can be identified and for which transactionprices are known.

81


82

Chapter 4Internet search behavior, liquidity,and prices in the housing market

It is tough to make predictions,especially about the future.

Yogi Berra

We employ detailed internet search data to examine price and liquid-ity dynamics of the Dutch housing market. We show that the number ofclicks on properties listed online proxies demand and the number of listedproperties proxies supply. From this internet search behavior we create amarket tightness indicator and we find that this indicator Granger causeschanges in both house prices and housing market liquidity. The resultsof a panel VAR suggest that a demand shock results in a temporary in-crease in liquidity and a permanent increase in prices in urban areas.This is in accordance with search and matching models.1

1This chapter is based on Van Dijk and Francke (2018). Acknowledgments: Theauthors would like to thank both Funda and the NVM for supplying the data. Ourspecial thanks go to Ruben Scholten for collecting and organizing the Funda data.We are also grateful to seminar participants of the 2015 Global Real Estate Sum-mit, the NVM research department, De Nederlandsche Bank, and Ortec Finance forproviding useful comments. Finally, we would like to thank Martijn Droes, Jakob deHaan, Irma Hindrayanto, Ide Kearney, Alex van de Minne, Rob Sperna Weiland, Si-mon Stevenson, Patrick Tuijp, and an anonymous referee for providing many helpfulsuggestions.

83

CHAPTER 4. INTERNET SEARCH BEHAVIOR, LIQUIDITY,AND PRICES IN THE HOUSING MARKET

4.1 Introduction

The internet proves to be a valuable source of information that foreshad-ows economic developments. The basic idea is that future consumptionis preceded by information gathering. Askitas and Zimmerman (2009)call this behavior “preparatory steps to spend”. In line with Wu andBrynjolfsson (2014), this chapter employs internet search data to exam-ine spatiotemporal housing market dynamics. We argue that potentialhome buyers start their search for a house by browsing the internet.The availability of detailed data in the Netherlands allows us to exam-ine the relationship between online search behavior and housing marketdevelopments on a local scale.

The largest housing website in the Netherlands is Funda.nl, whichhas a stable market share of around 60% of all housing websites (Ker-ste et al., 2012). Furthermore, 83% of potential buyers use Funda tofind a suitable home (Conclusr, 2014). Therefore, the activity on thiswebsite, more specifically search behavior, could give a useful indicationof current or future demand. Moreover, the number of listed propertiescould be a useful supply indicator. By combining these data on demandand supply, we develop an indicator of market tightness. The advan-tage of these data is that they provide a detailed panel, both over time(quarterly) and over the cross-section (municipalities).

Other papers like Carrillo et al. (2015) provide empirical evidencefor the relationship between measures of liquidity (sale probability andsellers’ bargaining power) and subsequent price appreciation. Our keycontribution is that we introduce a novel measure of market tightness.Using this measure, we are able to provide empirical evidence for therelationship between market tightness on the one hand and liquidity andprices on the other hand. This is important because it empirically val-idates the use of theoretical search and matching models. Our secondcontribution is from a practitioners’ point of view. We provide evidencethat internet data can be used as a leading indicator of prices and liq-uidity. As such this could potentially be of interest for forecasters inthe real estate market who wish to incorporate consumer sentiment inforecasting models.

Our main three research questions are: (1) How does market tight-ness influence market liquidity, (2) how does market tightness influencehouse prices, and (3) what are the differences in temporal dynamics?In our view, the current literature cannot answer these questions ade-

84


quately since there are no studies, as far as we are aware, that combinemarket tightness, market liquidity, and prices in a single empirical frame-work. Furthermore, we are also able to account for local housing marketcharacteristics as we employ data on a detailed geographical scale.

We approach these questions as follows. First, we employ a the-oretical search and match framework that includes market tightness,market liquidity, and prices. We generalize the framework of Carrilloet al. (2015) to allow for different numbers of sellers in each period.Second, we provide empirical evidence for these theoretical channels byestimating a panel VAR model that includes these variables.

Market tightness is estimated as follows. Funda provides number ofclicks per month for each individual house listed online. Due to privacyissues, the lowest level at which the internet search data can be linkedto transaction data is the ZIP code level. Unfortunately, the resultingnumber of observations per ZIP code level is too small to generate re-liable high-frequency series. Therefore, we aggregate the data at themunicipal and quarterly level. The result is a quarterly panel of all 403Dutch municipalities that contains (1) the number of houses which arefor sale on Funda, and (2) the number of clicks on these houses in thecorresponding quarter. By dividing the number clicks by the numberof online listed houses, we generate a measure of demand versus supply(i.e. a market tightness indicator). We denote this variable clicks perhouse or, in short, cph.

Following Genesove and Mayer (2001) and De Wit et al. (2013),among others, we employ the rate of sale (i.e. sales in a given perioddivided by the number of houses for sale at the beginning of the period)as a market liquidity measure. We denote this variable the rate of saleor ros in short. The rate of sale can be seen as an ex post matchingprobability. We construct quarterly house price indices at the munici-pal level using detailed transaction data based on the methodology ofFrancke and De Vos (2000). We denote this variable pr in short.

Our study is related to literature that examines the relationship be-tween prices and liquidity. Booming housing markets are typically char-acterized by more liquidity, while bust markets with declining prices usu-ally show less liquidity (Stein, 1995; Clayton et al., 2010; De Wit et al.,2013). A large part of the literature that examines the relationship be-tween house prices and market liquidity uses shocks in, for example, thelabor market (Clayton et al., 2010) or the mortgage rate (Hort, 2000;De Wit et al., 2013) to determine the price-volume correlation. These

85


shocks can be either demand or supply shocks. We propose a measurefor market tightness (demand versus supply) based on internet searchbehavior. Whereas there are other sources that provide either local orfrequent data, a major advantage of using internet search data is thatshocks can be measured both locally and frequently. Furthermore, thedata also directly measure consumer sentiment, which until recently wasdifficult to obtain at this granular level at this frequency.

To a lesser extent, our work is also related to the use of big datain real estate research. For example, Wu and Deng (2015) use inter-net search data from Google to detect information flows regarding pricediscovery from larger cities to smaller cities. Lee and Mori (2016) em-ploy data from Google Insights to examine housing premiums for luxuryhomes based on consumers’ preferences for non-housing luxury goods.Finally, Wu and Brynjolfsson (2014) use internet search data to forecasttransactions and price developments in the housing market. As far aswe are aware, this study is the first to use internet data to generate amarket tightness indicator.

To answer the three questions, we estimate impulse response func-tions based on a panel VAR. We find that (1) market liquidity increaseswhen market tightness increases, (2) an increase in market tightnessresults in an increase in prices, and (3) the effect on market liquidityis temporary, while the effect on prices is permanent. These empiricalfindings are in line with the theoretical search and matching frameworkin which buyers respond more quickly to a demand shock than sellersdo (Genesove and Han, 2012; Carrillo et al., 2015).

However, we find geographical differences in these dynamics. Forboth urban and rural areas our results suggest that the permanent effecton prices is positive. For urban areas, we find that liquidity increasestemporary, but not in rural areas. This may reflect that during ourrelatively short sample period, house prices in rural areas did not recoverfrom the declining trend. A further subdivision of the sample accordingto price trends confirms this: The effect of market tightness is morepronounced in regions in which prices started recovering in the finalyear of our sample than in regions with the largest negative returns. Inthe latter regions, the number of transactions remained relatively low,and therefore price discovery is slower in these regions.

In the next section we discuss some of the literature on the rela-tionship between market tightness, prices, and liquidity in the housingmarket; subsequently we offer a theoretical search and matching model.

86


Next, we describe the data used and the econometric model, after whichwe discuss the results and several robustness checks in which we devoteattention to geographic variation in responsiveness.

4.2 Related literature

4.2.1 Market liquidity

The financial economics literature generally distinguishes between mar-ket liquidity and funding liquidity (Brunnermeier and Pedersen, 2009).Market liquidity is defined as the ease at which assets can be traded,while funding liquidity refers to the ease with which they can be fi-nanced. In the housing market literature, examples of market liquidityinclude the rate of sale (Genesove and Mayer, 2001; Hort, 2000; De Witet al., 2013), (seller) time on market (Jud et al., 1996; Kang and Gard-ner, 1989; Glower et al., 1998), and the number of transactions (Wu andBrynjolfsson, 2014). An example of funding liquidity is the ease withwhich a mortgage can be obtained (i.e. credit constraints, see Ducaet al., 2011; Mian and Sufi, 2009; Francke et al., 2014.). In this studywe are interested in market liquidity and we employ the rate of sale asour measure. The rate of sale is defined as the number of transactionsin a given period divided by the number of houses on the market at thebeginning of that period. Hence, the rate of sale can be seen as an expost sale probability.

4.2.2 Search and matching models

De Wit et al. (2013) identify three groups of theories in the literaturethat link prices and liquidity: (i) Search and matching models, (ii) the in-teraction between downpayment constraints, mobility and house prices,and (iii) behavioral explanations. The authors stress, however, that thethree approaches are not mutually exclusive.

The theoretical link between market tightness, liquidity, and pricesis embedded in the search literature. Since housing markets are not per-fectly efficient (Case and Shiller, 1990) and no central housing exchangeexists, the housing market can be characterized as a search market, inwhich buyers and sellers look for each other until they are matched(Wheaton, 1990). If there is a match, a trade will occur which meansa house will be transacted. The chances of getting a match depend on

87


the ratio between buyers and sellers (i.e. market tightness). An increasein this ratio will result in a higher matching probability (i.e. rate ofsale). As sellers observe this increase, they will increase their reserva-tion prices, and prices will go up (Genesove and Han, 2012). For a moreelaborate discussion on search and matching models see the theoreticalframework in the next section.

4.2.3 Downpayment constraints

The fundamentals of the downpayment constraints lie within the workof Stein (1995), who introduced the downpayment hypothesis. Accord-ing to this hypothesis, homeowners who would like to buy a house areconstrained by a downpayment that they have to make in order to buythe new house. We expect that the downpayment hypothesis is not veryapplicable to the Dutch situation as the current LTV limit is relativelyhigh.2

4.2.4 Behavioral explanations

Finally, there are behavioral explanations for the relationship betweenprices and liquidity. Loss aversion is generally thought to hold in bothcommercial (Bokhari and Geltner, 2011) and residential (Genesove andMayer, 2001) markets. As prices go down, homeowners are reluctantto sell their houses for a sum below the amount they paid. Van derCruijsen et al. (2018) show that the principle of loss aversion leads to anoverestimation of the value of the house by the homeowner. The resultis that reservation prices of sellers and asking prices remain too high inbad times. As a consequence, market liquidity will dry up.

2Although the LTV limit will be lowered to 100% by 2018, it is still high comparedto other countries (Almeida et al., 2006) An exception might be current homeownerswho are underwater and are constrained by negative equity in buying their next home.

88


4.3 Theoretical framework

We first briefly discuss a theoretical search and matching model. Thiscan help to put our empirical findings in perspective and to examine howmarket tightness leads to subsequent changes in prices and liquidity. Themodel also provides insights into differences in the dynamics of markettightness, prices, and market liquidity.

4.3.1 Market tightness and prices

The relationship between market tightness (i.e. the ratio of buyers tosellers) and subsequent price appreciation is set out by Carrillo et al.(2015). They argue that the ex ante sale probability (based on the timeon market) and the sellers’ bargaining power (based on list price, saleprice and time on market) respond to changes in market tightness. Sim-ilar to Carrillo et al. (2015), we examine the relationship between mar-ket tightness and price appreciation. However, whereas Carrillo et al.(2015) provide empirical evidence for the relationship between ex antesale probability, bargaining power, and price changes, we propose a di-rect measure for market tightness based on internet search data. Con-sequently, we seek to provide empirical evidence for the link betweenchanges in market tightness, changes in sale probability, and changes inprices.

The model is a standard search and matching model with informa-tion asymmetries and is based on Carrillo et al. (2015) and Novy-Marx(2009). We have generalized the model with respect to the number ofactive sellers per period, allowing for different numbers of sellers in eachperiod. The full model is included in the Appendix; here we focus onthe economic interpretation of the model.

We first define the ratio of buyers to sellers as market tightness indi-cated by λ. Because of information asymmetries, buyers and sellers reactdifferently to a shock in market tightness. Buyers and sellers set theirreservation prices at which they are willing to buy or sell (Yavas andYang, 1995; Knight, 2002). If demand increases, the number of buyerswilling to pay the sellers’ reservation prices increases. Hence, the prob-ability that a transaction occurs increases. Genesove and Han (2012)show that sellers react to a demand shock with a lag. In other words,sellers gradually adjust their reservation prices upwards (downwards)when demand increases (decreases).

89


4.3.2 Market tightness and liquidity

If market tightness (λ) increases, agents will gradually adjust their per-ception regarding market tightness (λ∗) towards true market tightness.Sellers set their reservation prices ε based on their perception of markettightness. Prices change gradually since they are dependent on perceivedmarket tightness only. As the matching probability ω(λ, λ∗) dependsboth on true and perceived market tightness, market liquidity respondsdirectly to an increase in market tightness.

4.3.3 Temporal differences

To determine the effect of an increase in market tightness on prices andmarket liquidity, we increase λ and simulate the model. The calibratedparameters are presented in the Appendix. Figure 4.1 shows the resultsof the simulation. Here we increase market tightness from 1 to 2 in week11. In panel (A) the process of adapting expectations regarding λ isshown, panel (B) includes the effect on the threshold reservation price(ε), and panel (C) includes the effect on the sellers’ matching probability.We observe that after a shock in λ the matching probability increasesimmediately. When sellers gradually learn about the situation, pricesstart increasing, and the matching probability goes down again. Hence,the model predicts that an increase in market tightness results in animmediate increase in market liquidity and a more gradual increase inprices. Furthermore, market liquidity goes down after some time asagents adapt to the new situation.

90


Figure 4.1: Response of prices (B), matching probability (C) to a change in markettightness (A) in week 11.

(A)

10

20

30

40

0

1

2

3

Time in Weeks

Mark

etTightn

ess

Real Market Tightness (λ)

Perceived Market Tightness (λ∗)

(B)

10

20

30

40

400

420

440

460

480

500

Time in Weeks

Tre

shold

Pri

ce

in1000e

Prices (ε)

(C)

10

20

30

40

0.15

0.2

0.25

Time in Weeks

Seller’sM

atchin

gPro

bability

Seller’s Matching Probability (ω)

91


4.4 Data

The search and match model of the previous section includes markettightness (λ), threshold prices (ε), and sale probability (ω). In this sec-tion, we explain how we create empirical measures for these variables(see Table 4.1 for the link between the symbols in the empirical andtheoretical models). We estimate a price index as a measure for pricesand use the rate of sale as sale probability. To estimate market tight-ness, we use click data. We use individual transactions data to estimatethe former two and individual click data to estimate the latter. Bothtransaction data and click data are available at the individual level, butdue to privacy issues the lowest scale at which the data can be matchedis the ZIP code level.3 We aggregate the data further to municipal andquarterly levels because the ZIP code level is too detailed to generaterepresentative price, liquidity, and internet search indices. Hence, weestimate a price index and a liquidity measure from individual transac-tions data, and a market tightness measure from individual click datafor each municipality on a quarterly basis (Table 4.1).

