· pdf file · 2013-06-05title: microsoft word - s. piano dissertation.docx

Department of Economics University of Warwick

Dissertation

Student name Stefano Piano

Topic

The impact of European Economic Policy Uncertainty on the economic performance of the Eurozone

Abstract Recent events, such as the Eurozone debt crisis or the uncertainty surrounding the

European fiscal stimulus, have highlighted an unprecedented increase in uncertainty about economic policymaking in Europe. In this study, I construct a VAR model comprising of a European economic policy uncertainty index, devised by Baker, Bloom and Davis (2012), industrial production and control variables for the three biggest economies in the Eurozone: France, Italy and Germany. Impulse Response Function analysis suggests that a shock in the European policy uncertainty Index leads to a robust and persistent decline in industrial production in all countries. This ought to prove that European economic policy uncertainty has a negative impact on the economic performance of the Eurozone and that European policymakers ought to be blunter in their policy decisions.

Word count: 4997

Stefano Pisno The impact of EPU on the economic performance of the Eurozone

2

Table of Contents

1. Introduction 3

2. The existing literature 4 2.1 The literature about uncertainty and economic activity 4 2.2 The literature about economic policy uncertainty 4 2.3 The contribution of this study 5

3. Variables and their properties 6 3.1 Dataset: an overview 6 3.2 Stationarity analysis 7

4. The Model 10 4.1 Model set-‐up 10 4.2 Diagnostic tests 11

5. Impulse Response Functions 14 5.1 IRF results 14 5.2 Discussion 16

6. Concluding remarks 18

7. Bibliography 19 7.1 Cited works 19 7.2 Data sources and other useful references 20

8. Appendix 21 8.1 Appendix I: dataset description 21 8.2 Appendix II: summary statistics 22 8.3 Appendix III: tests 24 8.4 Appendix IV: Impulse Response Functions 26

Acknowledgements I am very grateful to Professor Christian Soegaard for his tireless support and to Professors Michael Clements and Gianna Boero for their suggestions. All errors are entirely my responsibility.


3

1. Introduction Several commentators have highlighted that ‘lingering policy uncertainty’ in Europe – as Pisani-‐Ferry (2012) put it – has seriously hampered the possibility of faster recovery and positive growth in the Eurozone. The European Central Bank, for instance, has argued that uncertainty about the Greek bailout and the resolution of the various stages of the Eurozone debt crisis has discouraged investment efforts and jeopardised the possibility of better economic performance over and on top of the financial consequences of these events (ECB, 2011). Yet, few scholars have discussed the matter with rigorous econometric methodology. In a recent working paper, Baker, Bloom and Davis (2012) have constructed an index to measure economic policy uncertainty in the United States and shown that it causes a consistent decline in economic performance. Using the European version of their index, this study aims to provide evidence that European economic policy uncertainty has a sizeable impact on the economic performance of the Eurozone. In order to achieve this objective, I will construct a Vector Autoregressive Model comprising of the European policy uncertainty index, interest rate, industrial production indexes – as a proxy for economic performance – the stock market price index – to control for overall economic and credit conditions – a realised volatility index – to control for the level of economic uncertainty – for the three biggest economies of the Eurozone: France, Germany and Italy. Impulse Response Functions computed from these models will show that a shock in European policy uncertainty generates a persistent decline in industrial production in all countries, robust to increasing the number of lags and changing the Cholesky ordering. I will argue that the persistence of this decline is explained by the special relationship between economic policy uncertainty and financial frictions. These results should primarily highlight the importance of blunt policymaking for European politicians and scholars interested in the European Union. However, they should also confirm that it is relevant to investigate the concept of economic policy uncertainty empirically and they should provide relevant insights to the scholars involved in this area. The remainder of this study is organised as follows. Section 2 provides a review of the literature on overall economic uncertainty and on economic policy uncertainty, and considers the contribution of this study. Section 3 offers a description of the dataset and tests the stationarity of the variables. Section 4 describes the model set-‐up and conducts diagnostic tests to identify the best specification. Section 5 computes Impulse Response Functions and proposes the financial frictions hypothesis to explain the findings. I conclude with some remarks on the necessary extensions to refine the conclusions of this study.


