2009:034 master's thesis - diva portal1016092/fulltext01.pdf · i abstract the purpose of this...
TRANSCRIPT
2009:034
M A S T E R ' S T H E S I S
Unit Root Tests and Structural Breaksin the Swedish Electricity Price
Isabelle Nilsson
Luleå University of Technology
D Master thesis Economics
Department of Business Administration and Social SciencesDivision of Social sciences
2009:034 - ISSN: 1402-1552 - ISRN: LTU-DUPP--09/034--SE
”Unit Root Tests and Structural Breaks in the
Swedish Electricity Price”
Isabelle Nilsson
Social Science and Business Administration Programmes
ECONOMICS PROGRAMME
Department of Business Administration and Social Sciences
Division of Economics
Supervisor: Linda Wårell
I
ABSTRACT
The purpose of this thesis was to examine the Swedish electricity price in order to determine
whether there are any structural breaks in the time series. Both the nominal and the real price
of electricity under the period 1900 to 2006 are examined. The theoretical framework used in
this thesis is different tests for unit root and structural breaks in time series. Tests for both
exogenously and endogenously determined breaks are performed. Different quantitative
methods for measuring structural breaks are applied in order to compare any differences. The
results showed that both the nominal and real electricity price exhibits unit root and that
breaks are most likely to occur rapidly. Both series exhibits a break around 1945 a time when
the Swedish economy was growing and the demand for electricity was increasing. The results
implies that more than one test should be used and compared when testing for structural
breaks in time series.
II
SAMMANFATTNING
Syftet med den här uppsatsen var att undersöka det svenska elpriset för att se huruvida det
finns några strukturella förändringar i denna tidsserie. Den teoretiska referensram som
användes var olika test för icke-stationaritet och strukturella förändringar i tidsserier. De olika
testen tillät både exogent och endogent bestämda förändringar, detta för att kunna jämföra
eventuella skillnader. Analysen baserades på data för både det reala och det nominella elpriset
under perioden 1900 och 2006. Resultaten visar att både den nominella och reala tidsserien
för elpriset är icke-stationär och serierna lämpligast förklaras av snabba förändringar. Båda
serierna uppvisade strukturella förändringar kring 1945, en period då den svenska ekonomin
växte och efterfrågan på elektricitet steg. Resultaten visar också på att fler än ett test bör
användas och jämföras vid test för strukturella förändringar i tidsserier.
III
TABLE OF CONTENTS
ABSTRACT ................................................................................................................................ I
SAMMANFATTNING ..............................................................................................................II
TABLE OF CONTENTS ......................................................................................................... III
LIST OF FIGURES AND TABLES ........................................................................................ IV
Chapter 1 INTRODUCTION ..................................................................................................... 1
1.1 Background....................................................................................................................... 1
1.2 Purpose ............................................................................................................................. 2
1.3 Methodology .................................................................................................................... 3
1.4 Scope and Limitations ...................................................................................................... 3
1.5 Outline .............................................................................................................................. 3
Chapter 2 BACKGROUND ....................................................................................................... 4
2.1 The Swedish Electricity Price .......................................................................................... 4
Chapter 3 THEORY ................................................................................................................... 9
3.1 Time Series Analysis ........................................................................................................ 9
3.2 Stationary and Non-stationary Time Series ..................................................................... 9
3.3 Unit Roots Test: The ADF Test ..................................................................................... 12
3.4 Unit Root Tests with Structural Breaks .......................................................................... 13
3.4.1 Exogenous Structural Break .................................................................................... 14
3.4.2 Endogenous Structural Break .................................................................................. 15
3.4.3 Multiple Structural Breaks ...................................................................................... 17
Chapter 4 EMPIRICAL RESULTS AND ANALYSIS ........................................................... 19
4.1 Reliability and Validity .................................................................................................. 19
4.2 Unit Root Tests ............................................................................................................... 20
4.3 Exogenous Structural Break Test ................................................................................... 21
4.4 Testing Unit Root Allowing for Endogenous Breaks .................................................... 22
4.5 Testing Unit Root Allowing for Two Breaks ................................................................. 24
4.6 Comparison .................................................................................................................... 27
Chapter 5 CONCLUSIONS ..................................................................................................... 30
REFERENCES ......................................................................................................................... 32
IV
LIST OF FIGURES AND TABLES
Figure 2.1 The Swedish Electricity Price and the Oil Price, Nominal, 1900- 2000 .................. 5
Figure 2.2 The Swedish Electricity Price, Nominal and Real, 1990- 2006 ............................... 6
Table 4.1 ADF Unit Root Test ................................................................................................. 20
Table 4.2 Zivot-Andrews Unit Root Test Allowing for One Break ......................................... 22
Figure 4.1 Zivot-Andews Test, Nominal Electricity Price ....................................................... 23
Figure 4.2 Zivot-Andews Test, Real Electricity Price ............................................................. 24
Table 4.3 Clemente-Montañés-Reyes Unit Root Test with Double Mean Shift ...................... 24
Figure 4.3 Clemente-Montañés-Reyes AO Test, Nominal Electricity Price ........................... 25
Figure 4.4 Clemente-Montañés-Reyes IO Test, Nominal Electricity Price ............................. 25
Figure 4.5 Clemente-Montañés-Reyes AO Test, Real Electricity Price .................................. 26
Figure 4.6 Clemente-Montañés-Reyes IO Test, Real Electricity Price .................................... 27
Table 4.4 Summary of Unit Root Tests .................................................................................... 27
Table 4.5 Summary of Break Points ........................................................................................ 28
1
Chapter 1
INTRODUCTION
1.1 Background
For about 100 years ago it became possible to use electricity in Sweden. Until the First World
War not many households had electricity. It was not until the 1920’s the demand for
electricity increased significantly. The first two decades in the previous century is
characterized by varying electricity prices. In 1945 almost every household in Sweden had
electricity. Between the years 1930 and 1960 the prices stayed at a low level (Energy Markets
Inspectorate, 2007). Following the oil crises in 1973 the nominal price of electricity increased
rapidly and continued to do so until a couple of years after the deregulation in the Swedish
electricity market in 1996.
The main purpose of the deregulation was to increase the competition on the electricity
market and thereby decrease the price of electricity. First, there was only Sweden and Norway
that had a common electricity market, a couple of years later both Finland and Denmark
joined the market. This resulted in a common electricity market called Nord Pool which was
the first multinational exchange for trade in electric power contracts (Flatabo et al., 2003).
Around two years after the reform in the Swedish market the prices fell. In 2000 the price
increased and reached a peak during 2003. It decreased a little in 2004 but then went up again
in 2005 (Swedish Energy Agency, 2006).
When looking at the all over pattern for the Swedish electricity price the last century it
fluctuates in correspondence with international crises, such as the Great Depression in 1929
and the first oil crisis in 1973. These years can be treated as points of structural breaks in the
electricity price. This means that at least one of the parameters affecting the electricity price
might have changed at these dates in the sample period. These effects can be transitory or
permanent (Hansen, 2001). However, even if it is reasonable to believe that structural breaks
2
have occurred at these dates, it would be interesting to see whether the break dates of these
crises fits endogenously determined break dates.
The Ordinary Least Square (OLS) method of estimating a standard regression model assumes
that mean and variance of the variables being tested are constant over time. Most macro
economic variables have means and variances that changes over time, these are called non-
stationary or unit root variables. The OLS method therefore gives misleading interpretations
when the regression equation contains non-stationary variables (Glynn et al., 2007).
Literature treating time series analysis have for a long time stressed the importance of testing
for structural breaks in economic data sets. Structural breaks can reflect institutional,
legislative or technical change. These breaks can also reflect changes in economic policies or
large economic shocks such as the oil crises in 1973. One particular problem that is common
when dealing with empirical modelling is that often one or more of the time series being used
have a unit root. Most macroeconomic data exhibits unit root processes therefore it is
suggested that it might be necessary to isolate some unique economic events and see them as
a permanent change in the pattern of the time series (Perman and Byrne, 2006).