Table 4.1: Overview of variables.

Variable Description Symbol* Period Source

pr House price index ε 2000 - 2014 NVM**ros Rate of sale ω 2000 - 2014 NVM**clc Clicks N/A 2011 - 2014 Fundalh Number of listed houses N/A 2011 - 2014 Fundacph Clicks per house λ 2011 - 2014 Funda**

*Correspondence symbol in theoretical model in Literature review/Appendix, **Owncalculations. The price index and rate of sale are based on individual transaction dataof the Dutch Brokerage Association (NVM), see Appendix for the index estimation.All variables except rate of sale in logs. All variables on quarterly and municipal scale.

4.4.1 Transaction data

We use detailed data from the Dutch Association of Real Estate Brokersand Real Estate Experts (NVM) to construct the quarterly house priceindices at the municipal level. The data include the sale price, date ofsale, and several house-specific characteristics (see Appendix). In total,

3The ZIP code is on a 4-digit level.

92


over 1.6 million transactions between 2000 and 2013 are included.4

Although a large share of all transactions in the Netherlands is in-cluded, the data do not cover all transactions. The data set includesapproximately 69% of all transactions over the sample period (2011 -2013).5 Kerste et al. (2012) and De Wit et al. (2013) report percentagesof 75% in 2010 and 55-60% in 2007, respectively.

The house price index is estimated using a Hierarchical Trend Model(HTM) as proposed by Francke and De Vos (2000) and Francke and Vos(2004). A HTM is a hedonic price model that specifically addressesthe spatial and temporal dependence of selling prices and is well suitedto construct constant-quality house price indices in thin markets. Themodel and price indices of the municipalities within one COROP region(Amsterdam region) are presented in the Appendix.6

Figure 4.2 contains a map with the values of a standardized home ineach municipality between 2011 and 2013. The map shows that centralareas in the Randstad or areas close to the Randstad and cities are moreexpensive.7 The constant-quality prices are the lowest in the northernprovinces of Friesland, Groningen, and Drenthe.

The rate of sale per municipality per quarter is determined by di-viding the number of sales by the number of houses for sale at thebeginning of the quarter. The indices are the trend components fromindividual unobserved component models per municipality (local levelwith a stochastic seasonal component, see Equations B.11a - B.11c in theAppendix).8 Figure B.1 in the Appendix includes rate of sale estimatesof municipalities within one COROP region.

During the sample period, house prices generally declined (Table4.2). During 2013 some areas started recovering, but on average house

4A transaction is denoted as “transaction” in the NVM database at the time ofthe signing of the purchase contract. In other Dutch databases, such as that of theKadaster (Dutch Land Registry) a transaction is included when the legal transfertakes place. Therefore it is generally found that NVM transaction data lead otherdata sources.

5The NVM data set used consists of roughly 1.6 million transactions between2000 and 2013. Statistics Netherlands (CBS) which records all transactions, reportedjust under 2.4 million transactions over this same period.

6A COROP region is the Dutch equivalent of a NUTS-3 region, comparable tothe MSA classification in the US.

7The Randstad is the most urbanized area in the Netherlands and includes citieslike Amsterdam, Rotterdam, The Hague, and Utrecht.

8We also experimented with a local linear trend model, but the drift componentwas insignificant for most municipalities.

93


prices declined. The rate of sale declined in 2011 and 2012, but startedincreasing in 2013.

The results of Fisher’s combing p-values test (Maddala and Wu,1999), in which separate ADF regressions are run for each municipality,are shown in Table 4.3. House prices are I(1): The series are non-stationary in levels but stationary in first-differences. The rate of saleseries is I(0).

Table 4.2: Descriptive statistics 2011 - 2013 per year.

Year Variable Average σ Min Max

2011 pr 8.4786 0.4572 7.1014 9.6625ros 0.0421 0.0135 0.0103 0.1158CPH 248.91 92.98 39.47 679.36cph 5.4379 0.3788 3.8421 6.3550∆pr -0.0123 0.0059 -0.0367 0.0080∆ros -0.0026 0.0048 -0.0282 0.0212∆cph -0.0471 0.0652 -0.3382 0.2145

2012 pr 8.4127 0.4577 7.0827 9.6298ros 0.0372 0.0109 0.0147 0.0903CPH 238.72 80.51 29.67 621.41cph 5.4043 0.3733 3.5568 6.1941∆pr -0.0174 0.0052 -0.0407 0.0027∆ros -0.0004 0.0040 -0.0255 0.0184∆cph 0.0012 0.0501 -0.2347 0.2693

2013 pr 8.3667 0.4568 7.0768 9.5625ros 0.0423 0.0148 0.0161 0.1242CPH 262.20 78.33 38.29 536.67cph 5.5137 0.3465 3.4934 6.2972∆pr -0.0062 0.0070 -0.0331 0.0164∆ros 0.0024 0.0048 -0.0162 0.0616∆cph 0.0556 0.0628 -0.2136 0.3889

2011-2013 pr 8.4193 0.4594 7.0768 9.6625ros 0.0405 0.0134 0.0103 0.1242CPH 249.95 84.72 29.67 679.36cph 5.4519 0.3692 3.4934 6.3550∆pr -0.0120 0.0076 -0.0407 0.0164∆ros -0.0002 0.0050 -0.0282 0.0616∆cph 0.0078 0.0719 -0.3382 0.3889

Average and σ depict the mean and standard deviations of the respective variable peryear of all municipalities. Min and Max describe the minimum and maximum valueof the variable of any municipality in the corresponding year. CPH and cph denotethe regular and log-transformed versions of the clicks per house variable respectively.∆ denotes the average quarterly change of the given year.

94


Figure 4.2: Map that depicts the value of a standardized home within each munici-pality between 2011 and 2013.

40

Kilo

me

ters

Pric

e

Figure 4.3: Map that depicts the average number of clicks per house of each munici-pality between 2011 and 2013.

40

Kilo

me

ters

Cli

cks

95


Table 4.3: Results of the combining p-value tests to test for a unit root in the specifiedvariables.

Variable 1 lag in ADF 3 lags in ADFp-statistic p-value p-statistic p-value

pr 488.08 1.0000 617.80 1.0000ros 926.72 0.0020∗∗∗ 1584.88 0.0000∗∗∗

cph 1118.20 0.0000∗∗∗ 1102.11 0.0000∗∗∗

∆pr 976.14 0.0000∗∗∗ 1732.69 0.0000∗∗∗

∆ros 1927.74 0.0000∗∗∗ 3707.65 0.0000∗∗∗

∆cph 1178.62 0.0000∗∗∗ 6341.29 0.0000∗∗∗

The regressions include a time trend and cross-sectional specific intercepts between2011 and 2013. H0: All panels contain unit roots, Ha: At least one panel is station-ary. We experimented with different lag-lengths in the separate ADF equations, lags1 and 3 are shown. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

4.4.2 Internet search data

Internet search popularity is based on data of the Funda housing web-site. In a survey conducted by Conclusr (2014), 93% of the respondentsmention “Funda” when they are asked to name a housing website. Ad-ditionally, 81% of the respondents would prefer Funda if they were tosell their home online. In 2013, the website registered 4.2 million uniquevisitors per month. Kerste et al. (2012) show that Funda is by far themost popular housing website in the Netherlands. According to their re-search, the website has a stable market share of around 60% of all Dutchhousing websites.9 The owner of Funda is NVM, from which the trans-action data originate. Kerste et al. (2012) point out that most of thereal estate brokers who are active on Funda are NVM members. Hencethe online listed properties are mostly properties that are brokered bythese NVM members.

The data measure the clicks per listed house per month. In 2011, forexample, the house at the 99th percentile is clicked almost 3,600 times,while the house that is clicked most often is clicked almost 315,000 times.These outliers are likely to be houses of celebrities or other special prop-erties that may be not representative for the “true” demand. Therefore,listings for which the number of clicks is in the top percentile of eachyear are removed. After aggregation and the removal of outliers, themean number of clicks per house per quarter in 2011 is 249. These

9The websites that have the second and third largest market share are Jaap.nl,and Huizenzoeker.nl with market shares of 9% and 8% respectively.

96


listed houses are subsequently linked to a municipality. In the next stepthe totals per municipality and per quarter are calculated. The resultsare (i) the amount of clicks and (ii) the number of listed houses permunicipality per quarter.

To measure “internet search popularity”, we generate our main vari-able of interest: clicks per house (cph). This variable measures howmany times listed properties have been clicked upon on average perquarter per municipality. This variable has an intuitive interpretation:A higher value for this variable could indicate a more popular area. Theconcept behind this interpretation is as follows. More clicks on a listedproperty could indicate that the property is relatively popular. There-fore, more clicks on houses (relative to the total number of listed houses)in a certain area could be evidence for a more popular area. The num-ber of clicks can therefore be characterized as a demand variable. Thenumber of listed houses per area can proxy for supply in a certain area,as these are roughly equal to the number of houses for sale in this area.The generated variable can therefore approximate market tightness.

With respect to the popularity of the municipalities on Funda, themap in Figure 4.3 suggests that areas in or close to larger cities aremore popular. The similarity between the maps regarding prices andclicks per house is striking. For example, houses in the Randstad areclicked on more often and are also more expensive. Likewise, houses inmunicipalities located in or near larger cities are clicked on more oftenand are more expensive.

Although this pattern is visible for most of the country, some areasin the north (e.g. Groningen) show a somewhat different pattern. Apossible explanation for this phenomenon is that this area was hit byseveral induced earthquakes.10 Although research by Francke and Lee(2013) suggests that prices changes in Groningen did not differ signifi-cantly from comparable neighboring areas, news regarding house pricesin Groningen may have triggered internet search behavior. This simul-taneous causality is explicitly taken into account as internet search be-havior is also included as dependent variable. Furthermore, we performan additional robustness check that excludes this area.

To deal with seasonal effects and noise, the clicks per house serieshave been seasonally adjusted and smoothed by estimating an unob-

10The earthquakes in the area are caused by natural gas extraction, for more in-formation (Dutch only) see http://www.rijksoverheid.nl/onderwerpen/aardbevingen-in-groningen/aardbevingen-door-gaswinning-in-groningen.

97


served components model. See the description and Equations (B.11a) -(B.11c) in the Appendix.

Over the sample period, the market tightness indicator decreased in2011, remained relatively stable in 2012, and started increasing in 2013(Table 4.2). Figure B.1 in the Appendix shows the development of theclicks per house in the Amsterdam region.

Finally, the unit root tests in Table 4.3 indicate that the clicks perhouse variable is I(0).

4.5 Empirical model

In order to examine the relationship between house prices, liquidity, andmarket tightness we specify a panel Vector Autoregression (VAR) modelin Equation (4.1). Table 4.1 describes the variables used. All variablesexcept the rate of sale are modeled in logs. We take the first differenceof these variables in the panel VAR.

∆Yi,t =

∆pri,t∆rosi,t∆cphi,t

=

Q∑q=1

Γq∆Yi,t−q +νt+εi,t, εi,t ∼ N(0,Σε). (4.1)

Here Y is a matrix of dependent variables which contains changes in loghouse prices (pr), changes in liquidity (ros), and changes in log numberof clicks per house (cph). All variables depend on their lags up to quarterQ, and the estimated coefficients are included in matrix Γ. Subscriptsi and t denote the municipality and quarter, respectively. Time fixedeffects are included in the models and are denoted by νt. By takingfirst differences, the time invariant unobserved heterogeneity betweenthe municipalities cancels out. Finally, ε is the error term.

By taking first-differences we control for this time invariant unob-served heterogeneity. Hence we effectively control for factors that aredifferent per municipality, but that are constant over time. Take, for ex-ample, residential mobility rates. Since there is a lot of intra-municipalresidential mobility, many sellers are also buyers in the same municipal-ity. Since these residential mobility rates exhibit geographical differences(Lee, 2014), it might be that demand in certain municipalities is under-or overestimated. Data from Statistics Netherlands (CBS) suggest thatmobility rates remained virtually the same between 2011 and 2013.11

11Intra-municipal movements in the Netherlands were 59.3% of all movements in

98


Hence, these are captured by the fixed effects. Similar concerns mightarise due to differences in internet access. Although, data on geograph-ical differences were unavailable, national figures on internet access re-mained virtually the same between 2011 and 2013. In 2011 and 2012,94% of all households had access, and in 2013 this was 95% (CBS).

In this model price changes, changes in liquidity and changes in clicksare modeled simultaneously. The simultaneous causality is explicitlytaken into account by the impulse responses based on the VAR model,and they therefore provide insights into this relationship.

It is generally accepted that house price changes exhibit positiveserial correlation in the short run (Capozza et al., 2004). This suggeststhat lags of the dependent variable should be included in the models.Hence, modeling the data as a dynamic panel is the obvious choice.A problem that arises when including lags of the dependent variablein the regression is that these lags are correlated with the error termand therefore will yield biased results (i.e. Nickell’s bias, see Nickell1981). This is especially the case in our setup, in which we have arelatively “small T”. In order to allow for these issues, the parametersof interest are estimated using Generalized Method of Moments (GMM,see Arellano and Bond, 1991, and Love and Zicchino, 2006).

4.6 Results

4.6.1 Estimation results

Table 4.4 presents the results of the quarterly panel VAR of the houseprice index, rate of sale, and clicks per house for Dutch municipalitiesbetween 2011 and 2013. In column (i) the change in house prices is thedependent variable, in column (ii) the change in the rate of sale is thedependent variable, and in column (iii) the change in clicks per house isthe dependent variable. The optimal number of lags as indicated by theinformation criteria is two quarters.12

The results indicate that the one-quarter lagged clicks per housevariable is positive and significant in the house price equation (at 5%)and positive and marginally significant (at 10%) in the rate of sale equa-

2011, 58.8% in 2012, and 58.3% in 2013.12MMSC-Bayesian information criterion (MBIC), MMSC-Akaike’s information cri-

terion (MAIC), and MMSC-Hannan and Quinn information criterion (MQIC), seeAndrews and Lu (2001).

99


Table 4.4: Panel VAR regression results.