4

2. The existing literature 2.1 The literature about uncertainty and economic activity Numerous papers have demonstrated that higher uncertainty produces a decline in output and investment at the micro level. The mechanism through which this occurs was already identified by Bernanke (1983) and defined as investment irreversibility. Higher uncertainty leads to an increase in adjustment costs and consequently renders firm more reluctant to make investment decisions. Building on such microeconomic findings, macroeconomists have investigated whether uncertainty can be considered as one of the factors shaping the business cycle. Measuring uncertainty with a time-‐varying component, Bloom (2009) has shown that uncertainty shocks generate a “wait-‐and-‐see” effect. In line with the notion of investment irreversibility, firms become more cautious in their investment decisions in the short-‐run and this causes output to decline. However, in the medium-‐term output and investment recover and overshoot, as the excess capacity is utilised. Bachman, Elstner and Sims (2013) have confirmed the existence of similar relationship in Germany, using business surveys, news and volatility indexes to measure uncertainty. Other macroeconomists recently departed from this interpretation and showed that financial frictions of different kinds can lead to a more persistent decline in economic activity. Christiano, Motto and Rostagno (2010) have augmented a standard DSGE model to show that an uncertainty shock in the presence of adverse dynamics in the financial system – particularly adjustments in credit supply and unfavourable lending conditions – leads to a persistent decline in economic activity. Arellano, Bai and Kehoe (2012) have suggested that this occurs because financial frictions accentuate the problem of investment irreversibility. They increase the cost of capital in the short-‐to-‐medium term. This renders firms more cautious to take decisions for a longer period of time than under an alternative scenario where these frictions were not present. 2.2 The literature on economic policy uncertainty Frequently borrowing hypotheses and tools from the literature on general uncertainty, numerous studies have analysed the relationship between uncertainty in economic policymaking and microeconomic fluctuations. The most recent contribution was offered by Yulio and Yook (2012), who have estimated that federal elections in the USA cause corporate investment to decrease by 8%. Elections increase uncertainty about policymaking because they present firms with different alternative developments in economic policy. This leads to a drop in investment coherent with the notion of irreversibility. Baker et al. (2012) have recently explored the relationship between economic policy uncertainty and economic activity on a macroeconomic level for the USA, employing a VAR model typical of studies on overall economic uncertainty. They have measured economic policy uncertainty with a continuous index – a weighted average of a news index and


5

disagreement measures for fiscal and monetary policy – and showed that it leads to a persistent decline in investment, output and industrial production. Broogard and Detzel (2012) have extended their index to a panel of 25 countries and showed that it generates a drop in stock returns, output and corporate investment. However, their assumption that news about policy uncertainty only affects the economy domestically might be a source of measurement error, because it omits possibly relevant information. In both studies, the decline in economic activity from a shock in economic policy uncertainty is persistent. Although this seems to be coherent with the financial frictions hypothesis, no formal reason has been proposed as to why this might be the case. 2.3 The contribution of this study This study provides valuable insights to the literature about economic policy uncertainty. It confirms that economic policy uncertainty has a persistent impact on economic activity and it identifies in its special relationship with financial frictions the reason for this effect. Furthermore, it shows that economic policy uncertainty shocks in one country can have significant effects beyond its borders. This discourages the use of measures of uncertainty that limit a priori this transmission channel, as the one employed by Broogard and Detzel (2012). However, the contribution of this study could clearly stretch beyond this body of literature. Its results highlight that there is a common uncertainty factor in Europe, which has a strong influence on economic fluctuations. This ought to be relevant for scholars interested in the political economy of the European Union, in the Eurozone crisis – such as Pisani-‐Ferry (2012) – and for policymakers alike.