This has implications for how the time series can be modelled. There are different kinds of
structural breaks, they can be determined endogenously or exogenously. They can also be
multiple, which means that the time series contains more than one structural break. Different
kinds of models are developed in purpose to capture and test for these breaks (Perman and
Byrne, 2006). Testing for unit root and at the same time allowing for structural breaks has
some advantages. Firstly, it prevents the test results from being biased towards non-
stationarity and unit root. Secondly, these tests can identify when the possible break occurred
(Glynn et al., 2007). The question of this thesis becomes what structural breaks can be
identified in the case of the Swedish electricity price? Do endogenously determined breaks
occur at the dates of international crises? If not what can be the underlying explanation for
these breaks?
1.2 Purpose
This thesis examines the Swedish electricity price in order to determine whether there are any
structural breaks in the time series. Different quantitative methods for measuring structural
breaks will be applied in order to compare any differences.
3
1.3 Methodology
The theoretical framework used in this paper considers an autoregressive model using time
series data. An econometric approach will be used to analyze the Swedish electricity price.
Statistical data on the price of electricity in Sweden from 100 years have been collected. One
of the methods used is the traditional Augmented Dickey-Fuller test to test whether the series
contains a unit root or not. To test for one exogenous structural break a Chow test will be
performed, this test does not test for unit root. The third test is a Zivot-Andrews unit root test
with one endogenously determined structural break. Finally a test proposed by Clemente,
Montañés and Reyes for two breaks is performed.
1.4 Scope and Limitations
The thesis will only consider the price of electricity in Sweden. The data that is received from
Kander (2002) concerns the nominal price of electricity and covers the years between 1892
and 2000, but to avoid some extreme values in the middle of 1890 the years used will be
limited to the time from 1900 to 2000. The statistics concerning the electricity price between
2001 and 2006 is received from SCB. The nominal prices will be converted into real prices
with the price level of 2006 as the base year. The data concerns the price of electricity without
taxes. Both the nominal and the real prices will be used and compared. The thesis will only
perform the structural breaks test that is presented above.
1.5 Outline
This thesis is divided into five chapters. The introduction in chapter 1 gives a general
description of the problem and further it presents the purpose, methodology, scope and
limitations of the thesis. Chapter 2 describes the background of the subject which includes a
historical approach of the Swedish electricity price. Chapter 3 presents the theoretical models
and tests used in this thesis. It also explains the concept of stationarity. In chapter 4 the results
from the tests for unit root and structural breaks are presented. It also includes a discussion
about the data used. The chapter ends with a comparison and an analysis of the results and the
methods used. Final conclusions can be found in chapter 5.
4
Chapter 2
BACKGROUND
This chapter will start with a historical presentation of the Swedish electricity price. The
presentation will focus on the years where there have been significant changes in the trend of
the price and look at what factors that might have caused these trend breaks.
2.1 The Swedish Electricity Price
The demand for electricity is decided by the general economic situation and the temperature.
Factors that affect the electricity price can be found in the conditions and factors that affect
the cost of producing electricity. In Sweden the most important of these is the hydrological
development, the development of capacity, the interaction on the Nordic electricity market
(since 1996), fuel prices, market competition and since 1995 the trade with emission permits
(Swedish Energy Agency, 2006).
For about 100 years ago it became practically possible to use electricity in Sweden. In the
beginning the electricity was produced by steam plants and only used for lighting. In this
period the electricity price was around today’s level in nominal terms. Around the turn of the
century Sweden started to build their first hydro power plants which became the main
electricity source. Private and communal power producing companies were given the
assignment to produce and distribute electricity. When it became a lack of paraffin during the
First World War the development of energy plants increased. Figure 2.2 shows that both the
real and the nominal electricity price increased to high levels during the First World War. In
1920 many household appliances were introduced in the Swedish homes and the households
became more dependent on electricity. This increased the demand for electricity and in 1945
almost every household in Sweden had electricity (Energy Markets Inspectorate, 2007).
Additionally, after the end of the Second World War the Swedish economy was strong. The
manufacturing industry developed and hence this sectors demand for electricity increased
(NE, 2009b).
5
As one can see in figure 2.1 the electricity price almost doubled between the years 1915 and
1920. With an increasing demand and supply during the early 1920 the prices started to
decline rapidly in both nominal and real terms. The decrease continued after the middle
1920’s but at a lowering rate (Energy Markets Inspectorate, 2007). An international crisis that
affected the world economy at this time was when the New York stock market crashed in
1929 which was the starting point of the Great Depression in the 1930’s (NE, 2008b).
In Sweden the first nuclear power plant was taken into production in 1972. However, with a
strong opinion against further development of nuclear power plants together with the oil crises
in 1973 and 1979 the Swedish government realised the weaknesses in their electricity
production. So they began to look for renewable energy sources (Rönnborg, 2003). Electricity
produced from nuclear power is relatively cheap compared to electricity that comes from
wind power for example. Although the demand for electricity has continued to increase since
1970 it has been doing so at a declining rate. At the same time the price of electricity has
increased significantly. One important factor contributing to the massive increase in the
electricity price and the demand for electricity is the relative price of oil, see figure 2.1.
Figure 2.1 The Swedish Electricity Price and the Oil Price, Nominal, 1900- 2000.
Source: Kander (2002).
0,0000
0,1000
0,2000
0,3000
0,4000
0,5000
0,6000
1900
1906
1912
1918
1924
1930
1936
1942
1948
1954
1960
1966
1972
1978
1984
1990
1996
Oil an
d electricity price SEK/KW
h
Oil price SEK/KWh Nominal
Electricity price SEK/KWh Nominal
6
After the oil crises in the 1970’s oil became substituted for electricity, especially when it
came to heating (Andersson, 1997). Another reason for the increase can be the less oil
intensive electricity production that could not generate the electricity that was being
demanded. Therefore it became an excess demand on the market which created an increase in
the price of electricity. Comparing the nominal and the real price of electricity one can see
that they are following different patterns, see figure 2.2. Before the oil crises in the 1970’s
both series are declining at different rates, after this period the nominal price increased
significantly while the real price continued to decline due to a lower rate. One explanation for
this is the high inflation rates in Sweden between 1970 and 1990 (SCB, 2007). One important
year in the Swedish economy is 1982 when the Swedish central bank devaluates the Swedish
currency by 16% after years of depression in the 1970’s. This action was named the “Big
bang”. At the same time the U.S. trade cycle went up and a long period of prosperity had
begun (Bergström, 1999). Furthermore, the Swedish economy went through a financial crisis
in the beginning of 1990. The crisis was caused by a major credit expansion that came after a
deregulation of the monetary system in 1989 (NE, 2009a).
Figure 2.2 The Swedish Electricity Price, Nominal and Real, 1990- 20061.
Source: Kander (2002) and SCB (2007).
1 To make the pattern more visible the nominal price has been multiplied by 10.
0,0000
2,0000
4,0000
6,0000
8,0000
10,0000
12,0000
14,0000
1900
1907
1914
1921
1928
1935
1942
1949
1956
1963
1970
1977
1984
1991
1998
2005
Electricity price SEK/KW
h
Electricity price SEK/10 KWh Nominal
Electricity price SEK/KWh Real (Base year: 2006)
7
In 1992 the first step towards a deregulated electricity market was taken when the
government’s hydroelectric power company was divided into Vattenfall AB and Svenska
Kraftnät. Until then the Swedish government was the only supplier of electricity in Sweden.