∆pr ∆ros ∆cph

∆prt−1 0.4860∗∗∗ 0.0091 -0.0342(14.8) (0.4) (-0.1)

∆prt−2 0.0455∗∗ -0.0285 -0.0173(2.1) (-1.2) (-0.1)

∆rost−1 0.0466∗ -0.0597∗ -0.7902∗∗∗

(1.8) (-1.6) (-3.4)∆rost−2 0.0120 -0.0024 -0.4074∗∗

(0.7) (-0.1) (-2.2)∆cpht−1 0.0036∗∗∗ 0.0031∗ 0.1445∗∗∗

(2.6) (1.9) (6.6)∆cpht−2 0.0003 -0.0020 0.0268

(0.2) (-1.4) (1.6)

Granger causality tests:∆pr N/A 1.814 0.033∆ros 3.449 N/A 12.175∗∗∗

∆cph 7.096∗∗ 6.187∗∗ N/AAll variables 10.974∗∗ 9.766∗∗ 12.343∗∗

Fixed effects Quarterly and municipalN 3224Number of panels 403Sample period 2011Q1 - 2013Q4Eigenvalue stability Condition Yes, panel VAR is stable

In (i) changes in log house prices are regressed on lagged changes in log house prices,lagged changes in the rate of sale and changes clicks per house, in (ii) changes in therate of sale are regressed on lagged changes in rate of sale, lagged changes in log houseprices and, changes in log clicks per house. Column (iii) includes the clicks per houseas dependent variable. The time and cross-sectional dimensions are 12 quarters be-tween 2011 and 2013 and 403 municipalities respectively. Coefficients are estimatedusing GMM and standard errors are robust to hetereoscedasticity, autocorrelation,and are clustered by municipality. Optimal number of lags based on information cri-teria. The table further depicts Granger causality tests of the variables in each equa-tion. T-statistics in parentheses, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

100


tion. This indicates that growth in clicks in the previous quarter leads togrowth in house prices and the rate of sale in this quarter. The Grangercausality tests confirm these findings. They indicate that a change inthe clicks per house Granger causes both changes in house prices andthe rate of sale. A one-quarter lag might seem too short (i.e. the timebetween browsing the internet and the sale might be longer), but recallthat the transaction date refers to the signing of the purchase contract;the actual legal transfer is approximately three months later.13

The model also provides evidence regarding the price-volume correla-tion. Changes in the lagged rate of sale have a positive effect on currentprice changes, although the effect is only marginally significant. Thisconfirms the findings of, for example, Miller and Sklarz (1986) and Car-rillo et al. (2015), who find that the changes in the rate of sale and saleprobability are leading indicators of price changes. The Granger causal-ity test, however, cannot be interpreted as significant. There seems tobe no effect of price changes on changes in liquidity. However, the lackof significance in these findings might also be attributed to the relativeshort sample (i.e. 3 years).

Interestingly, the change in the clicks per house responds significantlyto changes in the rate of sale, but not to changes in prices. Thereseems to be a negative relationship between changes in the rate of saleand changes in the clicks per house. The underlying mechanism mightbe the following. If more houses are sold in the previous quarter in amunicipality, there are fewer potential buyers left who are still lookingto buy a house in this municipality, hence the reduction in clicks. TheGranger causality test confirms that changes in the rate of sale Grangercause changes in the clicks per house.

4.6.2 Impulse responses

To interpret the results economically and compare them with the the-oretical predictions (Figure 4.1), this section looks at the impulse re-sponses based on the model in Table 4.4. The main advantage of in-terpreting the results by studying the impulse responses is that thesecapture the full dynamics of the model. If, for example, the clicks perhouse increases, the one-quarter lagged coefficient indicates a direct ef-

13For example, De Wit et al. (2013) use a lag of 3 months, see also (Dutch only):http://www.kadaster.nl/web/Themas/Themapaginas/dossier/Toelichting-op-de-cijfers-in-het-Vastgoed-Dashboard.htm.

101


fect on prices. There is, however, an additional effect running throughliquidity. Furthermore, the autoregressive components amplify the ef-fects. This is also clearly visible in Figure 4.4, which depicts the impulseresponses. The dashed lines represent the 95% confidence bounds andare obtained by Monte Carlo simulation. The size of a shock amountsto one standard deviation of the impulse variable.

In panel (G) of Figure 4.4, we shock the growth in the clicks perhouse, and the function depicts the response of changes in house prices.The graph shows that the largest growth in house prices is in the quarterafter the shock, and that the shock only dies out after roughly one year.The shock also has a positive impact on the change in the rate of salein the first quarter. The shock has a negative impact (although onlymarginally significant) on the rate of sale in the second quarter.

Figure 4.4: Impulse-response functions, impulse variable → response variable.

(A)

0 1 2 3 4 5 60

10

20

30

40

Time in quarters

Growth

in%

cph → cph (B)

0 1 2 3 4 5 6−2

−1.5

−1

−0.5

0

Time in quarters

Growth

in%

ros → cph

(C)

0 1 2 3 4 5 6−20

−10

0

10

20

Time in quarters

Growth

in%

pr → cph

(D)

0 1 2 3 4 5 6−0.2

−0.1

0

0.1

0.2

Time in quarters

Growth

in%

cph → ros (E)

0 1 2 3 4 5 6

0

0.5

1

Time in quarters

Growth

in%

ros → ros(F)

0 1 2 3 4 5 6

−2

0

2

Time in quarters

Growth

in%

pr → ros

(G)

0 1 2 3 4 5 60

0.1

0.2

0.3

Time in quarters

Growth

in%

cph → pr (H)

0 1 2 3 4 5 60

0.05

0.1

0.15

Time in quarters

Growth

in%

ros → pr(I)

0 1 2 3 4 5 60

10

20

30

40

50

Time in quarters

Growth

in%

pr → pr

The cumulative impulse responses as presented in Figure 4.5 depictthe cumulative growth of the response variable after a shock in the im-pulse variable. Hence, these can be interpreted as the level change ofthe response variable, and we can compare these with the theoreticalpredictions in Figure 4.1. Figure 4.5 indicates that a change in the

102


Figure 4.5: Cumulative impulse-response functions, impulse variable→ response vari-able.

(A)

0 1 2 3 4 5 630

35

40

45

50

Time in quarters

Growth

in%

cph → cph (B)

0 1 2 3 4 5 6−3

−2

−1

0

Time in quarters

Growth

in%

ros → cph (C)

0 1 2 3 4 5 6

−50

0

50

Time in quarters

Growth

in%

pr → cph

(D)

0 1 2 3 4 5 6

−0.2

−0.1

0

0.1

0.2

Time in quarters

Growth

in%

cph → ros (E)

0 1 2 3 4 5 61.2

1.3

1.4

1.5

Time in quarters

Growth

in%

ros → ros(F)

0 1 2 3 4 5 6

−5

0

5

Time in quarters

Growth

in%

pr → ros

(G)

0 1 2 3 4 5 60

0.2

0.4

0.6

0.8

Time in quarters

Growth

in%

cph → pr (H)

0 1 2 3 4 5 60

0.1

0.2

0.3

0.4

Time in quarters

Growth

in%

ros → pr(I)

0 1 2 3 4 5 640

60

80

100

120

Time in quarters

Growth

in%

pr → pr

number of clicks has a permanent effect on prices, but only a temporaryeffect on liquidity as measured by the rate of sale. An increase in onestandard deviation innovation (4.6%) in the clicks per house leads to apermanent price increase of roughly 0.4%. The rate of sale increases byapproximately 0.1 percentage points after 1 quarter. After 2 quarters,the rate of sale decreases to the pre-shock level. This is similar to thepredictions of the search and matching model (Figure 4.5). Liquidityincreases relatively quickly after a shock, but decreases after the agents’expectations adapt to reality (compare panel (D) of Figure 4.5 to panel(C) of Figure 4.1). The effect on prices is permanent, but much moregradual (compare panel (G) of Figure 4.5 to panel (B) of Figure 4.1).

The slow adjustment process of prices during a period in which pricesdecreased (Table 4.2) is also supportive of the principle of loss aversion(Genesove and Mayer, 2001). To illustrate this, consider the case of anegative demand shock. Because sellers keep their listing prices too highduring these bad times, fewer transactions occur. After sellers realizethe market has gone down, they lower their listing prices resulting in

103


further price decreases.To summarize: (1) Market liquidity responds positively to an in-

crease in market tightness and (2) the effect of market tightness onprices is also positive. Finally, (3) liquidity responds quickly to a pos-itive demand shock, and its effect is short-lived. Prices respond muchmore gradually and there seems to be a permanent increase in prices.This is in line with the theoretical findings. Finally, the results are alsoin line with the principle of loss aversion as documented by Genesoveand Mayer (2001).

4.7 Robustness checks

4.7.1 Earthquake area

The earthquakes in the northeastern part of the Netherlands (Gronin-gen) might have resulted in an increase in searches, while people werenot actually interested in buying in the area. This might bias the esti-mated coefficients. We performed a robustness check by excluding themunicipalities within the earthquake area. The results are shown in Ta-ble 4.5. All coefficients and their significance are very similar to thosepresented in Table 4.4, hence the coefficients are not biased due to theinclusion of the region.

104


Table 4.5: Panel VAR regression without earthquake areas.

∆pr ∆ros ∆cph

∆prt−1 0.4744∗∗∗ 0.1024 0.0054(14.2) (0.4) (0.0)

∆prt−2 0.0404∗ -0.0293 0.0153(1.8) (-1.2) (0.1)

∆rost−1 0.0497∗ -0.0590∗ -0.7665∗∗∗

(1.9) (-1.7) (-3.2)∆rost−2 0.0120 -0.0024 -0.4074∗∗

(0.7) (-0.1) (-2.2)∆cpht−1 0.0037∗∗∗ 0.0032∗ 0.1425∗∗∗

(2.6) (1.9) (6.3)∆cpht−2 0.0004 -0.0016 0.0361∗∗

(0.3) (-1.1) (2.1)

Granger causality tests:∆pr N/A 1.912 0.008∆ros 3.537 N/A 10.709∗∗∗

∆cph 7.274∗∗ 5.211∗ N/AAll variables 11.335∗∗ 8.970∗ 10.864∗∗

Fixed effects Quarterly and municipalN 3224Number of panels 380Sample period 2011Q1 - 2013Q4Eigenvalue stability Yes, panel VAR is stable

In (i) changes in log house prices are regressed on lagged changes in log house prices,lagged changes in the rate of sale and changes in log clicks per house, in (ii) changesin the rate of sale are regressed on lagged changes in rate of sale, lagged changes in loghouse prices and changes in log clicks per house. Column(iii) includes the clicks perhouse as dependent variable. The time and cross-sectional dimensions are 12 quartersbetween 2011 and 2013 and 380 municipalities respectively. The municipalities withinthe earthquake area have been left out. Coefficients are estimated using GMM andstandard errors are robust to hetereoscedasticity, autocorrelation, and are clusteredby municipality. Optimal number of lags based on information criteria. The tablefurther depicts Granger causality tests of the variables in each equation. T-statisticsin parentheses, ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

105


4.7.2 Geographical variation in responsiveness

By including detailed internet search data, we aim to account for localshocks in the housing market. Geographical differences might result indifferences in how the process is transmitted in liquidity and prices. Inthis section we divide the Netherlands in two parts, based on the degreeof urbanization, to see whether the process differs. We use the definitionof the degree of urbanization from Statistics Netherlands (CBS) to dividethe Netherlands into urbanized and rural areas.14

Table 4.6 presents the regression results for the two different samples.The size of the effect of a shock in market tightness is roughly similarin both samples.15 There is, however, a difference in how the shockin transmitted. The positive effect of a shock in market tightness onliquidity is present in the urban areas, but not in rural areas. In the latterareas the effect runs directly from market tightness to prices. Hence wedo not observe a temporary increase in liquidity in rural markets, onlyin urban markets.

This probably reflects that urban markets, in our sample, startedrecovering in the final year of the estimation period, whereas rural mar-kets did not. The rate of sale increased in 2013 by only 0.5 percentagepoints in rural areas compared to 1.6 percentage points in urban areas.Nevertheless, in rural areas the effect on prices is still positive and grad-ual, similar to the effect found in the full sample and in the sample withurban areas.16

To gain further insights into this process, we also split the samplebased on average annual returns. We cluster the 50% highest and the50% lowest returns and run separate models for each group. The resultsare shown in Table 4.7. We find that the effects on both prices and liq-uidity are somewhat stronger in regions that have had higher returns orless negative returns. This reflects that these regions started recovering

14CBS uses five categories, but as some categories are relatively small we regroupthe municipalities into two categories. Categories 1 and 2 of CBS are categorized asrural, and categories 3 to 5 are categorized as urban. There are 242 rural municipal-ities and 161 urban municipalities.

15The one-quarter lagged coefficient is larger in the urban sample, but the totaleffect as given by the cumulative impulse-response functions (not shown) is similar.The same holds for the effect on the rate of sale; in this equation both the one-quarter lagged (positive) and two-quarter lagged (negative) coefficients are bigger,but the cumulative effect is similar.

16These are the municipalities that are located within the COROP regions Oost-Groningen, Delfzijl en Omgeving and Overig Groningen.

106


Table

4.6

:R

obust

nes

sch

eck:

Panel

VA

Rre

gre

ssio

nre

sult

sof

urb

an

are

as

and

rura

lare

as.

Urb

an

Rura

l∆pr

∆ros

∆cph

∆pr

∆ros

∆cph

∆pr t

−1

0.4

948∗∗

∗0.0

190

-0.4

429

0.5

959∗∗

∗-0

.0108

0.2

238

(7.1

)(0

.3)

(-1.3

)(1

4.0

)(-

0.3

)(0

.6)

∆pr t

−2

0.0

809∗

-0.0

371

-0.1

359

0.0

342

-0.0

216

0.0

087

(1.9

)(-

0.8

)(-

0.6

)(1

.2)

(-0.7

)(0

.0)

∆ros t

−1

0.0

876∗∗

-0.0

849∗∗

-0.7

533∗∗

∗0.0

012

-0.0

229

-0.9

928∗∗

(2.0

)(-

2.0

)(-

2.9

)(0

.0)

(-0.5

)(-

2.2

)∆ros t

−2

0.0

242

0.0

366

-0.4

970∗∗

0.0

112

-0.0

593

-0.3

325

(0.8

)(0

.9)

(-2.2

)(0

.4)

(-1.5

)(-

0.9

)∆cpht−

10.0

067∗∗

0.0

093∗∗

0.0

947∗∗

∗0.0

033∗

0.0

002

0.1

766∗∗

∗

(2.0

)(2

.3)

(3.1

)(1

.9)

(0.1

)(6

.2)

∆cpht−

2-0

.0017

-0.0

072∗∗

0.0

285

0.0

012

0.0

001

0.0

279

(-0.5

)(-

2.0

)(1

.1)

(0.6

)(0

.1)

(1.2

)

Gra

nger

causa

lity

test

s:∆pr

N/A

1.7

73

1.7

71

N/A

0.6

38

0.3

11

∆ros

4.0

70

N/A

8.4

28∗∗

0.1

33

N/A

4.6

57∗

∆cph

4.6

60∗

10.4

90∗∗

∗N

/A

3.6

40

0.0

18

N/A

All

vari

able

s9.0

31∗

14.6

62∗∗

∗10.0

96∗∗

3.7

28

0.7

26

4.8

45

Fix

edeff

ects

Quart

erly

and

munic

ipal

Quart

erly

and

munic

ipal

N1288

1936

Num

ber

of

panel

s161

242

Sam

ple

per

iod

2011Q

1-

2013Q

42011Q

1-

2013Q

4E

igen

valu

est

abilit

yY

es,

panel

VA

Ris

stable

Yes

,panel

VA

Ris

stable

In(i

)/(i

v)

changes

inlo

ghouse

pri

ces

are

regre

ssed

on

lagged

changes

inlo

ghouse

pri

ces,

lagged

changes

inth

era

teof

sale

and

changes

inlo

gcl

icks

per

house

,in

(ii)

/(v

)ch

anges

inth

era

teof

sale

are

regre

ssed

on

lagged

changes

inra

teof

sale

,la

gged

changes

inlo

ghouse

pri

ces

and

changes

inlo

gti

mes

clic

ks

per

house

.C

olu

mns

(iii)/

(vi)

incl

ude

the

clic

ks

per

house

as

dep

enden

tva

riable

.T

he

tim

edim

ensi

on

is12

quart

ers

bet

wee

n2011

and

2013.