6

3. Variables and their properties 3.1 Dataset: an overview The present study employs a full monthly dataset ranging from January 1999 to December 2012 for the three biggest economies of the Eurozone: France, Germany and Italy. The time interval covers the years since the introduction of the ECB, when European Integration entered its current stage. The dataset features three country-‐specific variables – stock market price indexes, realised volatility indexes and industrial production indexes – the ECB interest rates and the European policy uncertainty Index. The stock market price indexes – 𝐶𝑎𝑐40,𝐷𝑎𝑥30 and 𝐹𝑡𝑠𝑒𝐼𝑡𝑎𝑙𝑖𝑎 – report the average of daily values for each month. The realised volatility indexes – V𝐶𝑎𝑐40, 𝑉𝐷𝑎𝑥30 and V𝐹𝑡𝑠𝑒𝐼𝑡 – were computed taking the monthly standard deviation of the stock market price indexes, following the method suggested by Broogard and Detzel (2012). As in Bachman et al. (2013) and Bloom (2009), they are employed as a proxy for overall economic uncertainty. For the industrial production variables – 𝐹𝑟_𝐼𝑛𝑑, 𝐺𝑒𝑟_𝐼𝑛𝑑 and 𝐼𝑡_𝐼𝑛𝑑 – I instead rely on the seasonally adjusted indexes provided by DataStream. The same goes of the ECB interest rates, 𝑖𝑟. The European economic policy uncertainty index – 𝐸𝑝𝑢 – is a weighted average of a Google news index and disagreement measures for inflation and budget deficit. An accurate description of the precautions adopted by Baker et al. (2012) for its construction is available in Appendix I. For the present purpose, it is essential to analyse two features of this dataset. First of all, it is important to show that the 𝐸𝑝𝑢 index is an accurate and specific measure of economic policy uncertainty. Secondly, it is crucial to consider whether the financial crisis that began in September 2009 causes anomalous fluctuations in the data, because this might cause problems of various kinds during data analysis and estimation (Perron, 1989). Figure 1 below provides some evidence to discuss both.

Figure 1. 𝑬𝒑𝒖


7

Figure 1 shows that the index captures very well European policy uncertainty dynamics. It peaks in relation to events that are associated to higher levels of policy uncertainty at the European level: wars, negotiations about the future of the EU, major policy proposals and various stages of the Eurozone debt crisis.1 Moreover, it shows that the level of European Economic policy uncertainty has increased dramatically after September 2008. Fluctuations of this kind are not restricted to the 𝐸𝑝𝑢 Index and are particularly apparent in the Industrial Production variables.2 Figure 2 below plots 𝐹𝑟_𝐼𝑛𝑑, 𝐺𝑒𝑟_𝐼𝑛𝑑 and 𝐼𝑡_𝐼𝑛𝑑.

Figure 2. Industrial production indexes

The three variables decline visibly after the bankruptcy of Lehman Brothers and then recover, albeit at very different rates. In both cases – for 𝐸𝑝𝑢 and the industrial production indexes –there is evidence that the financial crisis generates a structural break. This will pose various complications in the subsequent phases of this study. 3.2 Stationarity analysis Before constructing the VAR models in the next section, it is essential to clarify the order of integration of the variables in the dataset. This could be normally conducted through a standard Unit-‐Root test, such as the Augmented Dickey Fuller test. This is estimated on the three regressions

A: Δ𝑦! = 𝛽𝑦!!! + 𝛿!Δ𝑦!!!!!!! (1)

B: Δ𝑦! = 𝛼 + 𝛽𝑦!!! + 𝛿!Δ𝑦!!!

!!!! (2)

C: Δ𝑦! = 𝛼 + 𝛾𝑡 + 𝛽𝑦!!! + 𝛿!Δ𝑦!!!!!!! (3)

1 For an analysis of the relationships between 𝐸𝑝𝑢 and the volatility indexes see Appendix II. 2 Summary statistics and tests presenting evidence of a structural break for the other variables are available in Appendix.


8

where 𝛼 is a constant, 𝛾 is a time trend and under the null hypothesis 𝛽 = 1 the series is a Unit-‐root. However, the presence of a structural break in the variables renders this testing procedure inaccurate. The equations above do not account for the presence of the break. This might cause omitted relevant bias in the test statistic and lead to reject the null hypothesis less often than it is convenient to do so (Perron, 1989). Before addressing this complication, it is still useful to run a standard ADF test on all the variables that will be employed in the final VAR specifications. The table below reports these results.

Table 1. ADF testing* Variable

Test on the levels Test on the differences

p Model T-‐ Statistic 𝒑 Model T-‐ Statistic 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑) 3 C -‐2.556 2 B -‐6.040*** 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑) 5 C -‐2.602 4 B -‐5.144*** 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑) 3 C -‐2.562 2 B -‐4.699*** 𝑙𝑛(𝐶𝑎𝑐40) 7 C -‐2.711 6 B -‐3.739*** 𝑙𝑛(𝐷𝑎𝑥30) 6 C -‐2.046 5 B -‐4.884*** 𝑙𝑛(𝐹𝑡𝑠𝑒𝑖𝑡) 6 C -‐1.911 5 B -‐5.063*** 𝑉_𝐶𝑎𝑐40 1 B -‐6.112*** – – – 𝑉_𝐷𝑎𝑥30 1 B -‐6.019*** – – – 𝑉_𝐹𝑡𝑠𝑒𝑖𝑡 1 B -‐5.909*** – – – 𝑙𝑛(𝑖𝑟) 3 C -‐2.816 2 B -‐4.435*** 𝐸𝑝𝑢 0 C -‐3.903** – – –

*5% critical values are -‐2.886 for Model B and -‐3.442 for Model C. 1% critical values are -‐3.490 for Model B and -‐4.020 for Model C. ***, ** denote significance at the 1% and 5% level. I selected a preliminary value of 𝑝 using the Schwartz Information Criterion and I obtained the final value by running sequential LMAR tests for serial correlation. I have based the choice of the determinist trends on a visual inspection of the levels.