The deregulation continued in 1996. The purpose of the reform was to increase the
competition on the electricity market and in that way decrease the price of electricity. This
resulted in integration between the Swedish and Norwegian market into one deregulated
electricity market called Nord Pool (Andersson, 1997). In 1998 Finland entered the common
market and Denmark followed in 1999/2000. Nord Pool was the first multinational exchange
for trade in electric power contracts (Flatabo et al., 2003).
The market consists of one market operator (Nord Pool) and five system operators in the
member countries where each country has their own regulatory agency. In this market all
countries prices are equal to the system price. This system price is calculated by Nord Pool
and decided by the trade on the market. If the price calculation exceeds grid capacity limits
then two or more area prices are calculated. Generally, the spot price reflects the households’
prices which indicate that the retail market works efficient. The market for power trading is
called Elbas and the trade leads to physical delivery. It has the function of an aftermarket to
the spot market at Nord Pool. The involved actors in this market are the power producers,
distributors, industries and brokers from Sweden and Finland (Flatabo et al., 2003). In 2005
Nord Pool continued to grow when it opened the KONTEK bidding area in Germany, which
gave them access to the Vattenfall Europe Transmission control area (Nordpool, 2008).
The demand for electricity in Sweden increased about three percent during the years between
1996 and 2005. The interaction between the Nordic countries different production structures
is important for the electricity price on the Nordic market. The limited trade capacity between
the countries sometimes results in differences in prices between the price areas (Swedish
Energy Agency, 2006).
The system price was declining the years after 1996. In the year of 2000 the average price
was facing its lowest level. After 2000 the price increased and reached a peak during 2003. It
decreased a little in 2004 but then went up again in 2005. One of the main reasons for this is
the trade of emission permits that was introduced in 2005. Another important factor was the
very high prices on fossil fuels. The prices of oil and gas decreased between 1996 and 1999
8
but have thereafter increased, especially between the years 2004 and 2005. Furthermore, the
nuclear plant Barsebäcks second reactor was deactivated in May 2005 (Swedish Energy
Agency, 2006).
9
Chapter 3
THEORY
This chapter begins with discussing time series analysis and the autoregressive model that will
be used is presented. Further, the basic concepts of stationary and non-stationary time series
are explained. The chapter continues with tests for unit roots and the different kinds of
structural breaks.
3.1 Time Series Analysis
Time series analysis concerns observations that are ordered over time, the electricity price is
an example of a time series variable. This variable is random, i.e. its value cannot be
determined before it has actually been observed. Time series analysis is often used in the
finance sector, in research and the raw material market to name a few examples (Westerlund,
2005).
When estimating OLS regressions using time series data the time series is assumed to be
stationary. The problem is that most macroeconomic variables are non-stationary which can
cause serious econometric consequences. A common situation that occurs when dealing with
time series data is autocorrelation (Westerlund, 2005). The standard test for autocorrelation is
the Durbin-Watson (DW) test which is usually included in the basic diagnostic statistics. This
can however be corrected for by using an autoregressive model. Autocorrelation might
indicate omission of one or more lagged variables, or the omission of any important variable
in the model specification (Dougherty, 2007).
3.2 Stationary and Non-stationary Time Series
As mentioned above, performing OLS regressions assumes stationary time series. This means
that the time series exhibits a mean and variance that is constant over time. The problem with
most macroeconomic and financial variables is that they are non-stationary. This means that
10
the time series mean and variance varies over time, which can have serious econometric
consequences (Westerlund, 2005). The non-stationary series have no tendency to return to a
long-rung deterministic trend and the variance of the time series is dependent on time (Glynn
et al., 2007). The problem with for example two non-stationary variables is that they can
separate from each other over time and lose their joint relationship. This phenomenon is
usually called a spurious regression (Westerlund, 2005).
Most recent work has been focusing on the fitting of single variable processes where the
variable is a linear function of previous values of itself and the error term (Dougherty, 2007).
This model says how many lagged values of the dependent variable y that lies on the right
hand side of the equal sign in the regression model:
(1)
This time series is autoregressive of order ρ called the AR(ρ)-model. The model has an
intercept β0, the autoregressive parameters β1, β2,…, βρ, an error term εt and it does not include
any explanatory variables. t represents the time period. For simplicity assume only one lagged
value:
(2)
This thesis and the following discussion will focus on this simplified version of the previous
model. Though, the results from this AR(1)-model can normally be generalized to the regular
AR(ρ)-model in equation (1) (Westerlund, 2005).
If the regressive parameter β1 is:
1 1 (3)
Then yt is a stationary variable. Further, if β1 is equal to 1 then the AR(1)- model in equation
(2) and yt is said to be non-stationary. If β1 is greater then 1 this would cause yt to explode and
be infinitely large over time. However, this is not reasonable and therefore should the cases
where β1 > 1 and where β1 < -1 be ignored. If β1 = 1 the variable yt has what is called a unit
11
root and is therefore non-stationary. To test whether a time series variable is non-stationary or
not one can use the Augmented Dickey Fuller model (see section 3.4). This model tests if the
variable contains a unit root (Westerlund, 2005).
If the variable contains a unit root, i.e. β1 = 1, equation (2) becomes:
(4)
This is an example of a non-stationary process known as a random walk (Dougherty, 2007).
This means that the non-stationary time series suffer permanent effects from random shocks.
(Glynn et al, 2007). If the series starts at y0 at time 0, its value at time t become:
(5)
In this process the contribution of each shock ε is permanently built into the time series.
Because the series includes the sum of the shocks, it is said to be integrated. If expectations
are taken into account at time 0, the expected value of yt at any future time t is independent of
t but its variance is increasing over time:
(6)
(Dougherty, 2007). If a non-stationary variable is present this could be rewritten into
stationary variables before an estimation of the model is done. This implies that the
interpretation of the parameters should be slightly different (Westerlund, 2005). If one can
transform the non-stationary process in this way, it is said to be difference-stationarity. If it
becomes stationary by differencing y once it is integrated of order 1, I(1). If it becomes
stationary by differencing twice it is called I(2) and so on. Trend-stationarity, on the other
hand, is when a non-stationary time series can be transformed into a stationary process by
extracting a time trend. The difference between these two types of stationarity is important in
the analysis of time series (Dougherty, 2007).
There are two reasons why one should be careful when doing regressions with non-stationary
time series. The first is that the properties of the regression estimators are likely to be
12
critically affected. The second is regarding the risk of spurious regressions (Dougherty, 2007).
Signs of a non-stationary variable in a regression are that the DW statistics for the
autocorrelation is too small, that the R2-statistic is near 1 and that the explanatory variable is
very statistical significant, despite the fact that it is not a determinant of the dependent
variable (Westerlund, 2005). This type of spurious regressions can be avoided by detrending
the variables before doing the regression, i.e. by including a time trend in the regression. The
Monte Carlo experiment showed that spurious regressions also can occur in regressions with
integrated time series and with regressions using stationary time series, if autocorrelation in
the disturbance term is ignored (Dougherty, 2007).
Moreover, if the non-stationary variables are cointegrated, i.e. that the two variables includes
the same unit root or stochastic trend, the variables have a joint component. Therefore they
are not independent despite that they are non-stationary, which implies that they are related in
some way. Cointegration can be interpreted as a sort of long run equilibrium. For the
stochastic trend in the variables to take out one another the disturbance term in the
cointegration regression has to be stationary, which it will be if the variables have the same
stochastic trend. To test for cointegration is to test whether the disturbance term in the
regression model is stationary or not (Westerlund, 2005). There will however be no focus on
this phenomenon during this thesis, but worth mentioning.