Colu

mns

(i)-

(iii)

incl

ude

161

urb

an

munic

ipaliti

esand

colu

mns

(iv)-

(vi)

incl

ude

242

rura

lm

unic

ipaliti

es.

Coeffi

cien

tsare

esti

mate

dusi

ng

GM

Mand

standard

erro

rsare

robust

tohet

ereo

sced

ast

icit

y,au-

toco

rrel

ati

on,

and

are

clust

ered

by

munic

ipality

.O

pti

mal

num

ber

of

lags

base

don

info

rmati

on

crit

eria

.T

he

table

furt

her

dep

icts

Gra

nger

causa

lity

test

sof

the

vari

able

sin

each

equati

on.

T-s

tati

stic

sin

pare

nth

eses

,∗p<

0.1

0,∗∗p<

0.0

5,∗∗

∗p<

0.0

1.

107


in the final year of the estimation period. With respect to the dynam-ics, an increase in market tightness results in a temporary increase inliquidity and a permanent increase in prices.

For regions with the most negative returns the effect on prices issignificant, but more spread out over the first and second lags. Thisindicates that price discovery is slower in these regions.17 We suspectthat the underlying reason for this is the poor state of the market inthese regions. The (level of the) rate of sale was substantially lower inthese regions compared to regions with higher returns. Therefore, therewas less information that sellers could use in setting their reservationprices. This is also reflected in the coefficients of clicks per house in therate of sale equation. These have the same sign as in other regressions,but are insignificant.

17This would be reflected by a higher α in the theoretical model: Sellers attachmore weight to perceived market tightness in the previous period.

108


Table

4.7

:R

obust

nes

sch

eck:

Panel

VA

Rre

gre

ssio

nre

sult

sw

ith

the

sam

ple

split

intw

oacc

ord

ing

toth

eav

erage

annualize

dre

turn

s.

50%

hig

hes

tre

turn

s50

%lo

wes

tre

turn

s∆pr

∆ros

∆cph

∆pr

∆ros

∆cph

∆pr t

−1

0.4

614∗∗

∗0.0

258

-0.2

542

0.5

079∗∗

∗-0

.0101

0.2

483

(10.4

)(0

.8)

(-0.8

)(1

0.5

)(-

0.3

)(0

.7)

∆pr t

−2

0.0

113

-0.0

418

0.2

750

0.0

811∗∗

-0.0

116

-0.3

971∗

(0.5

)(-

1.5

)(1

.1)

(2.2

)(-

0.3

)(-

1.6

)∆ros t

−1

0.0

710∗

-0.0

600

-0.6

836∗∗

0.0

182

-0.0

592

-0.9

537∗∗

∗

(1.9

)(-

1.4

)(-

2.1

)(0

.6)

(-1.2

)(-

2.9

)∆ros t

−2

0.0

030

-0.0

378

-0.3

554

0.0

252

0.0

411

-0.4

394∗

(0.1

)(-

1.0

)(-

1.4

)(1

.0)

(0.8

)(-

1.7

)∆cpht−

10.0

053∗∗

∗0.0

042∗

0.1

201∗∗

∗0.0

015

0.0

022

0.1

735∗∗

∗

(2.7

)(1

.7)

(3.8

)(0

.8)

(1.0

)(4

.9)

∆cpht−

2-0

.0028

-0.0

011

0.0

491∗∗

0.0

039∗

-0.0

029

0.0

012

(-1.5

)(-

0.5

)(2

.2)

(1.9

)(-

1.6

)(0

.1)

Gra

nger

causa

lity

test

s:∆pr

N/A

3.6

16

1.6

42

N/A

0.1

41

2.8

26

∆ros

3.5

43

N/A

4.6

94∗

1.3

04

N/A

9.0

30∗∗

∗

∆cph

8.2

98∗∗

∗3.9

02

N/A

4.5

49∗

3.2

46

N/A

All

vari

able

s13.0

61∗∗

∗9.8

33∗∗

6.7

50

5.6

55

4.7

28

11.1

12∗∗

Fix

edeff

ects

Quart

erly

and

munic

ipal

Quart

erly

and

munic

ipal

N1616

1608

Num

ber

of

panel

s202

201

Sam

ple

per

iod

2011Q

1-

2013Q

42011Q

1-

2013Q

4E

igen

valu

est

abilit

yY

es,

panel

VA

Ris

stable

Yes

,panel

VA

Ris

stable

In(i

)/(i

v)

changes

inlo

ghouse

pri

ces

are

regre

ssed

on

lagged

changes

inlo

ghouse

pri

ces,

lagged

changes

inth

era

teof

sale

and

changes

inlo

gcl

icks

per

house

,in

(ii)

/(v

)ch

anges

inth

era

teof

sale

are

regre

ssed

on

lagged

changes

inra

teof

sale

,la

gged

changes

inlo

ghouse

pri

ces

and

changes

inlo

gcl

icks

per

house

.C

olu

mns

(iii)/

(vi)

incl

ude

the

clic

ks

per

house

as

dep

enden

tva

riable

.T

he

tim

edim

ensi

on

is12

quart

ers

bet

wee

n2011

and

2013.

Colu

mns

(i)-

(iii)

incl

ude

161

urb

an

munic

ipaliti

esand

colu

mns

(iv)-

(vi)

incl

ude

253

rura

lm

unic

ipaliti

es.

Coeffi

cien

tsare

esti

mate

dusi

ng

GM

Mand

standard

erro

rsare

robust

tohet

ereo

sked

ast

icit

y,au-

toco

rrel

ati

on,

and

are

clust

ered

by

munic

ipality

.O

pti

mal

num

ber

of

lags

base

don

info

rmati

on

crit

eria

.T

he

table

furt

her

dep

icts

Gra

nger

causa

lity

test

sof

the

vari

able

sin

each

equati

on.

T-s

tati

stic

sin

pare

nth

eses

,∗p<

0.1

0,∗∗p<

0.0

5,∗∗

∗p<

0.0

1.

109


4.8 Conclusion

In this chapter, we have provided theoretical and empirical evidence onthe links between market tightness, market liquidity, and house prices.We further provide insights into differences in temporal dynamics andgeographical variations.

We have shown that changes in market tightness based on internetsearch data clicks per house Granger causes (1) changes in market liquid-ity and (2) house price changes. Furthermore, we show that (3) liquidityresponds relatively quickly to a demand shock and that prices respondmore gradually. The effect on liquidity is temporary, and the effect onprices is permanent. These empirical findings confirm the theoreticalfindings of Genesove and Han (2012), Carrillo et al. (2015) and the pre-dictions from the search and matching model outlined in the Appendix.The underlying theoretical mechanism is that when more buyers enterthe market and market tightness increases, this will be reflected in anincrease in the matching probability. Because of information asymme-tries, sellers do not observe this increase immediately and will adapttheir reservation and corresponding listing prices gradually. This resultsin a lagged response of prices, whereas market liquidity as measured bythe matching probability increases instantaneously.

Internet search behavior does not seem to respond to price changes,but only to changes in liquidity. The relationship between changes inliquidity and clicks per house is found to be negative, possibly becausemore houses sold in the previous quarter indicates that there are fewerpotential buyers left who are searching in a particular municipality.

As for geographical differences, a demand shock is temporarily ab-sorbed in market liquidity as measured by the rate of sale in urban areas.Prices adapt more gradually, and the effect is permanent. Following thisprice adjustment, liquidity reverts back close to its original level. In ru-ral areas the effect runs straight from market tightness to prices, but theincrease is still gradual as predicted by the search and matching model.The reason is that urban areas were recovering in the final year of thesample, whereas rural areas were not. This is also reflected in the modelestimated over the 50% regions with the most negative returns. Here,price discovery is found to be slower than in other regions which can beattributed to the low rate of sale in these regions. Theoretically, thiswould be reflected in sellers attaching more weight to their perceptionof market tightness yesterday than to signals of market tightness today.

110


Finally, as the sample period is characterized by declining houseprices, homeowners may have kept their listing prices too high. Hence,the slow adjustment process of prices found in this study is supportiveof loss aversion in the housing market (Genesove and Mayer, 2001).

In the internet era vast amounts of data are produced which can beincorporated into economic models. The main advantage of the data isthe availability at a disaggregate level, both regionally and with respectto the frequency. We have shown that the housing market proves to beno different than, for example, the labor market or the stock marketwhen it comes to the added value of consumer sentiment measured byinternet data. It is only possible to use Funda data from 2011 onwards,hence the sample period is only three years. Out-of-sample forecast ex-ercises are therefore not possible yet and are also not the focus of ourstudy. As the Granger causality tests have shown that the data has pre-dictive power, using this type of data could be useful in future researchon forecasting developments in the housing market. Furthermore, inthese three years prices generally decreased, therefore it would be in-teresting to repeat the research when prices are generally increasing tosee whether the dynamics are different. Furthermore, although the dataare available at house level they cannot be linked to the postal addressdue to privacy issues. If the Funda data could be linked to a specificaddress, it would be possible to merge them with the database fromwhich the house price and liquidity indices originate. This would allowthe research to be performed on an individual house level rather than onan aggregated scale. Finally, it would also be interesting to be able touse data regarding the user of the website (i.e. the clicker). In doing so,it would be possible to distinguish unique users and to distinguish be-tween intra- and inter-municipal demand. Demand could be overstatedif a single user clicks multiple times, but we expect this phenomenon tobe relatively constant over time. We are therefore confident that thiseffect is captured by the fixed effects.

111


112

Chapter 5Conclusion and summary“Commercial and residential realestate market liquidity”

Je gaat het pas zien als je hetdoor hebt.

Johan Cruijff

The main question of this thesis is: “What is the role of market liquid-ity in real estate assets?” The role of market liquidity is fundamentallyrelated to price movements in real estate. As such, prices and liquidityshould not be viewed as separate concepts. This thesis offered sev-eral novel measures for real estate market liquidity. All the measuresused and constructed in this thesis show strong cyclical movements anda strong co-movement with prices. This indicates that not only pricemovements should be tracked by policymakers, market watchers, andmarket participants. A better and completer picture of the market canbe painted when analyzing both liquidity and price movements. Maybeeven more importantly, all chapters in this thesis show that liquiditymovements tend to lead price movements. This insight could help poli-cymakers in spotting crises and overheating markets early on.

113

CHAPTER 5. CONCLUSION AND SUMMARY “COMMERCIALAND RESIDENTIAL REAL ESTATE MARKET LIQUIDITY”

5.1 How can real estate market liquidity be mea-sured?

The lack of well-established ways to measure market liquidity led to subquestion (i): “How can real estate market liquidity be measured?” Allthree chapters use different measures that relate to market liquidity andprovide similar answers to this question.

Chapter 2 constructs a methodology that allows for the construc-tion of constant-quality liquidity indices when transaction data are sparse.The presented methodology addresses the problem that the quality ofsold houses differs over time. Furthermore, in some periods it might bethe case that more properties are withdrawn. The latter is treated as acensoring problem and is explicitly taken into account in the methodol-ogy. One of the main advantages of the presented method is that it canalso be used in thin markets (i.e. markets with few transactions). In thinmarkets, indices produced by conventional methods may be unreliable.The proposed methodology introduces a structure on the estimated coef-ficients that allows to create reliable indices in these markets. Moreover,the presented methodology corrects for both quality and withdrawals.The results suggest that withdrawals are the main driver of the differ-ence between the average time on market (TOM) of sold properties andthe proposed measure. Quality, however, also plays a significant role.The quality of sold properties is different over the cycle.

Chapter 3 extends the constant-liquidity price index methodologyintroduced by Fisher et al. (2003) in two ways: (i) by casting the methodin a repeat sales framework, and (ii) by estimating the model in a struc-tural time series format. As a result, it is possible to disentangle reserva-tion prices of buyers and sellers for commercial real estate at a city levelwithout the need of a substantial set of property characteristics. The dif-ference between the central tendencies of buyers’ and sellers’ reservationprices can be seen as a measure for market liquidity. This is analogousto the widely used bid-ask spread measure in the stock market. The pre-sented method is a very general method to construct constant-liquidityprice indices and liquidity indices as it can be applied to every data setin which repeat sales can be identified and for which transaction pricesare known. By estimating the model in a structural time series format,the model can also be estimated for thin markets.

Chapter 4 provides evidence that a measure for market tightness

114


based on internet search behavior –clicks per house– Granger causeschanges in market liquidity based on transaction data –rate of sale. Theunderlying reason is that internet search behavior signals preparatorysteps to spend. The response goes in two ways: internet search behavioralso responds to changes in liquidity. The relationship between changesin liquidity and clicks per house is found to be negative, possibly becausemore houses sold in the previous quarter indicates that there are fewerpotential buyers left who are searching in a particular municipality.

5.2 What is the relationship between prices andmarket liquidity in real estate markets?

In the real estate literature, there is a strong focus on the interactionbetween market liquidity and prices. The chapters in this thesis alsoshed light on this question. In general, prices and liquidity are verymuch pro-cyclical and liquidity tends to lead prices.

Using the constant-quality TOM indices of Chapter 2 it is shown forthe Dutch housing market that during busts the TOM is high and marketliquidity is low. Furthermore, it is shown that there exists a commonalitybetween the constructed liquidity indices and transaction price indices.Additionally, a Granger causality analysis shows that liquidity changeslead price changes. A novel finding, consistent with the general assetpricing literature, is that liquidity risk is also higher in busts. Chapter2 also looks at the relationship between the list price premium and theTOM. A higher list price premium (i.e. a higher list price compared tothe predicted market value) is related to a lower sale probability. Theeffect is shown to be varying over time, in busts both the average listprice premium and the total effect on sale probability increases. In themost recent years (since 2015), the list price premium turns, on average,into a list price discount. The reason is that sellers change their behaviordue to the extreme tightness of the market.

Chapter 3 suggests that in all major US commercial real estatemarkets –for most of the time– liquidity is pro-cyclical. Because liquid-ity varies over time, “normal” price indices (i.e. observed transactionprice indices) provide –in some sense– an apples versus oranges compar-ison. This chapter provides a way to correct repeat sales price indicesfor varying liquidity. The result is that the constant-liquidity indices(demand reservation price indices) are more cyclical than normal price

115


indices and supply reservation price indices. The results further indicatethat the supply and normal price indices tend to move slower than thedemand indices. Note that this chapter is based on US commercial realestate data, the other two are based on Dutch residential data. It mightbe the case that the relationship between prices and liquidity is differentfor commercial and residential real estate as well as for US and Dutchreal estate. For example, commercial real estate is more of an investmentgood than residential real estate, which is more of a consumption good.Additionally, (some) US markets are much less supply-constrained thanmost Dutch markets, which may also have implications for the relation-ship. Moreover, the mortgage markets are different. This might haveimplications for reservation price behavior. For example, one might ar-gue that anchoring behavior is stronger in markets with high mortgagedebt as the negative equity problem can become larger. Apart fromthese descriptive differences, I will leave a more thorough examinationof these differences for future research.

Chapter 4 presents a theoretical and empirical model where trans-action prices, market tightness, and market liquidity are allowed to in-teract. The chapter further provides insights into differences in temporaldynamics and geographical variations. The results show that changesin market tightness based on internet search data (clicks per house),Granger causes (1) changes in market liquidity (breath of the market),and (2) house price changes. Furthermore, the results suggest that (3)liquidity responds relatively quickly to a demand shock and that pricesrespond more gradually. The effect on liquidity is temporary, and theeffect on prices is permanent. The underlying theoretical mechanism isthat when more buyers enter the market and market tightness increases,this will be reflected in an increase in the matching probability. Becauseof information asymmetries, sellers do not observe this increase imme-diately and will adapt their reservation and corresponding listing pricesgradually. This results in a lagged response of prices, whereas marketliquidity increases instantaneously. As for geographical differences, a de-mand shock is temporarily absorbed in market liquidity in urban areas.Prices adapt more gradually, and the effect is permanent. Following thisprice adjustment, liquidity reverts back close to its original level. In ru-ral areas the effect runs straight from market tightness to prices, but theincrease is still gradual as predicted by the search and matching model.The reason is that urban areas were recovering in the final year of thesample, whereas rural areas were not. This is also reflected in the model

116


estimated over the 50% regions with the most negative returns. Here,price discovery is found to be slower than in other regions which can beattributed to the low rate of sale in these regions. Theoretically, thiswould be reflected in sellers attaching more weight to their perceptionof market tightness yesterday than to signals of market tightness today.

5.3 How can market liquidity be used for a bet-ter understanding and monitoring of realestate market?

This thesis has shown that understanding market liquidity can be of vi-tal importance for policymakers and other market watchers. In general,liquidity can be useful for both monitoring and forecasting purposes.Additionally, the pro-cyclicality of liquidity is related the reservationprice behavior of buyers and sellers. Especially the latter may be re-lated to mortgage markets, which makes studying real estate liquidityparticularly interesting for policymakers.

Chapter 2 shows that the average TOM of sold properties –whichis frequently used by practitioners as market liquidity indicator– mightbe misleading as the average TOM only considers properties that havebeen sold. Furthermore, traded properties are heterogeneous. Sincethe probability to withdraw and housing characteristics differ over thecycle, the average TOM could paint the wrong picture. These problemsare more severe in small markets or markets where properties tradeinfrequently (i.e. during a crisis). The results show that simply takingthe mean of the TOM of sold properties underestimates (overestimates)market liquidity in good (bad) times. In other words, the quality of theproperties that are sold is different and/or the probability to withdrawis different.

Chapter 3 shows that tracking demand and supply reservationprices gives more insight into the real estate market. For policymakers,the presented method is interesting from both a monitoring and forecast-ing perspective. The comparison of the demand and supply reservationprice indices can provide a more extensive view of the current and histor-ical state of the market than by simply observing normal price indices.Because the demand indices provide a collapsed metric of both pricesand liquidity, they tend to lead price indices. As such, they can be usefulfor forecasting purposes of price indices. The lead-lag relationship be-

117


tween the demand and supply indices seems to be about a year. The lagof “normal” price indices to the demand indices is roughly two quarters.

Chapter 4 provides theoretical and empirical evidence on the linksbetween market tightness, market liquidity, and house prices. Under-standing these links can help policymakers in comprehending the rela-tionship between different indicators that policymakers tend to use toassess the market situation. Additionally, the chapter provides a novelindicator based on internet search behavior that is shown to lead otherindicators. This can be useful to forecast housing markets on a regionalscale. The lead-lag relationship between the market tightness indicatorand market liquidity is about one quarter, prices respond more gradualand half of the shock is absorbed after about two quarters.

118


119


120

Appendices

121

Appendix AAppendix for Chapter 2

A.1 Comparison with Carillo and Pope (2012)

This appendix offers a comparison of the presented methodology withindices produced by applying the non-parametric methodology of Car-rillo and Pope (2012) (henceforth CP). The CP-method is very usefulfor large markets and can be estimated quickly. However, in smallermarkets the method provides somewhat noisy estimates.

The methodology takes the Kaplan-Meier estimator as basis to es-timate the (empirical) cumulative distribution functions and survivorfunctions for each time-period (Kaplan and Meier, 1958). Propertiesthat are sold are treated as “failure” in Kaplan-Meier terminology andproperties that are withdrawn from the market or those that are still onthe market are treated as “censored”.

In order to create an index, quantiles of these distributions are linked.It might be illuminating to consider an example with two years. First,TOM distributions of listed properties for both years are estimated usingthe Kaplan-Meier estimator. Next, the median TOM of each distribu-tion is taken. The line between these medians represents a liquidityindex. The index can in principle be estimated for each point in thedistribution. Note that the Kaplan-Meier estimate of the distributiondoes not take individual property characteristics into account. There-fore, the distributions are also estimated using a weighted Kaplan-Meierestimate. These will subsequently result in constant-quality liquidityindices. CP propose an extenstion to the methodology of DiNardo et al.(1996) (DFL), in which the DFL-decomposition is combined with the

123

APPENDIX A. APPENDIX FOR CHAPTER 2

Kaplan-Meier estimator to allow it to work with censored variables.The weights of the weighted Kaplan-Meier estimator are based on thedegree of similarity between the property in question and properties inthe reference year. Intuitively speaking, more (less) weight is attachedto properties that are more (less) similar to properties in the referencequarter. The weights are estimated by estimating a logit on the proper-ties sold or withdrawn in the reference quarter and the properties soldor withdrawn in the quarter in question. The dependent variable in thislogit takes 1 if the property is sold or withdrawn in the reference quarter.The independent variables are housing characteristics.1 Therefore, thepredicted probabilities of this logit provide a measure of how similar aproperty in a given quarter is to properties from the reference quarter(a higher predicted value indicates more similarity).

When the reference quarter and the comparison quarter consistsof many observations, the estimated logit will provide sensible results.However, when there are few transactions, the logit is based on too fewobservations and the coefficients become unstable. Another differencewith the presented methodology is that the methodology of CP looks atthe ex ante distribution of the TOM. In other words, the average TOMof the quarter when the house was listed is calculated. This obviouslyresults in indices that are leading compared to indices based on the av-erage TOM of the quarter when the house was sold. However, this alsomeans that close the end of the sample, the data will only include prop-erties that are sold or withdrawn quickly. In quarter N-1 there will onlybe properties that are sold or withdrawn in quarter N-1 or N. Hence,the indices will be biased downwards if these are estimated in the finalyears of the sample.

A comparison between the indices based on the CP methodology andthe presented methodology in this study with random walk structure isincluded in Figure A.1. Note that the CP indices indicate the ex antesale probability, hence the the expected TOM of a house in the period thehouse was listed. The indices based on the methodology from this studyare based on the realized sale probability, or the TOM of the period

1The control variables are the same as in the presented Bayesian methodology:log size, log size squared, dummies for gardens, parking places, landleases, mainte-nance (bad, normal and well-maintained), construction period (before 1905, 1906-1944, 1945-1990, 1991-2000 after 2001), and property type (terraced, back-to-back,corner, semi-detached, detached, ground floor split level apartment, upper floor splitlevel, other apartment), and list price premium.

124


in which the house was sold or withdrawn. The obvious consequenceis that the CP indices seem to lead to indices based on the presentedmethodology for most of the sample, but the x-axes are not the same.

In the small market it proves to be rather problematic to estimateconstant-quality liquidity indices with the CP methodology. However,in the medium, and especially in the large market, the results are muchmore comparable (Figure A.1).

125


Figure A.1: Comparison between constant-quality random walk indices and Carilloand Pope (2012).


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

0

100

200

300

400

500

600

700

Period of sale in constant-quality, RW index / Period of listing in CP index

Constant-quality

liquid

ity

(2005Q1

=100)

Constant-quality, RW CP


2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

100

200

300

400

500

600


Mark

etilliquid

ity

(2005Q1=100)



2005

2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

50

100

150

200

250

300


Mark

etilliquid

ity

(2005Q1=100)


126


A.2 Estimated transaction price at time of en-try

To determine the expected transaction price at time of entry, a hedonicprice model is estimated. Transaction prices are modeled as a function ofhousing characteristics that are also included as controls in the liquidityindices. The model is estimated separately for each municipality andincludes dummies for the quarter of sale. The predicted values of thismodel (with the dummy for quarter of sale replaced by the value itentered the market) represent the estimated transaction price. Notethat it is also possible to estimate the expected transaction price ofhouses that are eventually withdrawn.

The estimated coefficients for each market are included in Table A.1.As the model is used for the calculation of one control variable only(list price premium), the coefficients are not discussed in detail. Mostcoefficients have the expected sign and are significant. Although theanalysis of this study focuses on houses sold or withdrawn between 2005and 2016, it might be the case the house was listed before this period.Therefore, in order to determine the market value at time of entry, thehedonic model is estimated using all observations some of which are alsosold prior to the period of interest. Therefore the recorded number ofobservations per municipality in this Appendix are somewhat differentthan those reported in Table 2.1.

127


Table A.1: Hedonic estimation of the coefficients on log transaction price.

Small Market Medium Market Large MarketVariable β t β t β tConstant 4.29 21.9 9.94 114.3 3.2 110.9Bad Maint. (Omitted) (Omitted) (Omitted)Normal Maint. 0.10 10.0 0.11 28.2 0.1 73.4Good Maint. 0.18 15.1 0.24 46.9 0.2 117.1< 1905 (Omitted) (Omitted) (Omitted)1906− 1944 -0.07 -1.5 0.22 4.8 0.0 -27.51945− 1990 -0.05 -1.2 -0.07 -1.6 -0.1 -44.01991− 2000 0.07 1.4 0.26 5.5 0.0 0.0> 2001 0.10 2.2 0.34 7.2 0.0 -7.2HT Terraced (Omitted) (Omitted) (Omitted)HT Back-to-Back 0.19 6.4 0.27 15.0 0.1 7.1HT Corner 0.05 8.1 0.06 12.8 0.0 9.0HT Semi-Detached 0.19 19.0 0.39 39.6 0.2 27.0HT Detached 0.36 27.0 0.72 47.3 0.3 33.0AT Split (GF) 0.05 3.4 -0.39 -41.5 -0.1 -49.1AT Split (UF) 0.00 -0.1 -0.41 -45.2 -0.1 -39.3AT Other -0.03 -1.8 -0.44 -57.9 -0.1 -51.6log(size) 1.76 33.7 0.31 17.0 2.0 412.9log(size)2 -0.10 -32.2 -0.02 -16.7 -0.1 -383.8Garden 0.01 0.9 0.00 0.6 0.1 51.5Parking 0.10 13.6 0.16 35.1 0.1 46.2Landlease -0.04 -1.0 -0.58 -11.0 0.0 -9.3Market conditions Quarter of sale dummiesLocation ZIP-code dummiesObservations 3,667 21,405 143,299R2 0.9143 0.9096 0.9243RMSE 0.1477 0.1943 0.1816

HT = House type, AT = Apartment type, GF = Ground floor, UF, Upper floor.

A.3 Transaction prices indices

The transaction price indices are estimated using a hedonic model. Morespecifically, the indices are estimated using a Hierarchical Trend Model(HTM) (Francke and De Vos, 2000; Francke and Vos, 2004). This modelis well-suited to estimate constant-quality price indices in thin marketsand is also used in Van Dijk and Francke (2018) to estimate quarterlytransaction price indices in the Netherlands. The hedonic regression isperformed for the COROP-region in which the three municipalities arelocated.2 The common COROP-trend is modeled as local linear trendand the municipal trends are modeled as a random walk. The HTM isdefined as (Francke and Vos, 2004):

2A COROP-region is the Dutch equivalent of an MSA, it is however much smallerin geographical and population size.

128


yt = iµt +Dϑ,tθt +Xtβ + εt, εt ∼ N(0, σ2εI), (A.1a)

µt+1 = µt + κt + ηt, ηt ∼ N(0, σ2η), (A.1b)

κt+1 = κt + ζt, ζt ∼ N(0, σ2ζ ), (A.1c)

θt+1 = θt +$t, $t ∼ N(0, σ2$I). (A.1d)

Here yt is a vector of log selling prices. Next, µt is the common trendof the COROP-region, vector θt contains the municipal-specific trends,and matrix D is a selection matrix of the municipality. Finally, Xt is avector with house characteristics with coefficients β.

Results of coefficient estimates of the housing characteristics areshown in first two columns of Table A.2. As this estimation is onlyperformed to calculate the price indices that are used in one subsection,the coefficients are not discussed in detail. All coefficients, however,have the expected signs and the fit is satisfactory. The transaction priceindices for the three municipalities are the sum of the common trend(of the COROP-region) and the municipal trend. These indices are pre-sented in Figure 2.4.

129


Table A.2: Hedonic (HTM) estimation of the coefficients on log transaction price.

Transaction priceVariable β t

Bad Maint. (Omitted)Normal Maint. 0.097 45.0Good Maint. 0.194 77.5< 1905 (Omitted)1906− 1944 -0.117 52.81945− 1990 -0.388 155.71991− 2000 -0.235 81.2> 2001 -0.252 79.9HT Terraced (Omitted)HT Back-to-Back 0.130 17.5HT Corner 0.022 7.6HT Semi-Detached 0.132 33.8HT Detached 0.231 46.1AT Split-Level (Ground or multiple) 0.113 40.7AT Split-Level (Upper floor) 0.117 37.1AT Other 0.025 7.9log(size) 2.148 298.1log(size)2 -0.120 215.2Garden 0.067 26.4Parking 0.067 33.4Landlease -0.110 64.8Market conditions Common trend (Local Linear Trend)Location Muncipal trends (Random Walk)Observations 132243R2 0.7871RMSE 0.2239

HT = House type, AT = Apartment type

130

Appendix BAppendix for Chapter 4

B.1 Theoretical framework

This appendix offers a theoretical model that links increases (decreases)in market tightness to a subsequent increase (decrease) in market liquid-ity and prices. Table 4.1 shows the link between the symbols used in thetheoretical and empirical framework. The model further aims to provideinsights into differences in the temporal dynamics. We closely follow themodel of Carrillo et al. (2015), which in turn is based on standard searchand matching models in the housing market (Yinger, 1981; Novy-Marx,2009; Genesove and Han, 2012). Unlike standard search and matchingmodels but in line with Carrillo et al. (2015) and Berkovec and Good-man (1996), we allow for a lagged response of buyers and sellers dueto imperfect information. We differ from Carrillo et al. (2015) with re-gard to the following aspects. In the model of Carrillo et al. (2015) thenumber of sellers is constant over time. We generalize this and allowfor different number of sellers in each period. Thereby we can link themodel more closely to our empirical setup, in which both the numberof clicks and the number of houses listed online vary in each quarter.Furthermore, we introduce different parametric assumptions.