3.3 Unit Roots Test: The ADF Test
One way of detecting non-stationarity is often described as testing for unit roots. The standard
test by Dickey and Fuller (1979) is based on the model:
(7)
rewritten as
∆ 1 (8)
Where Δyt = yt – yt-1. The time series is non-stationary if the coefficient of yt-1 is zero, the
series is then difference-stationary. If the coefficient of t not is zero the series is trend-
stationary. One property of the Dickey-Fuller test is that the disturbance term is not
autocorrelated. However, if it is autocorrelated more lagged values of yt should be included on
13
the right hand side of equation (7). Rewriting equation (8) yields the so-called Augmented
Dickey-Fuller (ADF) model (Dougherty, 2007):
∆ 1 ∆ (9)
This model tests whether a time series is affected by transitory or permanent shocks, by
testing the null hypothesis that the (β1 + β2 – 1) = 0. This means that the series exhibits unit
root, i.e. it is non-stationary. This hypothesis is tested against the alternative hypothesis that
(β1 + β2 – 1) < 0, i.e. that the series is stationary (Perman and Byrne, 2006). Here Δ denotes
the first difference of the time series yt that is being tested (Glynn et al., 2007). It should
however by mentioned that in practice the test tends to have low power and rejecting the null
hypothesis does not always mean that the time series is non-stationary (Perman and Byrne,
2006). Perron (1989)2 showed that without considering a possible existing break when doing
the ADF test could lead to a bias that reduces the ability to reject a false unit root null
hypothesis. To correct for this type of failure Perron suggested an ADF test allowing for a
known or exogenous structural break (Glynn et al., 2007).
3.4 Unit Root Tests with Structural Breaks
A structural break may be the change in the time series as a result of some unique economic
event (Glynn et al., 2007). Structural breaks can reflect institutional, legislative or technical
change. It can also be changes in economic policies or large economic shocks, such as the oil
crises in 1973. Most macroeconomic time series variables are non-stationary, i.e. they contain
a unit root (Westman, 2005). This means that a structural break can have a permanent affect
on the pattern of the time series (Perman and Byrne, 2006). Testing the unit root hypothesis
allowing for structural breaks has some advantages. One is that it prevents test results from
becoming biased towards unit root. Another advantage is that it can identify when the possible
break occurred. Two important issues with this type of procedure is that some authors claims
that there is a trade-off between the power of the test and the amount of information included
with respect to the choice of break date. A second issue is that the tests only capture the single
most significant break in the series, this can however be corrected for by using tests allowing
for multiple breaks (Glynn et al., 2007).
2 Hereafter referred to as Perron.
14
3.4.1 Exogenous Structural Break
Structural breaks in time series can affect the results of unit root tests (see section 3.4). Perron
argued that the ADF tests become biased towards the non-rejection of the null hypothesis
when a structural break is present. Further, he claims that most macroeconomic series do not
actually exhibit unit root. This characteristic has rather come from large and infrequent shocks
which indicate that the economy returns to a deterministic trend after small and frequent
shocks. Perron therefore modified the standard ADF test to a procedure that includes dummy
variables to account for a single exogenous (known) break. This test allows for a break under
both the null and the alternative hypothesis (Glynn et al., 2007). When an ADF test indicates
that a unit root is present in a time series, the unit root should be considered as a single
permanent break in a deterministic part of a stationary or trend-stationary process (Perman
and Byrne, 2006).
The following three trend break models are developed from Perron. These tests for the
existence of unit root taking three types of structural breaks into account. At first there is a
crash model that allows for a break in the intercept (level) of the time series:
∑ ∆ (10)
The second model is called a changing growth model that allows for a change in the slope
(rate of growth) of the linear trend and the two segments joined at the break date:
∑ ∆ (11)
An example of this can be a productivity slowdown. The third model allows for both an
intercept and a slope change to occur simultaneously, this means a time change in both the
level and the slope.
∑ ∆ (12)
TB represents the break date. The intercept dummy DUt shows a change in level where DUt =
1 if (t > TB) and DUt = 0 otherwise. The slope dummy DTt and DTt* corresponds to a change
in the slope of the trend function DT* = t-TB or DTt* = t if t > TB and zero otherwise. These
15
three models have a unit root with break under the null hypothesis because the dummy
variables are included in the regression under the null hypothesis. The alternative hypothesis
indicates a broken trend stationary process, i.e. that the break was identified ex ante by
economic information (Glynn et al., 2007). This test will not be used in this thesis, however,
the Zivot and Andrews test that will be presented in section 3.5.2 is based upon one of these
models suggested by Perron.
Another test for structural change is the Chow test. The testing procedure divides the sample
into two sub-periods and estimates the same equation for each sub-period separately. Then the
equality of the two sets is tested using the F-statistic to see whether there is some structural
change between the periods before and after the exogenously chosen break date. There are
two ways to pick the date: either by picking an arbitrary candidate break date or picking a date
based on some known characteristic of the data. There are however problems with both ways.
In the first case the true break date can be missed so the test may be uninformative. In the
second case the test can be misleading because the break date is correlated with the data and
the test might indicate a break even though no break exists (Hansen, 2001). The null
hypothesis is that the coefficient for the two sub-periods is equivalent and that there is no
structural change between the two periods. A rejection of the null hypothesis indicates that
there is a structural change between the two sub-periods (Dougherty, 2007).
3.4.2 Endogenous Structural Break
The break date can also be endogenous, i.e. it is correlated with the data (Hansen, 2001). The
test for these kinds of breaks is the same as for exogenous breaks. One tests for each possible
break date in the sample, or some specific part of the sample, and then chooses the date with
strongest evidence against the null hypothesis of unit root. That is where the t-statistic from
the ADF test of unit root is at a minimum (Perman and Byrne, 2006).
Perron and Vogelsang (1992) and Perron (1997) suggested a test that allows for two different
forms of structural break called the Additive Outlier (AO) and the Innovational Outlier (IO)
models. The AO model changes are assumed to take place rapidly allowing for a break in the
slope, while in the IO model changes are assumed to take place gradually. The latter one
allows for a break in both the intercept and the slope. Common for both models is the
assumption of no break under the null hypothesis of unit root (Glynn et al., 2007).
16
Zivot and Andrews (1992)3 transforms Perron’s unit root test that is based upon an
exogenously determined break date into an unconditional unit root test, i.e. instead of treating
the break date as fixed these authors purpose a test where the break date is estimated. The test
allows for a single break in the intercept and the trend (slope) of the series (Zivot and
Andrews, 1992). They suggest a sequential test using the full sample and a different dummy
variable for each possible break date (Perman and Byrne, 2006).
Zivot and Andrews uses the intervention outlier model for the changing growth model in
equation (11) instead of the additive outlier model used by Perron. The regression is written:
ŷ â ŷ ∑ ĉ ∆ŷ ê (13)
where ŷBt is the residuals from a regression with yt as the dependent variable and where the
explanatory variables contains a constant, time trend and DT*t. Further, they treat the
structural break as an endogenous occurrence and the null hypothesis for the three models
presented in equations (10), (11) and (12) is:
(14)
The selection of the breakpoint for the dummy variable in equations (10), (11) and (12) is
viewed as an outcome of an estimation procedure that is designed to fit yt to a certain trend
stationary representation. In other words, Zivot and Andrews assume that the alternative
hypothesis specifies that yt can be a trend stationary process with one break in the trend that
occurs at an unknown point in time. The goal with this test is to find the breakpoint that
supports this alterative hypothesis the most (Zivot and Andrews, 1992). The time of the break
is detected based on the most significant t-statistics for α (Saatcioglu and Korap, 2008). This
is where the t-statistic from the ADF test of unit root is at a minimum, i.e. a break date is
selected where the strongest evidence against the null hypotheses of unit root is found. A
break under the null hypothesis is not allowed in this test (Perman and Byrne, 2006).
3 Hereafter referred to as Zivot and Andrews.
17
3.4.3 Multiple Structural Breaks
Some variables do not show just one break. Relaxing the assumption about only one structural
break makes it possible to test for multiple breaks. In this case the distinction between a non-
stationary and stationary series becomes less clear (Perman and Byrne, 2006).
Clemente, Montañés and Reyes (1998) extend the Perron and Vogelsang (1992) statistics to
the case of two changes in the mean. They wish to test the null hypothesis H0 against the
alternative hypothesis H1.
: (15)
: (16)
In these equations DTBit is a pulse variable equal to one if t = TBi +1 and becomes zero
otherwise. Further, DUit = 1 if t > TBi (i = 1,2) and zero otherwise. TB1 and TB2 represents the
time periods when the mean is being modified. For simplicity, assume that TBi = λiT (i = 1,2)
where 0 < λi < 1 and λ2 > λ1 (Clemente et al, 1998).
If the two breaks belong to the innovational outlier one can test the unit root hypothesis by
estimating the following model:
∑ ∆ (17)
From this estimation the minimum value of the simulated t-ratio is obtained and this value can
be used for testing if the autoregressive parameter is 1 for all break time combinations. To
make it possible to derive the asymptotic4 distribution of this statistic assume that 0 < λ0 < λ1,
λ2 < 1 – λ0 < 1. By implementing the largest possible window, λ1 and λ2 take values in the ((t
+ 2)/T, (T – 1)/T)) interval. Further, assume that λ2 > λ1 + 1 which implies that cases where
the breaks occur in consecutive periods are eliminated (Clemente et al, 1998).
4 Definition asymptote: asymptote to a curve is a straight line that the curve approaches unrestrictive without
coincides with it (NE, 2008a).
18
If the shifts are better described as additive outliers the unit root hypothesis can be tested
through a two-step procedure. First, eliminate the deterministic part of the variable by
estimating the following model:
ỹ (18)
The second step involves taking out the test by searching for the minimal t-ratio for the
hypothesis that ρ = 1 in the following model:
ỹ ∑ ∑ ỹ ∑ ∆ỹ (19)
The dummy variable DTBit is included in the model to make sure that min ^ (λ1, λ2)
converges to the distribution (Clemente et al, 1998):
min ^ , ½ ½ (20)
The thesis continues to examine whether there are structural breaks in the Swedish electricity
price and if these coincide with the international crises. First, a Chow test for one exogenously
determined break will be used. Second, the Zivot and Andrews test that assumes one
endogenous break is performed. Finally, the Clemente, Montañés and Reyes test is used to
test for two structural breaks.
19
Chapter 4
EMPIRICAL RESULTS AND ANALYSIS
The chapter begins with a discussion of the reliability and validity of the data and the
methods. Five different methods have been used to detect unit root processes and structural
breaks in both the nominal and the real Swedish electricity price. Firstly, a Dickey-Fuller and
an Augmented Dickey-Fuller test not allowing for any breaks is used to determine whether
the variables exhibits unit root. Secondly, exogenously determined breaks are detected by the
Chow test for structural change. Furthermore, a Zivot and Andrews test is performed to test
for unit root allowing one endogenous structural break and also a test for two breaks purposed
by Clemente, Montañés and Reyes. The results from these tests are presented in order and
analyzed in a comparison in the end of the chapter.
4.1 Reliability and Validity
The data is based on the nominal value of the electricity price, which implies bias in form of
inflation included in the data. However, when examining structural breaks it is also interesting
to analyse time series with much variation. Furthermore, it is interesting to compare any
differences that might occur between the two time series. That is why tests on both nominal
and real prices will be done. Another limitation is that the years between 1892 and 1900 are
excluded. The main reason is to avoid some extreme values, which might warp the results. In
addition, the use of electricity was very low during these years. The statistics considering the
years between 1892 and 2000 is received in nominal prices from Kander (2002) and is
believed to be reliable. To capture the effects from the deregulation in 1996 the time series
has been extended to 2006. The additional data have been collected from the Statistics of
Sweden and is assumed to be reliable as well. The data concerns the spot price of electricity
without taxes.
The Chow test is a commonly used test for exogenously determined structural change.
Though there are some problems concerning the choice of break point. The true break date
20
can be missed so the test becomes uninformative or the date can be correlated with the data
which gives misleading results (Hansen, 2001). The Dickey-Fuller and the ADF test are
considered as the traditional unit root test. However, there exists some criticism against these
as mentioned before. Firstly, some authors claim that in practice the test tends to have low
power and rejecting the null hypothesis does not always mean that the time series is non-
stationary (Perman and Byrne, 2006). Secondly, not considering a possible existing break
could lead to a bias that reduces the ability to reject a false unit root null hypothesis (Glynn et
al., 2007). That is why the Zivot-Andrews is used. This method has been criticised for not
allowing for break under the null hypothesis, gradual breaks or multiple breaks (Perman and
Byrne, 2006). It is reasonable to think that the series might exhibit more than one structural
break and that breaks occur gradually. That is why the Clemente-Montañes-Reyes test is
performed. For estimation purposes Stata has been used for all tests performed and the tests
have been chosen from those tests that have been available in this programme.
4.2 Unit Root Tests
First, equation (9), the Augmented Dickey-Fuller test, was used to test for unit root without
allowing for any structural breaks. The variables tested are the logarithm of the nominal
electricity price (LELN) and the logarithm of the real electricity price (LELR) under the
period 1900 to 2006. Table 4.1 shows the results from the tests.
Table 4.1 ADF Unit Root Test
Levels
T-stat (lags)
1st difference
T-stat (lags)
Decision
LELN
Intercept
-0.841 (0) Non-stationary
-4.146 (2) Stationary
Unit root
Intercept + trend
None
LELR
Intercept
Intercept + trend
None
-1.240 (0) Non-stationary
-0.622 (0) Non-stationary
-1.437 (0) Non-stationary
-2.136 (0) Non-stationary
-4.033 (2) Stationary
-4.361 (2) Stationary
-4.157 (2) Stationary
-9.006 (1) Stationary
-9.109 (1) Stationary
-4.488 (2) Stationary
Unit root
Unit root
Unit root
Unit root
Stationary
21
The ADF test for unit root in levels shows that the series are non-stationary both in intercept
and trend. Taking the first difference of the series and testing these with and without intercept
and trend makes the series stationary. These results indicate that both of series exhibits unit
root processes. The criticism against the ADF test was that it did not allow for existing
structural breaks and therefore became biased towards unit root which reduces the ability to
reject a false unit root null hypothesis (Glynn et al., 2007). It should also be mentioned that
this test tends to have low power and rejecting the null hypothesis does not always mean that
the series is non-stationary (Perman and Byrne, 2006).
4.3 Exogenous Structural Break Test
To perform a Chow test on the nominal price an autoregressive model of order one was used
to regress the whole sample between 1900 and 2006. The AR(1) model was used because
including more lagged variables did not make the test statistics more significant so for
simplicity only one lag was included. A break date at 1977 was chosen, this because when
looking at figure 2.2 it seems that the series change pattern after this date. The first sub-period
therefore includes the observations between 1900 and 1976. The second sub-period includes
the observations from 1977 to 2006. The same AR(1) model was used to regress the two sub-
periods separately. The result of the residual sum of squares (RSS), the number of parameters
(k) and the number of observations for each period (n1 and n2) was used to perform an F-test.
/ /
(21)
An estimation of the equation above gave the F-value: 7.8483 and the critical F-value5 gives
4.82. The estimated F-value is larger than the critical value and therefore the null hypothesis
of no structural change can be rejected. The alternative hypothesis indicates that there is a
structural change between the two sub-periods in the nominal electricity price. The Chow test
therefore indicates that a structural break in this time series occurs at 1977. The test was also
performed on the real electricity price. A regression of an autoregressive model of order one
was used on the whole sample and the sub-periods. Including more than one lagged variables
made these variables not statistical significant. The chosen break date was at 1968. This break
5 Critical F-value with 1% significance level.
22
date seemed most reasonable when studying figure 2.2. At this point the pattern changes
significantly in the real price series. The estimated F-value becomes 14.2391 which is higher
than the critical F-value6: 4.82. One can reject the null hypothesis which indicates that 1968 is
a structural break in the real electricity price series.