We start by defining market tightness, λ for each period t:

λt =mb,t

ms,t. (B.1)

Here mb and ms are the number of buyers and sellers at a given pointin time t, respectively. A higher (lower) λ indicates more (fewer) buyers

131

APPENDIX B. APPENDIX FOR CHAPTER 4

relative to sellers, and this in turn increases (decreases) the probabilitythat a seller meets a potential buyer. Conversely, for buyers a higher(lower) λ results in a lower (higher) probability to find a suitable seller.We denote these probabilities qb(λ) for buyers and qs(λ) for sellers and letqs(λ) = exp(−ρs/λ) and qb(λ) = 1−exp(−ρb/λ). Here, scalars ρs, ρb > 0measure how market tightness affects the matching probabilities. Hence,∂qs∂λ = exp(−ρs/λ)/λ2 > 0 and ∂qb

∂λ = − exp(−ρb/λ)/λ2 < 0, so a highermarket tightness results in a higher matching probability for sellers anda lower matching probability for buyers.

In each period, agents choose to transact or continue their search. Atransaction will occur if the gain from a transaction for both parties islarger than zero. Let ε be the offer price. Following Carrillo et al. (2015),we assume this value to be independent and identically distributed.Next, let the value of search be Vb,t and Vs,t for the buyer and seller,respectively. Then a transaction will occur only if ε > Vb,t + Vs,t = ε,that is, when the offer ε exceeds the reservation price of the seller ε. Forthe seller, the value of search is equal to the expected gain of a matchtimes the probability of a match plus the complement of the probabilityof finding a match times the discounted search value of the next period.

Vs,t =qs

1 + rE[max(Vs,t + θ(ε− Vb,t − Vs,t), Vs,t+1)− cs]

+1− qs1 + r

Vs,t+1,(B.2)

Vb,t =qb

1 + rE[max(Vb,t + (1− θ)(ε− Vb,t − Vs,t), Vb,t+1)− cb]

+1− qb1 + r

Vb,t+1.(B.3)

Here, r is the interest rate agents use to discount future gains or costs.The bargaining power affects the size of the potential gain for the buyerand seller and is denoted by θ(λ) ∈ [0, 1]. Following Carrillo et al.(2015), we assume this to be positively (negatively) related to λ for theseller (buyer): θ(λ) = k exp(λ)/(1 + k exp(λ)), where k > 0 is a scalarto measure how bargaining power responds to λ. We assume a Log-normal distribution of ε, since the value of a match cannot be negative(i.e. transaction prices cannot be negative). Furthermore, in our sam-ple transaction prices prove to be Log-normally distributed rather than

132


normally distributed. Finally, cs and cb are fixed costs for the seller andbuyer.

In the steady state we can rewrite Equations (B.2) and (B.3) as:

Vs =qsr

(θE[max(ε− Vb − Vs, 0)]− cs), (B.4)

Vb =qbr

((1− θ)E[max(ε− Vb − Vs, 0)]− cb). (B.5)

A seller will accept a match if ε exceeds a certain reservation thresholdε, hence a match occurs if ε > Vb + Vs = ε. By rewriting the latterequality we obtain:

rε = (qsθ + qb(1− θ))E[max(ε− ε, 0)]− (qbcb + qscs). (B.6)

Next, we allow for information asymmetries. Genesove and Han (2012)show that sellers react to a demand shock with a lag. If, for example,more buyers enter the market, this is not directly observed by all sellers.Therefore, in their eyes, market tightness is lower than the true markettightness. As true market tightness increases, agents gradually adjusttheir expectations regarding market tightness:

λ∗t = αλ∗t−1 + (1− α)λ, (B.7)

where, α ∈ [0, 1]. Hence, the perceived market tightness is a weightedcombination of the lagged perceived market tightness and the currenttrue market tightness.

If fixed matching costs are small relative to the potential benefitsand since the term E[max(ε− ε, 0)] =

∫∞ε (ε− ε)dF (ε) is decreasing in ε

there exists a unique solution (Carrillo et al., 2015; Novy-Marx, 2009).Assuming a Log-normal distribution of ε, we can rewrite this term:

E[max(ε− ε, 0)] = E(ε)N(d1)− εN(d2)

= exp(µ+ σ2/2)N(d1)− εN(d2),(B.8)

where

d1 =log[exp(µ+ σ2/2)/ε] + σ2/2

σ,

d2 =log[exp(µ+ σ2/2)/ε]− σ2/2

σ,

133


and N(d) denotes the cumulative normal density function, µ is the meanand σ the standard deviation of ε.

Note that sellers set their reservation prices ε based on matchingprobabilities and bargaining power, which in turn depend on perceivedmarket tightness rather than on true market tightness. Hence, pricesdepend on perceived market tightness only. We follow Carrillo et al.(2015) in the specification of the probability that a seller finds a suc-cessful match in a given period:

ωt = qs(λ)[1−G(ε(λ∗t ))], (B.9)

here G denotes the cumulative density function. Note that part of thematching probability depends on true market tightness (λ) and partdepends on perceived market tightness (λ∗).

Since the matching probability depends on both true and perceivedmarket tightness, market liquidity responds directly to a change in mar-ket tightness. Prices, however, depend on perceived market tightnessonly and change more gradually.

We calibrated the parameters as follows: The average transactionprice in our sample is e 250,000 and the standard deviation is e 160,000.This corresponds to a µ of 12.258 and a σ of 0.586.1 Other calibratedparameters are cs = cb =e 10,000, r = 0.0847%, ρs = 1, ρb = 0.1,k = 0.5 and α = 0.75.2

B.2 House price index and rate of sale estima-tion

The house price index is estimated using a Hierarchical Trend Model(HTM). The HTM is estimated recursively over 40 COROP regions inthe Netherlands to allow for different effects of house characteristicson prices in each region. To estimate a quarterly price index at themunicipal level, each COROP region is subdivided into municipalities

1The Log-normal parameters are related to a set of normal parameters by: µ =

ln

(m√

1+ vm2

)and σ =

√ln(1 + v

m2

), where m and v are the mean and standard

deviation of a non-logarithmized sample, respectively.2The mortgage rate r corresponds to an annual rate of 4.5%, which is the average

mortgage rate in the sample.

134


each containing their own trend which is modeled as a random walk. TheCOROP region trend is modeled as a local linear trend. By summing themunicipal trend and the COROP region trend, the quarterly price indexof 403 municipalities is estimated. The HTM is defined in Equations(B.10a) - (B.10d) (Francke and Vos, 2004):

yt = iµt +Dϑ,tθt +Xtβ + εt, εt ∼ N(0, σ2εI), (B.10a)

µt+1 = µt + κt + ηt, ηt ∼ N(0, σ2η), (B.10b)

κt+1 = κt + ζt, ζt ∼ N(0, σ2ζ ), (B.10c)

θt+1 = θt +$t, $t ∼ N(0, σ2$I). (B.10d)

Here yt is a vector of log selling prices within a COROP region, µt isthe COROP trend, and vector θt contains the municipal-specific trends.Furthermore, matrix D is a selection matrix to select the municipality inwhich the transaction has taken place. Finally, Xt is a vector containinghouse characteristics with the estimated coefficients β.

Table B.1 shows summary statistics of the estimates for all COROPregions. Additionally, results for one COROP region (COROP region23 Amsterdam region) are presented. Table B.2 presents the estimatedcoefficients for this region, and panel (A) of Figure B.1 presents theestimated price index for six municipalities within this COROP region.

Finally, the rate of sale is estimated by dividing the number of trans-actions by the houses for sale at the beginning of the quarter. The es-timated rate of sale for the municipalities within the Amsterdam regionare depicted in panel (B) of Figure B.1.

The rate of sale and clicks per house series are the trend compo-nents from unobserved components models (Local Level). This processis depicted in Equations (B.11a) - (B.11c) in which yt is the observa-tion vector, µt is the trend component (i.e. smoothed series), and γt isa stochastic seasonal component. Moreover, dummies for 2012Q4 and2013Q1 are included in these unobserved component models as there wasa sudden increase and subsequent drop in these quarters due to the abol-ishment of the tax-deductibility of interest payments under interest-onlymortgages.

135


yt = µt + γt + εt, εt ∼ N(0, σ2εI), (B.11a)

µt+1 = µt + ηt, ηt ∼ N(0, σ2η), (B.11b)

γt = −s−3∑j=1

γt−j + ζt, ζt ∼ N(0, σ2ζ ). (B.11c)

136


Table B.1: Summary statistics of HTM estimations of the price index.

COROP R2 RMSE N1 0.818 0.177 11,0972 0.813 0.176 4,5613 0.834 0.178 31,4104 0.844 0.172 23,8445 0.845 0.177 11,5856 0.875 0.147 16,8407 0.852 0.157 20,8088 0.826 0.167 15,9799 0.847 0.146 12,06410 0.850 0.140 28,96011 0.854 0.150 13,48012 0.859 0.152 42,32213 0.855 0.149 54,03314 0.868 0.149 27,55215 0.846 0.160 55,82816 0.877 0.138 18,13017 0.861 0.160 107,48218 0.839 0.166 21,77119 0.888 0.143 25,25520 0.857 0.150 15,87121 0.882 0.191 22,01922 0.848 0.134 13,35023 0.842 0.175 61,61824 0.900 0.177 26,19425 0.878 0.145 30,55126 0.882 0.197 43,68927 0.860 0.137 12,61828 0.883 0.136 25,48629 0.841 0.169 72,99230 0.869 0.143 25,47231 0.860 0.190 7,15732 0.864 0.170 14,64533 0.848 0.160 45,38334 0.876 0.144 37,38835 0.867 0.140 43,74436 0.876 0.141 50,90737 0.864 0.143 8,17138 0.830 0.152 12,14539 0.833 0.181 18,50240 0.869 0.123 37,110Average 0.857 0.158 29,200

The R2, RMSE and N denote the R-squared Root Mean Squared Error and the number of observations ofthe HTM in the respective COROP region.

137


Table B.2: Estimates of the coefficients of house characteristics on the log of trans-action price in COROP region 23 (Amsterdam region).

Dep var: Log transaction priceVariable Beta T-stat

Log house size smaller than 350m3 0.655 144.2Log house size 350 - 500m3 0.030 57.8Log house size larger than 500m3 0.056 80.8Log lot size smaller than 500m3 0.103 48.9Log lot size 500 - 1500m3 0.001 1.5Log lot size larger than 1500m3 0.001 0.5Number of rooms 0.027 41.2Built before 1905 0.338 93.8Built 1906 - 1930 0.173 57.1Built 1931 - 1944 0.149 39.5Built 1945 - 1959 0.046 13.3Built 1960 - 1970 -0.031 11.7Built 1971 - 1980 -0.045 16.5Built 1981 - 1990 (omitted) (omitted)Built 1991 - 2000 0.071 28.5Built after 2000 0.087 23.1HT Simple -0.029 5.9HT Single-family (omitted) (omitted)HT Canal House 0.453 55.2HT Mansion 0.153 53.3HT Living Farm 0.135 10.0HT Bungalow 0.266 42.5HT Villa 0.310 61.6HT Manor 0.338 23.0HT Estate 0.399 3.1HT Ground floor ap. 0.134 28.6HT Top floor ap. 0.086 25.5HT Multiple level ap. 0.054 7.7HT ap. w/porch 0.077 15.7HT ap. w/gallery -0.017 2.9HT Nursing home -1.091 31.4HT Top and ground floor ap. 0.173 11.2Very poor maintenance -0.230 16.3Very poor to poor maintenance -0.208 7.6Poor maintenance -0.145 25.4Poor to average maintenance -0.166 13.2Average maintenance -0.094 32.4Average to good maintenance -0.081 16.6Good maintenance (omitted) (omitted)Good to excellent maintenance 0.088 17.4Excellent maintenance 0.084 40.0No parking (omitted) (omitted)Parking 0.068 34.5Market conditions Common trend (Local Linear Trend)Location Municipal trends (Random Walk)R2 0.842RMSE 0.175Observations 61,618

HT = House type, ap. = apartment

138


Figure B.1: Price indices (A), rate of sale (B) and clicks per house (C) of six munic-ipalities within COROP region 23 (Amsterdam region).

(A)

2011

2012

2013

2014

80

85

90

95

100

House

pricein

dex

Aalsmeer Amstelveen

Amsterdam Graft − De Rijp

Beemster Diemen

(B)

2011

2012

2013

2014

0

0.02

0.04

0.06

0.08

0.1

0.12

Rate

ofsa

le

Aalsmeer Amstelveen


Beemster Diemen

(C)

2011

2012

2013

2014

100

200

300

400

500

Clicksperhouse

Aalsmeer Amstelveen


Beemster Diemen

139


140

Bibliography

Acharya, V. V., Y. Amihud, and S. T. Bharath (2013). Liquidity risk ofcorporate bond returns: conditional approach. Journal of FinancialEconomics 110 (2), 358–386.

Acharya, V. V. and L. H. Pedersen (2005). Asset pricing with liquidityrisk. Journal of Financial Economics 77 (2), 375–410.

Almeida, H., M. Campello, and C. Liu (2006). The financial acceler-ator: Evidence from international housing markets. Review of Fi-nance 10 (3), 321–352.

Ametefe, F., S. Devaney, and G. Marcato (2016). Liquidity: A reviewof dimensions, causes, measures, and empirical applications in realestate markets. Journal of Real Estate Literature 24 (1), 1–29.

Andrews, D. and B. Lu (2001). Consistent model and moment selectionprocedures for GMM estimation with application to dynamic paneldata models. Journal of Econometrics 101 (1), 123–164.

Arellano, M. and S. Bond (1991). Some tests of specification for paneldata: Monte Carlo evidence and an application to employment equa-tions. The Review of Economic Studies 58 (2), 277–297.

Askitas, N. and K. Zimmerman (2009). Google econometrics and unem-ployment forecasting. Applied Economics Quarterly 55 (2), 107–120.

Bailey, M. J., R. F. Muth, and H. O. Nourse (1963). A regression methodfor real estate price index construction. Journal of the American Sta-tistical Association 58, 933–942.

141

BIBLIOGRAPHY

Belkin, J., D. J. Hempel, and D. W. McLeavey (1976). An empiri-cal study of time on market using multidimensional segmentation ofhousing markets. Real Estate Economics 4 (2), 57–75.

Berkovec, J. and J. Goodman (1996). Turnover as a measure of demandfor existing homes. Real Estate Economics 24 (4), 421–440.

Betancourt, M. and M. Girolami (2015). Hamiltonian Monte Carlo forhierarchical models. Current trends in Bayesian methodology withapplications 79, 30.

Bokhari, S. and D. Geltner (2011). Loss aversion and anchoring incommercial real estate pricing: Empirical evidence and price indeximplications. Real Estate Economics 39 (4), 635–670.

Brunnermeier, M. and L. Pedersen (2009). Market liquidity and fundingliquidity. Review of Financial Studies 22 (6), 2201–2238.

Cameron, A. C. and P. K. Trivedi (2005). Microeconometrics: Methodsand Applications. Cambridge University Press.

Capozza, D., P. Hendershott, and C. Mack (2004). An anatomy ofprice dynamics in illiquid markets: Analysis and evidence from localhousing markets. Real Estate Economics 32 (1), 1–32.

Carpenter, B., A. Gelman, M. Hoffman, D. Lee, B. Goodrich, M. Be-tancourt, M. A. Brubaker, J. Guo, P. Li, and A. Riddell (2016). Stan:A probabilistic programming language. Journal of Statistical Soft-ware 20, 1–43.