4.4 Testing Unit Root Allowing for Endogenous Breaks
Equation (13), the Zivot-Andrews test, was used to test for unit root allowing for one
endogenously determined structural break. Both the logarithm of the nominal electricity price
(LELN) and the logarithm of the real electricity price (LELR) under the period 1900 to 2006
were estimated. Table 4.2 shows the results of the tests.
Table 4.2 Zivot-Andrews Unit Root Test Allowing for One Break
LELN LELR
Lags included* 3 3
Minimum t-statistics -3.571 -3.009
at year 1977 1990
1% Critical Value -5.43 -5.43
5% Critical Value -4.80 -4.80
* lags for the difference of the series selected via TTest.
When using the Zivot-Andrews test the break date is chosen where the t-statistics for α is
most significant, this is where the t-statistic from the ADF test of unit root is at a minimum.
This is the break date where there is strongest evidence against the null hypothesis of unit
root. For the nominal electricity price the null hypothesis of unit root cannot be rejected
despite that a break has been included. The breakpoint for this variable does not exactly
coincide with any international crises. The break date lies between the first and the second oil
crises in 1973 and 1979. Additionally, the years between 1976 and 1978 were characterized
by high inflation rates. As mentioned before the Zivot-Andrews test does not allow for a
break under the null hypothesis which makes the chosen break date for the nominal electricity
price not significant. This does not necessarily mean that there is no break in 1977. However,
the break date is chosen from the minimum t-statistics and because the t-statistics is calculated 6 Critical F-value with 1% significance level.
from th
fully tru
with the
Figure
This ho
rejected
the one
exogeno
test was
1990 th
also tw
explain
endogen
series fo
he variance
usted. This
e exogenous
4.1 Zivot-A
olds for the
d and theref
es found in
ously and t
s 1968. The
hat was caus
wo years be
a change
nous determ
or the real p
that varies
result is in
sly determin
Andews Tes
real electr
fore the resu
the nomin
the endogen
e Zivot-And
sed by a de
efore the fir
in pattern i
mined break
price of elec
over time w
nteresting b
ned break in
st, Nomina
icity price
ult cannot b
nal series w
nously deter
drews test in
eregulation
rst deregula
in the elect
k differs mi
ctricity.
23
when a seri
because the
n the Chow
al Electricit
as well. Th
be fully trus
where the br
rmined bre
ndicates a b
of the mon
ation of th
tricity price
ight be that
ies is non-s
break date
test.
ty Price
he null hyp
sted. Furthe
reak could
ak differs.
reak during
netary system
he Swedish
e. The reas
t there exis
stationary th
e chosen by
othesis of u
ermore, this
detect exo
The date c
g the financi
m and a cre
electricity
son why th
ts more tha
his result ca
y this test c
unit root ca
s result diffe
genously. H
hosen in th
ial crisis in
edit expansi
market wh
he exogenou
an one brea
annot be
coincides
annot be
fers from
Here the
he Chow
Sweden
ion. It is
hich can
usly and
ak in the
Figure
4.5 Test
Equatio
two stru
logarith
Table 4
Append
Table 4
Variable
LELN
LELR
* Min. t
Remem
break in
for a br
price th
AO mo
4.2 Zivot-A
ting Unit R
on (19), the
uctural brea
hm of the re
4.3 shows
dix 2.
4.3 Clement
e
M
-
-
t is the mini
mber that in
n the slope.
reak in both
e null hypo
del picks 1
Andews Tes
Root Allowi
Clemente-M
aks. Both t
eal electricit
the results
te-Montañé
Additiv
Min t *
-4.700
-3.684
imum t-stati
the AO mo
In the IO
h the interce
othesis of un
945 and 19
st, Real Ele
ing for Two
Montañes-R
the logarith
ty price (LE
from the
és-Reyes U
ve Outlier (
Optimal B
1945 an
1943 an
istics calcul
odel chang
model chan
ept and the
nit root cann
982 as the o
24
ectricity Pr
o Breaks
Reyes test, w
m of the n
ELR) under
tests, for m
Unit Root T
(AO)
reakpoints
nd 1982
nd 1966
lated. 5% cr
es are assu
nges are ass
e slope. Des
not be rejec
optimal brea
rice
was used to
nominal ele
r the period
more test s
est with Do
In
Min t
-3.726
-2.795
ritical value
med to tak
sumed to ta
spite the bre
cted in eithe
ak dates for
o test for un
ctricity pric
1900 to 20
statistics fro
ouble Mean
nnovational
* Opti
6 1
5 1
e for two bre
e place rap
ake place gr
eaks in the
er the AO or
r this series.
nit root allow
ce (LELN)
006 were es
om these t
n Shift
Outlier (IO
imal Breakp
917 and 19
938 and 20
eaks: -5.490
pidly allowi
radually and
nominal el
r the IO mo
. The t-stati
wing for
and the
stimated.
tests see
O)
points
75
00
0.
ing for a
d allows
lectricity
odel. The
istics for
the dum
shows t
IO mod
place ra
Figure
1945 w
Sweden
period a
demand
currency
nominal
rates at
the 1970
Figure
mmy variabl
that the AO
del. This ind
apidly rather
4.3 Clemen
as the last y
n had electr
after 1945, t
ded more e
y by 16%.
l electricity
26,2% and
0’s.
4.4 Clemen
les included
O model is m
dicates that
r than gradu
nte-Montañ
year of the
ricity which
the manufac
electricity.
The IO mo
y price. In
47,0%. The
nte-Montañ
d in the mod
more signif
t the series
ually.
ñés-Reyes A
Second Wo
h increased
cturing indu
In 1982 th
odel suggest
1917 and 1
e second br
ñés-Reyes I
25
dels and the
ficant in exp
is more lik
AO Test, N
orld War an
the demand
ustry in Sw
he Swedish
ts 1917 and
1918 Swed
reak in 1975
IO Test, No
e statistics fo
plaining the
kely to exh
Nominal Ele
nd at this ti
d for electr
eden develo
h central b
d 1975 as th
en suffers
5 can be exp
ominal Elec
for the nomi
e nominal p
ibit structur
ectricity Pr
ime almost
ricity. Addi
oped and he
bank deval
he optimal b
from extrem
plained by t
ctricity Pri
inal electric
price series
ral breaks t
rice
every hous
tionally, du
ence also th
luates the
break point
mely high
the two oil
ice
city price
than the
that take
sehold in
uring the
his sector
Swedish
s for the
inflation
crises in
In the c
reject th
the dum
explaini
graduall
exhibits
specific
model i
and 196
Figure
The yea
model. T
Great D
electrici
power c
level. O
market
end of
internat
Sweden
the dem
real ele
increase
substitu
case of the r
he null hypo
mmy variabl
ing the beh
ly is less li
s a unit ro
cation break
s significan
66.