Carrillo, P., E. De Wit, and W. Larson (2015). Can tightness in the hous-ing market help predict subsequent home price appreciation? RealEstate Economics 43 (3), 609–651.

Carrillo, P. and B. Williams (2015). The repeat time-on-the-marketindex. SSRN Working Paper 2584558 .

Carrillo, P. E. (2013). To sell or not to sell: Measuring the heat of thehousing market. Real Estate Economics 41 (2), 310–346.

Carrillo, P. E. and J. C. Pope (2012). Are homes hot or cold potatoes?the distribution of marketing time in the housing market. RegionalScience and Urban Economics 42 (1), 189–197.

142

BIBLIOGRAPHY

Case, K. and R. Shiller (1990). Forecasting prices and excess returns inthe housing market. Real Estate Economics 18 (3), 253–273.

Clapp, J. M., R. Lu-Andrews, and T. Zhou (2017). Is the behaviorof sellers with expected gains and losses relevant to cycles in houseprices? SSRN Working paper 2877546 .

Clayton, J., N. Miller, and L. Peng (2010). Price-volume correlationin the housing market: Causality and co-movements. The Journal ofReal Estate Finance and Economics 40 (1), 14–40.

Conclusr (2014). Makelaardijmonitor 2014: Benchmark huizensites [InDutch]. Conclusr Research Report .

Dari-Mattiacci, G., O. Gelderblom, J. Jonker, and E. C. Perotti (2017).The emergence of the corporate form. The Journal of Law, Economics,and Organization 33 (2), 193–236.

De Wit, E., P. Englund, and M. Francke (2013). Price and transactionvolume in the Dutch housing market. Regional Science and UrbanEconomics 43 (2), 220–241.

De Wit, E. R. and B. Van der Klaauw (2013). Asymmetric informationand list-price reductions in the housing market. Regional Science andUrban Economics 43 (3), 507–520.

Dellaportas, P. and A. F. Smith (1993). Bayesian inference for gen-eralized linear and proportional hazards models via Gibbs sampling.Applied Statistics 42 (3), 443–459.

DiNardo, J., N. M. Fortin, and T. Lemieux (1996). Labor market insti-tutions and the distribution of wages, 1973-1992: A semiparametricapproach. Econometrica 64 (5), 1001–1044.

DNB (2016). Overzicht Financiele Stabiliteit Herfst 2016 [In Dutch].

Dube, J. and D. Legros (2016). A spatiotemporal solution for the si-multaneous sale price and time-on-the-market problem. Real EstateEconomics 44 (4), 846–877.

Duca, J., J. Muellbauer, and A. Murphy (2011). House prices and creditconstraints: Making sense of the US experience. The Economic Jour-nal 121 (552), 533–551.

143

BIBLIOGRAPHY

Fisher, J., D. Gatzlaff, D. Geltner, and D. Haurin (2003). Controllingfor the impact of variable liquidity in commercial real estate priceindices. Real Estate Economics 31 (2), 269–303.

Fisher, J., D. Geltner, and H. Pollakowski (2007). A quarterlytransactions-based index of institutional real estate investment per-formance and movements in supply and demand. The Journal of RealEstate Finance and Economics 34 (1), 5–33.

Francke, Marc Geltner, D., A. Van de Minne, and B. White (2017). Revi-sions in granular repeat sales indices. SSRN Working paper 3016563 .

Francke, M. and A. De Vos (2000). Efficient computation of hierarchicaltrends. Journal of Business & Economic Statistics 18 (1), 51–57.

Francke, M. and K. Lee (2013). De waardeontwikkeling op de woning-markt in aardbevingsgevoelige gebieden rond het Groningenveld [InDutch]. Ortec Finance Research Report .

Francke, M. and G. Vos (2004). The hierarchical trend model for prop-erty valuation and local price indices. The Journal of Real EstateFinance and Economics 28 (2), 179–208.

Francke, M. K. (2010). Repeat sales index for thin markets. The Journalof Real Estate Finance and Economics 41 (1), 24–52.

Francke, M. K. and A. van de Minne (2017). The hierarchical repeatsales model for granular commercial real estate and residential priceindices. The Journal of Real Estate Finance and Economics 55 (4),511–532.

Francke, M. K., A. M. Van de Minne, and J. Verbruggen (2014). Theeffect of credit conditions on the Dutch housing market. DNB WorkingPaper No. 447 .

Gatzlaff, D. H. and D. R. Haurin (1997). Sample selection bias andrepeat-sales index estimates. The Journal of Real Estate Finance andEconomics 14 (1), 33–50.

Geltner, D., A. Kumar, and A. van de Minne (2017). Riskiness of realestate development: A perspective from urban economics & optionvalue theory. SSRN Working paper 2907036 .

144

BIBLIOGRAPHY

Genesove, D. and L. Han (2012). Search and matching in the housingmarket. Journal of Urban Economics 72 (1), 31–45.

Genesove, D. and C. Mayer (2001). Loss aversion and seller behavior:Evidence from the housing market. The Quarterly Journal of Eco-nomics 116 (4), 1233–1260.

Glower, M., D. Haurin, and P. Hendershott (1998). Selling time andselling price: The influence of seller motivation. Real Estate Eco-nomics 26 (4), 719–740.

Goetzmann, W. and L. Peng (2006). Estimating house price indexesin the presence of seller reservation prices. Review of Economics andStatistics 88 (1), 100–112.

Haurin, D. (1988). The duration of marketing time of residential hous-ing. Real Estate Economics 16 (4), 396–410.

Heckman, J. J. (1979). Sample selection bias as a specification error.Econometrica: Journal of the Econometric Society , 153–161.

Heidelberger, P. and P. D. Welch (1981). A spectral method for confi-dence interval generation and run length control in simulations. Com-munications of the ACM 24 (4), 233–245.

Hekwolter of Hekhuis, M., R. Nijskens, and W. Heeringa (2017). Dewoningmarkt in de grote steden [In Dutch]. DNB Occasional Stud-ies 15 (1).

Hoffman, M. D. and A. Gelman (2014). The No-U-turn sampler: Adap-tively setting path lengths in Hamiltonian Monte Carlo. Journal ofMachine Learning Research 15 (1), 1593–1623.

Hort, K. (2000). Prices and turnover in the market for owner-occupiedhomes. Regional Science and Urban Economics 30 (1), 99–119.

Jenkins, S. P. (2005). Survival analysis. Unpublished manuscript, Insti-tute for Social and Economic Research, University of Essex, Colch-ester, UK 42, 54–56.

Johnson, N. and S. Kotz (1972). Distributions in Statistics: ContinuousMultivariate Distributions. Wiley: New York, NY .

145

BIBLIOGRAPHY

Jud, D., T. Seaks, and D. Winkler (1996). Time on the market: The im-pact of residential brokerage. Journal of Real Estate Research 12 (2),447–458.

Kang, H. and M. Gardner (1989). Selling price and marketing time in theresidential real estate market. Journal of Real Estate Research 4 (1),21–35.

Kaplan, E. L. and P. Meier (1958). Nonparametric estimation fromincomplete observations. Journal of the American Statistical Associ-ation 53 (282), 457–481.

Kerste, M., B. Baarsma, N. Roosenboom, and P. Risseeuw (2012).Ongezien, onverkocht? [In Dutch]. SEO Research Report 2012-08.

Knight, J. (2002). Listing price, time on market, and ultimate sellingprice: Causes and effects of listing price changes. Real Estate Eco-nomics 30 (2), 213–237.

Koehler, E., E. Brown, and S. J.-P. Haneuse (2009). On the assess-ment of monte carlo error in simulation-based statistical analyses.The American Statistician 63 (2), 155–162.

Koster, H. and J. Rouwendal (2017). Verkopen boven de vraagprijs-buitensporig of nieuwe werkelijkheid? [In Dutch]. ASRE ResearchPapers 2017 (05).

Lee, K. (2014). Why do renters stay in or leave certain neighborhoods?The role of neighborhood characteristics, housing tenure transitions,and race. Journal of Regional Science 54 (5), 755–787.

Lee, K. O. and M. Mori (2016). Do conspicuous consumers pay higherhousing premiums? spatial and temporal variation in the unitedstates. Real Estate Economics 44 (3), 726–763.

Love, I. and L. Zicchino (2006). Financial development and dynamic in-vestment behavior: Evidence from panel VAR. The Quarterly Reviewof Economics and Finance 46 (2), 190–210.

Maddala, G. and S. Wu (1999). A comparative study of unit root testswith panel data and a new simple test. Oxford Bulletin of Economicsand Statistics 61 (S1), 631–652.

146

BIBLIOGRAPHY

Mian, A. and A. Sufi (2009). The consequences of mortgage credit ex-pansion: Evidence from the US mortgage default crisis. The QuarterlyJournal of Economics 124 (4), 1449–1496.

Miller, N. and M. Sklarz (1986). A note on leading indicators of housingmarket price trends. Journal of Real Estate Research 1 (1), 99–109.

Nickell, S. (1981). Biases in dynamic models with fixed effects. Econo-metrica 49 (6), 1417–1426.

Novy-Marx, R. (2009). Hot and cold markets. Real Estate Eco-nomics 37 (1), 1–22.

NVM (2016). Analyse Woningmarkt Q1 2016 [In Dutch].

Real Capital Analytics (2017). USCapital trends July 2017.

Stein, J. (1995). Prices and trading volume in the housing market:A model with downpayment effects. The Quarterly Journal of Eco-nomics 110 (2), 379–406.

Van der Cruijsen, C., D.-J. Jansen, and M. van Rooij (2018). The rose-tinted spectacles of homeowners. Journal of Consumer Affairs 52 (1),61–87.

Van Dijk, D., D. Geltner, and A. van de Minne (2018). Revisiting supplyand demand indexes in real estate. DNB Working paper 583.

Van Dijk, D. W. and M. K. Francke (2018). Internet search behavior, liq-uidity and prices in the housing market. Real Estate Economics 46 (2),368–403.

Vehtari, A., A. Gelman, and J. Gabry (2017). Practical Bayesian modelevaluation using leave-one-out cross-validation and WAIC. Statisticsand Computing 27 (5), 1413–1432.

Watanabe, S. (2010). Asymptotic equivalence of Bayes cross validationand widely applicable information criterion in singular learning theory.Journal of Machine Learning Research 11 (Dec), 3571–3594.

Wheaton, W. C. (1990). Vacancy, search, and prices in a housing marketmatching model. Journal of Political Economy 98 (6), 1270–1292.

147

BIBLIOGRAPHY

Wu and E. Brynjolfsson (2014). The future of prediction: HowGoogle searches foreshadow housing prices and sales. In A. Gold-farb, S. Greenstein, and C. Tucker (Eds.), Economic Analysis of theDigital Economy, pp. 89 – 118. University of Chicago Press.

Wu and Y. Deng (2015). Intercity information diffusion and price discov-ery in housing markets: Evidence from Google searches. The Journalof Real Estate Finance and Economics 50 (3), 289–306.

Yavas, A. and S. Yang (1995). The strategic role of listing price in mar-keting real estate: Theory and evidence. Real Estate Economics 23 (3),347–368.

Yinger, J. (1981). A search model of real estate broker behavior. TheAmerican Economic Review 71 (4), 591–605.

148

Nederlandse samenvatting(Summary in Dutch)

Marktliquiditeit in residentieel en commercieel vastgoed

De hoofdvraag in dit proefschrift is “Wat is de rol van markt-liquiditeit in vastgoed?”. Marktliquiditeit is sterk gerelateerd aanprijsdynamiek in vastgoed. Daarom dienen prijzen en liquiditeit niet alsgescheiden concepten te worden gezien. Dit proefschrift biedt een aan-tal nieuwe maatstaven voor het meten van marktliquiditeit in vastgoed.De ontwikkeling in alle maatstaven laat een sterke cyclische samenhangzien met prijsontwikkelingen. Dit toont aan dat beleidsmakers en an-dere marktvolgers niet louter prijsontwikkelingen in acht moeten nemen.Een completer beeld ontstaat als zowel liquiditeits- als prijsontwikkelin-gen worden geanalyseerd. Daarbij is van belang dat liquiditeitsdynamiekleidend is voor prijsdynamiek. Deze inzichten kunnen bijvoorbeeld vanbelang zijn voor beleidsmakers in het duiden van crises en oververhittemarkten.

De hoofdvraag is onderverdeeld in drie deelvragen, waarvan de eer-ste is: “Hoe kan marktliquiditeit in vastgoed worden gemeten?”Drie hoofdstukken in deze thesis gaan hier op in. Hoofdstuk 2 biedteen manier om marktliquiditeitsindices gebaseerd op constante kwaliteitvan de woningen te schatten als transactiedata schaars is. De maatstafhoudt rekening met het feit dat de kwaliteit van verkochte woningenniet constant is over de tijd. Daarnaast wordt in acht genomen dat insommige periodes meer of minder woningen uit de verkoop worden te-ruggetrokken. Een van de belangrijkste voordelen van de methodologieis dat deze ook gebruikt kan worden in dunne markten (dat wil zeg-

149

NEDERLANDSE SAMENVATTING (SUMMARY IN DUTCH)

gen, in markten waarin weinig transacties plaatsvinden). Indices vanmarktliquiditeit gemeten op basis van conventionele methodes kunnennamelijk onbetrouwbaar zijn in dunne markten. De ontwikkelde me-thode introduceert een structuur op de geschatte coefficienten waardoorer ook betrouwbare indices kunnen worden geschat voor deze markten.Daarnaast houdt de methode rekening met verschillen in kwaliteit enhet aantal terugtrekkingen. De resultaten suggereren dat terugtrekkin-gen de belangrijkste oorzaak zijn van de verschillen tussen gemiddeldeverkooptijd van verkochte woningen en de ontwikkelde maatstaf. Daar-naast zorgt kwaliteit voor significante verschillen.

Hoofdstuk 3 breidt de methode van Fisher et al. (2003) voor hetbepalen van constante liquiditeits prijsindices verder uit en wel op tweemanieren: (i) door het plaatsen van de methode in een herhaalde verko-pen raamwerk en (ii) door het schatten in een structureel tijdreeksmo-del. Door deze uitbreidingen is het mogelijk om reserveringsprijzen vankopers en verkopers voor commercieel vastgoed op een regionaal (ste-delijk) niveau te bepalen, zonder dat een uitgebreide verzameling vanpandkarakteristieken voorhanden dient te zijn. Het verschil tussen decentrummaten van de reserveringsprijzen van kopers en verkopers is eenmaatstaf voor liquiditeit. Dit kan worden gezien als een bid-ask spreadvoor de vastgoedmarkt. De ontwikkelde maatstaf is een zeer algemenemaatstaf die geschat kan worden met iedere database waarin herhaaldeverkopen (inclusief de transactieprijzen en transactiedata) kunnen wor-den geıdentificeerd. Omdat de maatstaf in een structureel tijdreeksmo-del is geplaatst, kan deze ook worden gebruikt in dunne markten.