4.5 Clemen
ars of 1938
The first br
Depression i
ity market a
contracts. D
One can say
started to s
the Secon
tional crises
n had electri
mand of elec
ctricity pric
e in the pric
uted for ele
real electric
othesis of un
les included
haviour of th
ikely to occ
oot process
ks are assum
nt in explain
nte-Montañ
8 and 2000
eak point ca
in the 1930
and Nord P
During this b
y that the e
show at this
nd World W
s. An additi
icity and the
ctricity. The
ce follows
ce of oil, s
ctricity and
ity price the
nit root. Ho
d shows tha
he real elec
cur in this
s despite t
med to take p
ning the real
ñés-Reyes A
have been
an only be e
’s. In 2000
ool became
break date t
ffects of th
s point. In t
War which
onal explan
e manufactu
e second stru
an even pa
ee figure 2
d since the 26
e results in
owever, in th
at only one
ctricity pric
series. The
the presenc
place rapidl
l electricity
AO Test, R
picked out
explained by
both Finla
e the first m
the average
he deregulat
he AO mod
h supports
nation is tha
uring indust
uctural brea
attern, see f
.1. When th
price of o
table 4.4 sh
he case of th
of the dum
e. This indi
e AO mode
ce of struc
ly. The dum
price and th
Real Electri
t as years o
y the beginn
and and Den
multinational
e price of el
tion and th
del the first
the theory
at around 19
try develope
ak takes pla
figure 2.2.
he oil price
oil has cont
hows that th
he IO mode
mmy variab
icates that b
el also sugg
ctural brea
mmy variabl
he test sugg
city Price
of structural
ning of the
nmark had j
l exchange
lectricity w
e more com
t break poin
y of breaks
945 almost
ed which ha
ace in 1966,
This can b
e started to
tinued to ri
he IO mode
el, the t-stati
bles is signi
breaks takin
gests that th
aks. In this
les indicate
gests breaks
l change by
recovering
joined the c
for trade in
was facing it
mpetitive el
nt occurs du
s occurring
every hous
ad a major a
, after this b
be explained
increase it
ise the dem
el cannot
istics for
ficant in
ng place
he series
s model
that this
s in 1943
y the IO
after the
common
n electric
ts lowest
lectricity
uring the
g during
sehold in
affect on
break the
d by the
became
mand for
electrici
Additio
where t
seemed
Figure
4.6 Com
In the fi
the nom
the nom
significa
tests the
breaks i
the nom
results i
at a five
and real
Table 4
Variable
LELN
LELR
ity has als
nally, this
the break w
likely that
4.6 Clemen
mparison
irst stage tes
minal and the
minal electri
ance level.
e results ten
in the Zivot
minal price f
in table 4.4
e percent si
l price of el
4.4 Summar
es AD
Unit
Unit
so continue
break date
was chosen
this series w
nte-Montañ
sts for unit r
e real electr
icity price
Perron arg
nd to be bia
t-Andrews
for electrici
shows that
gnificance
ectricity exh
ry of Unit R
DF test
Root U
Root U
ed to incr
coincides
from study
would exhib
ñés-Reyes I
root withou
ricity price.
cannot reje
gued that wi
ased toward
and Clemen
ity exhibits
for all mod
level. The c
hibits unit r
Root Tests
Zivot-And
Unit Root
Unit Root
27
rease which
with the br
ying the figu
bit two brea
IO Test, Re
ut allowing f
The results
ect the null
thout accou
ds unit root.
nte et al. tes
a unit root
dels the null
concluding
root even wh
drews
U
U
h is reflec
reak date s
ure of the s
aks.
eal Electric
for any stru
s in table 4.4
hypothesis
unting for st
However,
sts the resu
t process. F
hypothesis
remark is t
hen structur
Clemente e
AO mod
nit Root
nit Root
cted in a
suggested fr
series. As m
city Price
uctural break
4 indicate th
of unit roo
tructural bre
allowing fo
ults indicate
or the real
s of unit roo
therefore tha
ral breaks a
et al.
del
Un
Un
decrease in
from the Ch
mentioned b
ks was used
hat the ADF
ot at a five
eaks in the u
or both one
s the same,
electricity p
ot cannot be
at both the
are included
Clemente e
IO mode
nit Root
nit Root
n price.
how test
before it
d on both
F test for
e percent
unit root
and two
i.e. that
price the
rejected
nominal
d.
et al.
el
28
Table 4.5 Summary of Break Points
Variables Chow test Zivot-Andrews Clemente et al.
AO model
Clemente et al.
IO model
LELN 1977 1977 1945 and 1982 1917 and 1975
LELR 1968 1990 1943 and 1966 1938 and 2000
The results in table 4.5 for the nominal electricity price shows that the exogenously
determined break date in the Chow test coincides with the endogenously determined break
date in the Zivot-Andrews test. For the nominal electricity price the dummy variables
included in the Clemente et al. IO model only one were statistical significant. The only
significant endogenously determined breaks that can be identified in this series are the break
dates purposed by the Clemente et al. AO model. That is 1945 and 1982. This result is
interesting as one might have expected that the years around the First World War or at some
date near the oil crises to be characterized by structural change according to the data in figure
2.2. As mentioned before, almost every household in Sweden had electricity at 1945 and the
manufacturing industry developed demanding more electricity, which increased the demand
for electricity and the big devaluation of the Swedish krona is likely to be the reason for the
break in 1982.
The results for the real electricity price in table 4.5 shows that the Zivot-Andrews test
indicates a break during the financial crisis in Sweden in 1990, which does not coincides with
the exogenously determined break in the Chow test where the break date was chosen from
studying the figure of the real electricity price in chapter 2. However, looking at this figure
one might expect that the real electricity price exhibit more than one break. As mentioned
before only one of the dummy variables in the IO model was significant in explaining the real
electricity price. However, the second break in 2000 proposed by this model seems likely
according to the circumstances in the electricity market around this year. The AO model
allowing for two breaks taking place rapidly indicates two significant breaks. The first break
in 1943 cannot easily be detected in figure 2.2, while the second break almost coincides with
the exogenously determined break in the Chow test. After the break in 1968 the series clearly
changes pattern, it becomes almost flat. The change in pattern can be explained by the
substitution from oil towards electricity as the price of oil increases significantly from the late
29
1960’s and forward. This break is the one that coincides with the exogenously determined
break in the Chow test.
Comparing the results for the nominal and the real electricity price one concluding remark is
that something happened with the price of electricity during the period after the Second World
War that is around 1945 were the Clemente et al. AO model suggest a break for both the real
and nominal price. The exogenously and endogenously determined break coincides for both
series but while the nominal exogenously determined break coincides with the one-break
model the real electricity price series exogenously determined break coincides with the two-
break model. These results indicate that the nominal price most likely exhibits one break and
that the real electricity price is better explained by two breaks. This might be expected when
studying the series in figure 2.2 looking at parts where the series follows an even pattern.
30
Chapter 5
CONCLUSIONS
The purpose of this thesis was to examine the Swedish electricity price between the years of
1900 and 2006 in order to determine whether there are any structural breaks in the time series.
This was done by applying different quantitative methods for measuring structural breaks,
which made it possible to compare potential differences.
The analysis shows that for the nominal electricity price that exhibits much variation, none of
the tests that have been applied can reject the null hypothesis of unit root. The exogenously
determined break date suggested in the Chow test is being supported by the Zivot-Andrews
test for one endogenously determined break and partly by the Clemente-Montañés-Reyes IO
model. However, the significance of the two latter tests was not that high. It turned out that
the AO model was most significant in explaining the endogenous determined breaks in the
nominal electricity price series. This model assumes that the breaks takes place rapidly and
suggested 1945 and 1982 as break points in the series. The concluding remark according the
nominal price of electricity is that it exhibits unit root and that breaks are taking place rapidly.
Furthermore, the exogenously determined break did not coincide with the statistical
significant endogenous determined break date in this model. The expected break date was
between the two oil crises in the 1970’s. However, the AO model suggests 1945 and 1982 as
points of structural change.
The results suggested that the real electricity price is possessing unit root. The Zivot-Andrews
test indicates a break in 1990 that is during a financial crisis in Sweden that was caused by a
deregulation of the monetary system and a credit expansion. It is also two years before the
first deregulation of the Swedish electricity market which can explain a change in pattern in
the electricity price. The Clemente-Montañés-Reyes IO model indicates unit root in the real
electricity price but this model was not that significant in explaining the series. The AO model
was significant in explaining the series indicating two breaks taking place gradually in 1943
and 1966. The exogenous determined break in the Chow test takes place around the second
31
break suggested by the AO model. A concluding remark regarding the real electricity price is
that it exhibits unit root and that breaks most likely taking place rapidly. Most of the break
dates suggested by the different tests coincides with international and national crisis, policy
changes on the electricity market and changes in demand. The end of the Second World War
and the increase in demand during this period, the financial crisis in 1990, the massive
increase in the price of oil during the last four decades and the expansion of the common
Nordic electricity market in 2000 are changes that has affect the real price of electricity.
Because the exogenously determined break does not coincides with the endogenous
determined break in the one-break Zivot-Andrews model for the real electricity price one
concluding remark is that it is not always easy to detect a break only by studying a series
pattern. In the case of the real electricity price it seems more likely that the series exhibits
more than one break. These results implies that more than one test should be used and
compared when testing for structural breaks in time series.
As one might expect the nominal price is most affected by fluctuations in the inflation rate.
Both series are exhibits unit root and are best explained by structural changes that is taking
place rapidly. Additionally, the results indicate a break around 1945 for both series. This was
a period when the Swedish economy was strong, most households had electricity, the
manufacturing industry was growing and hence the demand for electricity was increasing.
32
REFERENCES
Andersson, B. (1997). Essays on the Swedish Electricity Market. Stockholm School of
Economics, Stockholm.
Bergström, V. (1999). Därför har Sverige bytt stabiliseringspolitisk regim. Address of First
Deputy Governor for the Swedish Riksbank.
Clemente, J., A. Montanés, A. and M. Reyes. (1998). Testing for a Unit Root in Variables
with a Double Change in the Mean. Economic Letters, vol. 59, pp.175-182.
Dougherty, C. (2007). Introduction to Econometrics third edition. Oxford University Press,
New York.
Energy Markets Inspectorate. (2007). Utveckling av nätpriser. 1 januari 1997- 1 januari
2007, report, ISSN 1653-8056. Swedish Energy Agency publication service, Eskilstuna.
Flatabo, N., G. Doorman, O.S. Grande, H. Randen and I. Wangensteen. (2003). Experience
with the Nord Pool design and implementation. IEEE Transactions on Power Systems, vol.18,
pp. 541-547.
Glynn, J., N. Perera and R. Verma. (2007). Unit Root Tests and Structural Breaks: A Survey
with Applications. Revista de metodos cuantitativos para la economia y la empresa, vol. 3,
pp. 63-79.
Hansen. E, B. (2001). The New Econometrics of Structural Change: Dating Breaks in U.S.
Labor Productivity. Journal of Economic Perspectives, vol. 15, pp.117-128.
Kander, A. (2002). Economic growth, energy consumption and CO2 emissions in Sweden
1800-2000. Doctor’s thesis, Department of Economic History, Lund University.
33
NE - National encyclopedia. (2008a). www.ne.se/sok/asymptot. 2008-12-09.
NE – National encyclopedia. (2008b). http://www.ne.se/artikel/1057810. 2008-12-10.
NE – National encyclopedia. (2009a).
http://www.ne.se/lång/riksbanken/293644/29364401?i_h_word=1990+ekonomi+svensk.
2009-02-19.
NE – National encyclopedia. (2009a). http://www.ne.se/kort/sverige. 2009-03-05.
Nord Pool. (2008). http://www.nordpoolspot.com/about/History/. 2008-11-16.
Perman, R. and J.P. Byrne. (2006). Unit Roots and Structural Breaks: A Survey of the
Literature. Department of Economics, University of Glasgow, Glasgow.
Perron, P. (1989). The Great Crash, the Oil Price Shock, and the Unit Root Hypothesis.
Econometrica, vol. 57, pp. 1361-1401.
Rönnborg, P. (2003). Borta med vinden. En analys av konkurrensen mellan leverantörer av
vindkraftverk i Sverige. Report nr. 43. Department of Technology and Society, Lund
University.
Saatcioglu, C. and L. Korap. (2008). Long-run Relations Between Money, Prices and Output:
The Case of Turkey. ZKÜ Sosyal Bilimler Dergisi, vol. 4, pp. 33-54.
SCB – Statistics of Sweden. (2008). Erik Marklund (Personal e-mail). Sender
([email protected]). (2008-11-14). Price of electricity between the years 2001 and 2006.
E-mail to the receiver. ([email protected]).
SCB – Statistics of Sweden. (2007). Inflation 1831-2007. Statistics,
www.scb.se/Statistik/PR/PR0101/2007M12/PR0101_2007M12_DI_06-07_SV.xls (2008-11-
14).
34
Swedish Energy Agency. (2006). Prisbildning och konkurrens på elmarknaden, report, ISSN
1403-1892. Swedish Energy Agency publication service, Eskilstuna.
Westerlund, J. (2005). Introduktion till ekonometri. Studentlitteratur, Lund.
Zivot, E. and W.K. Andrews. (1992). Further Evidence on the Great Crash, the Oil-Price
Shock, and the Unit-Root Hypothesis. Journal of Business & Economic Statistics, vol. 10, pp.
251-270.
Databases: EconLit, JSTOR, Google Scholar, Swedish Energy Agency, SCB.
Keywords: time series analysis, structural breaks, structural change, unit roots, Swedish
electricity market, Swedish electricity price.
Appendix 1
ADF test statistics
Nominal ADF Levels
T-stat
Prob
Lag length
Decision
Intercept -0.841 0.803 0 Non-stationary
Intercept + slope -1.240 0.897 0 Non-stationary
None -0.622 0.446 0 Non-stationary
Nominal ADF 1st Diff
T-stat
Prob
Lag length
Decision
Intercept -4.146 0.001 2 Stationary
Intercept + slope -4.361 0.004 2 Stationary
None -4.157 0.000 2 Stationary
Real ADF Levels
T-stat
Prob
Lag length
Decision
Intercept -1.437 0.5617 0 Non-stationary
Intercept + slope -2.136 0.5196 0 Non-stationary
None -4.033 0.0001 2 Stationary
Real ADF 1st Diff
T-stat
Prob
Lag length
Decision
Intercept -9.006 0.0000 1 Stationary
Intercept + slope -9.109 0.0000 1 Stationary
None -4.488 0.0000 2 Stationary
Appendix 2
Clemente-Montañés-Reyes Regression Results
Clemente-Montañés-Reyes Unit Root Test for LELR, IO model
AR(2) Du1 Du2 (rho – 1) Constant
Coefficients 0.07633 0.17002 0.04225 0.22080
T-statistics -2.422 3.485 -2.795
P-values 0.018 0.001 -5.490*
* 5% critical value.
Clemente-Montañés-Reyes Unit Root Test for LELN, IO model
AR(5) Du1 Du2 (rho – 1) Constant
Coefficients -0.05645 0.14603 -0.13117 -0.22272
T-statistics -1.573 4.318 -3.726
P-values 0.119 0.000 -5.490*
* 5% critical value.
Clemente-Montañés-Reyes Unit Root Test for LELR, AO model
AR(12) Du1 Du2 (rho – 1) Constant
Coefficients -1.31289 -1.14709 -0.20113 6.28187
T-statistics -14.632 -12.570 -3.684
P-values 0.000 0.000 -5.490*
* 5% critical value.
Clemente-Montañés-Reyes Unit Root Test for LELN, AO model
AR(3) Du1 Du2 (rho – 1) Constant
Coefficients -0.43087 0.99141 -0.34527 -1.71438
T-statistics -7.900 15.315 -4.700
P-values 0.000 0.000 -5.490*
* 5% critical value.