Hoofdstuk 4 laat zien dat een maatstaf voor krapte op de woning-markt op basis van zoekgedrag op internet (clicks per woning) leidendis voor een maatstaf van marktliquiditeit op basis van transactiedata.De achterliggende reden is dat zoekgedrag op internet een signaal is vanvoorbereidende stappen die uiteindelijk kunnen leiden tot de koop vaneen huis. Het effect gaat echter twee kanten op: internetzoekgedrag re-ageert negatief op veranderingen in marktliquiditeit. De reden is dateen stijging van het aantal verkochte woningen in het afgelopen kwar-taal reflecteert dat er minder woningzoekers over zijn in een bepaaldegemeente in het huidige kwartaal.

De tweede deelvraag is: “Hoe zijn prijzen en liquiditeit gere-lateerd in vastgoedmarkten?” De liquiditeitsindices gebaseerd opconstante kwaliteit van de woningen voor de Nederlandse woningmarktuit Hoofdstuk 2 laten zien dat marktliquiditeit laag is tijdens een cri-

150


sis. De ontwikkelde liquiditeitsmaatstaf laat verder zien dat er sterkeovereenkomsten bestaan tussen prijsontwikkelingen in ontwikkelingen enliquiditeit. Een Granger causaliteit analyse laat verder zien dat veran-deringen in liquiditeit eerder plaatsvinden dan veranderingen in prijzen.Een nieuwe bevinding, consistent met de asset pricing literatuur, laatzien dat het liquiditeitsrisico tevens hoger is tijdens een crisis. Hoofd-stuk 2 gaat ook in op de verhouding tussen de vraagprijspremie en deverkooptijd. Een hogere premie (dat wil zeggen een hogere vraagprijsten opzichte van de geschatte marktwaarde) is gerelateerd aan een klei-nere verkoopkans. Het effect is niet constant over de tijd: tijdens decrisis is zowel de vraagprijspremie als het totale effect op de verkoop-kans groter (de gemiddelde vraagprijspremie vermenigvuldigd met decoefficient). Recentelijk (sinds 2015) is de vraagprijspremie gemiddeldgezien veranderd in een vraagprijskorting. De reden is dat verkopershun gedrag veranderen door de extreme krapte op de markt.

Hoofdstuk 3 laat zien dat liquiditeit in alle grote commerciele vast-goedmarkten meestal procyclisch is. Omdat liquiditeit niet constant isover de tijd, vergelijkt een normale prijsindex, tot op zekere hoogte,appels met peren. Dit hoofdstuk ontwikkelt een methode om prijsindi-ces gebaseerd op herhaalde verkopen te corrigeren voor tijdsvarierendeliquiditeit. Deze constante liquiditeit prijsindices (reserveringsprijsin-dices van vragers) zijn meer cyclisch dan gewone prijsindices en reser-veringsprijsindices van aanbieders. De resultaten laten verder zien datgewone prijsindices en prijsindices van het aanbod trager reageren dande reserveringsprijsindices van vragers. Merk op dat dit hoofdstuk in-gaat op commercieel vastgoed in de Verenigde Staten, de andere tweehebben betrekking op de Nederlandse woningmarkt. Het zou zo kunnenzijn dat er verschillen zijn in de verhouding tussen prijzen en liquiditeittussen commerciele en residentiele vastgoedmarkten enerzijds en tussenAmerikaanse en Nederlandse vastgoedmarkten anderzijds. Bijvoorbeeldomdat commercieel vastgoed over het algemeen meer een investerings-goed is dan residentieel vastgoed. Dit laatste wordt over het algemeenmeer gezien als een consumptiegoed. Verder is het aanbod in (sommige)markten in de Verenigde Staten veel minder gerestricteerd dan in Neder-landse markten. Dit zou ook van invloed kunnen zijn op de verhoudingtussen prijzen en liquiditeit. Daarnaast verschillen de hypotheekmarktenin beide landen van elkaar. Dit kan implicaties hebben voor het gedragvan vragers en aanbieders bij het bepalen van hun reserveringsprijzen.Zo kan bijvoorbeeld worden beargumenteerd dat anchoring sterker aan-

151


wezig is in markten met hoge hypotheekschulden, aangezien de kans opeen restschuld hier groter is. Een uitgebreidere analyse van deze onder-werpen kan in toekomstig onderzoek worden opgepakt.

Hoofdstuk 4 presenteert een theoretisch en empirisch model waar-bij transactieprijzen, marktkrapte en marktliquiditeit elkaar beınvloeden.Het hoofdstuk biedt naast inzichten in de variatie over tijd ook inzichtenin geografische verschillen. De resultaten laten zien dat veranderingenin marktkrapte gebaseerd op zoekgedrag op internet (clicks per woning)voorlopen op (1) veranderingen in marktliquiditeit en (2) veranderingenin prijzen. Daarnaast laat het hoofdstuk zien dat liquiditeit sneller danprijzen reageert op schokken in marktkrapte. Het effect op liquiditeit istijdelijk en het effect op prijzen permanent. Het onderliggend theore-tisch mechanisme is dat een toestroom van meer kopers zorgt voor eenkrappere markt, hetgeen leidt tot een toename in de kans op een trans-actie. Door een asymmetrie in informatie observeren verkopers dit nietdirect en passen zij hun reserveringsprijzen langzamer aan. Het gevolgis een vertraagde respons van transactieprijzen, terwijl marktliquiditeitdirect omhoog gaat. Er zijn geografische verschillen in dit proces. In ste-delijke gebieden wordt een vraagschok geabsorbeerd in marktliquiditeiten passen prijzen zich langzaam aan. In landelijke gebieden daarentegenis er geen effect te zien op marktliquiditeit, maar alleen een vertraagdeffect op prijzen. De reden is dat stedelijke gebieden herstellende warenvan de crisis in het laatste jaar van de steekproef, terwijl dit nog niethet geval was in landelijke gebieden. Dit is ook zichtbaar in het modeldat geschat is op 50% van de gebieden met de meest negatieve rende-menten. Price discovery is deze gebieden is trager dan in andere regio’s.De reden is dat het aantal transacties lager ligt in deze gebieden. In hettheoretisch model wordt dit gereflecteerd in het feit dat verkopers meerwaarde hechten aan de perceptie van marktkrapte van gisteren dan vanvandaag.

De derde deelvraag luidt “Hoe kan marktliquiditeit worden ge-bruikt voor een beter begrip en een betere monitoring van devastgoedmarkt?”. Dit proefschrift toont aan dat het begrijpen vanmarktliquiditeit van cruciaal belang is voor beleidsmakers en anderemarktvolgers. Liquiditeit kan sterk van pas komen bij zowel de monito-ring als het voorspellen van vastgoedmarkten.

Hoofdstuk 2 laat zien dat de gemiddelde verkooptijd van verkochtewoningen –een maatstaf die regelmatig wordt gebruikt als indicator voormarktliquiditeit– misleidend kan zijn. De reden is dat bij de berekening

152


van dit gemiddelde alleen verkochte woningen worden gebruikt. Daar-naast zijn verhandelde woningen heterogeen. Aangezien de kans op te-rugtrekking en woningkenmerken over de tijd verschillen, kan varen opalleen de gemiddelde verkooptijd tot verkeerde conclusies leiden. Dezeproblemen zijn groter in kleine markten of in markten waar weinig wo-ningen worden verhandeld (bijvoorbeeld tijdens een crisis). De resulta-ten suggereren dat in een neergaande (opgaande) markt, de gemiddeldeverkooptijd van verkochte woningen marktliquiditeit onderschat (over-schat).

Hoofdstuk 3 biedt een dieper inzicht in de vastgoedgoedmarkt doorte kijken naar de reserveringsprijzen van kopers en verkopers. Voor be-leidsmakers is het model interessant bij het monitoren en voorspellenvan de vastgoedmarkt. De vergelijking tussen de vraag- en aanbod re-serveringsprijsindices geeft een completer beeld van de historische enhuidige marktsituatie dan gewone prijsindices. Omdat de vraagindiceseen maatstaf bieden voor zowel prijs- als liquiditeitsontwikkelingen, lo-pen deze voor op prijsindices. Daarom kunnen deze tevens nuttig zijnbij het voorspellen van prijsindices. De vertraging van het aanbod tenopzichte van de vraag is ongeveer een jaar. De vertraging van normaleprijsindices is ongeveer twee kwartalen.

Hoofdstuk 4 biedt theoretisch en empirisch bewijs voor de relatietussen marktkrapte, marktliquiditeit en huizenprijzen. Het begrijpenvan deze relatie kan belangrijk zijn voor beleidsmakers die deze markt-indicatoren bij het duiden van de marktsituatie vaak los van elkaar ge-bruiken. Daarnaast ontwikkelt dit hoofdstuk een nieuwe indicator ge-baseerd op internetzoekgedrag die voorloopt op andere indicatoren. Ditkan nuttig zijn bij het voorspellen van regionale woningmarkten. Devertraging van marktliquiditeit ten opzichte van deze indicator is onge-veer een kwartaal. Prijzen reageren langzamer, ongeveer de helft van detotale schok is geabsorbeerd na twee kwartalen.

153


154

List of Co-Authors

Chapter 3 is based on the working paper by Van Dijk, Geltner & Vande Minne (2018) with the title “Revisiting supply and demand indexesin real estate”, which can be found on the website of De NederlandscheBank. The input and work of the co-authors on this chapter is very muchappreciated and acknowledged, and the co-authors recognized Dorinthvan Dijk as lead author.

Chapter 4 is based on the article by Van Dijk & Francke (2018) withthe title “Internet Search Behavior, Liquidity and Prices in the HousingMarket”, and is published in Real Estate Economics 46(2), which canbe found in the Wiley Online Library. The input and work of the co-author on this chapter is very much appreciated and acknowledged, andthe co-author recognized Dorinth van Dijk as lead author.

155

LIST OF CO-AUTHORS

156

Acknowledgments

This thesis is the result of four enjoyable, exciting, entertaining, enlight-ening, difficult, sometimes filled with blood, sweat and tears, but aboveall inspirational, years. It wouldn’t have been possible without the help,advice, and distraction of many individuals. Here I will try to acknowl-edge everyone who made this thesis possible –or to put in the words ofYogi Berra: “Thank you all for making this day necessary”.

First and foremost, thanks to my supervisors Marc Francke andJakob de Haan for their guidance of this thesis.

Marc, your unparalleled combination of knowledge on real estate,econometrics, and warm personality was helpful on all occasions, bothon work-related and personal issues. I appreciate the trust you put in meto pursue a PhD even before I finished my master’s thesis. I’m lookingforward to continue working on our existing projects and to start newones.

Jakob, thank you for your valuable and comprehensive comments onmy work. Even though you always claim that real estate and econo-metrics are not your fields, your knowledge of these topics for sure isimpressive and was very helpful. I hope that we can start a new jointproject soon and I am looking forward to the next Sigaren met Jakob.

I would further like to thank everyone at the UvA Finance Group.In particular thanks to the real estate colleagues for making me feelpart of a group that combines an innovative academic climate with adown-to-earth industry perspective. Martijn Droes, for your sugges-tions on my papers, for showing me the way at many conferences, andfor the ice cream. Peter van Gool, for your detailed comments on mythesis, your wisdom in general, and, of course, the wine. Johan Conijn,

157

ACKNOWLEDGMENTS

Erasmo Giambona, Gianluca Marcato, Frans Schilder, Marcel Theebe,I’m grateful for the many insightful interactions. Thank you JeroenLigterink and, recently, Esther Eiling for your valued suggestions andmanagerial support. Also, thanks to Bas Bouten, Deyanira Gonzalez,and Lorena Zevedei for providing guidance in the administrative jungle.

I thank Johan Verbruggen and again Jakob de Haan, for puttingconfidence in me during my very first internship that eventually allowedme to write this thesis in combination with my work at De NederlandscheBank (DNB). During the recent years at DNB, I benefited a lot fromthe detailed suggestions and superb economic vision of Peter van Els.Thanks, for reading and commenting on the vast amount of (not alwayspolished) pages in the past years. Thanks to all other colleagues at EBOand DNB for the professional suggestions and the even more professionalkaraoke nights. Especially, thanks to the Backstreet Boys for having myback on stage.

This thesis was made possible by several inspiring overseas visits, ofwhich the one-semester visit at the MIT Center for Real Estate was thehighlight. In particular, David Geltner and Alex van de Minne, thankyou for showing that everything can be solved under beech trees. Theliquidity seminars in the Muddy Charles surely played an important rolefor this thesis. David, your knowledge on real estate, economics, finance,history, politics, Cruijff and Berra quotes, and even Dutch history isimpressive and I have benefited a lot from this. Alex, besides showingmy appreciation for sharing your vast knowledge and creative ideas, Ialso would like to thank you for joining me to Funda the very first timeand helping me with my master’s thesis. Thereby you constructed theunderlying foundations of this thesis like a true architect and you pavedthe way for me in many respects.

Also thanks to the other members of the defense committee. I amhonored that you agreed to be on my committee. John Clapp, manythanks for your useful suggestions on several chapters of my thesis andthe vriendelijke welcome at UConn. Piet Eichholtz, our enjoyable con-versations at conferences, seminars, and symposia, were always valuableto my research.

Data-driven is a buzzword that appears on many, not always appro-priate, occasions. I do think it is safe to say that it is appropriate in thecase of this thesis; without data there would not be a thesis. Therefore,my gratitude goes out to the data suppliers, Funda (Ruben Scholten),NVM/Brainbay (Frank Harleman), RCA (Jim Costello, Willem Vlam-

158

ACKNOWLEDGMENTS

ing, Bob White), for making this research possible.A big shout-out goes out to my paranymphs Coen Modderman and

Rob Sperna Weiland for providing the indispensable support during thefinal months. Coen, thanks for an inspirational way of showing me thatit is possible to combine PhD-life, working life, and jaarclublife. Rob,thanks for being my PhD-buddy, for your valuable input on numeroustopics, and for completing #dreamteam. But most of all, I’m proud ofour collaboration on the “Working 50 project”.

Many thanks to my other colleagues for making life at the officepleasant. Finance PhDs and other (former) M4.04 inhabitants, Felipe,Fleur, Guilherme, Ieva R., Ieva S., Joris, Magda, Merve, Richard, Rob,Robin, Ruobing, Pascal, Sin Man, Yumei, thank you for the vivid dis-cussions during already famous flamingo lunches. Team Phoenix (Eloisaand Hannah), thank you for all the fun during and beyond office times(and for losing at almost everything). Thanks to Joris, Leonard, and myother TI-friends for showing me what smart is supposed to be. Thanksto other ABS PhDs and the “drinks THIS Friday” group for the manyactivities, dinners, and weekends outside the office.

I couldn’t have focused during work without the support and distrac-tion provided by friends. Thanks to Il Situatione (Bob, Michiel, Pieter,Sjo, Tim, Tristan), Jaarclub Blaffer and disciples (Coen, David G.,David S., Dirk, Edward, Floris, Gust, Jeroen, Joost, Joris, Luuk, Marc,Patrick, Robin, Sven, Thomas), Menno, Thom, and other friends forthe many dinners, drinks, weekends, housewarmings, vacations, squashmatches, and Vastelaovende.

Finally, thanks to my parents, sister, in-laws, and other family mem-bers for your support, for pushing me to keep on following my intuition,and for telling me to slow down every now and then. Evelien, withoutyour warm love I wouldn’t have been able to do it.

159

uva-dare (digital academic repository) commercial and ... · van dijk, d.w. link to publication...

Documents