We fail to reject the null for 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑), 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑), ln(It_Ind) and 𝑙𝑛(𝑖𝑟) and the stock market price indexes, but they can all be thought to be I(1) with more than 99% confidence. However, we are able to reject the null of Unit Root on 𝐸𝑝𝑢 at the 5% level and on the realized volatility indexes at the 1% level. To account for the structural break, I will use the testing procedure developed by Zivot and Andrews (1993) and codified for Stata by Baum (2005). The procedure employs the augmented ADF equation

Δ𝑦! = 𝛼 + 𝜃𝐷! 𝜆 + 𝛾𝑡 + 𝛽𝑦!!! + 𝛿!Δ𝑦!!!

!

!!!

(4)

where 𝐷!(𝜆) = 1 𝑖𝑓 𝑡 > 𝑛𝜆,𝐷!(𝜆) = 0 otherwise and λ is the breakpoint fraction, such that 𝜆 = 𝑛!"#$% 𝑛. Using this equation, the Zivot-‐Andrews procedure computes the t-‐test statistic 𝛽 = 1 for every value of 𝜆. The minimum value from these tests is taken to be the final test statistic and to describe the date of the break. This approach has the advantage of identifying


9

the break endogenously instead of relying on an arbitrary identification through visual inspection as in Perron (1989). This increases the power of the test and provides evidence in support of the breakpoint. Below, I report the results of this test for the industrial production variables. I voluntarily omit the test results on the other variables because they do not introduce any significant innovation with respect to the standard ADF tests – however, they are available in Appendix III.

Table 2. Zinov-‐Andrews tests * Variable p Break point Test statistic 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑) 3 10/2008 -‐6.204*** 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑) 3 09/2008 -‐5.154** 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑) 3 08/2008 -‐5.096**

* 5% and 1% critical values are -‐4.80 and -‐5.43 respectively. ** and *** denote significance at the 5% and 1% level. p was selected using sequential t-‐tests, as suggested by Baum (2005).

At the 5% significance level, we reject the null hypothesis for 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑), 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑) and ln(It_Ind). Hence, the Industrial Production variables can be characterised as stationary with a break rather than as Unit Roots. Moreover, as anticipated by the visual inspection of the data earlier, the breakpoint occurs around the time of the financial crisis. These results are fully consistent with the findings of Zivot and Andrews (1993) and will be crucial for the estimation of the VAR models in the next section.


10

4. The model 4.1 Model set-‐up For each country, I intend to set-‐up the following VAR model:

𝑌! = 𝐴!𝑌!!! + 𝛾𝑡 + 𝜑!𝐷! + 𝑈! !

!!!

(5)

where the vector 𝑌! comprises of ln 𝐶𝑎𝑐40 , 𝑉𝐶𝑎𝑐40 , ln 𝑖𝑟 , 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑) ,𝐸𝑝𝑢 for France, ln 𝐷𝑎𝑥30 , 𝑉𝐷𝑎𝑥30 , ln 𝑖𝑟 , 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑) , 𝐸𝑝𝑢 for Germany, ln 𝐹𝑡𝑠𝑒𝐼𝑡 , 𝑉𝐹𝑡𝑠𝑒𝐼𝑡 , ln 𝑖𝑟 , 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑), 𝐸𝑝𝑢 for Italy, 𝐴! is the coefficient matrix, 𝑡 is a time trend and 𝐷! is a country-‐specific dummy variable accounting for the financial crisis.3 Taking log approximations for 𝑖𝑟, stock market and industrial production helps in the interpretation of results (Bloom, 2009). This model set-‐up ought to possess the two essential properties necessary to conduct meaningful IRF analysis: stationarity and a correct identification of European policy uncertainty shocks (Stock and Watson, 2001). Sim, Stock and Watson (1990) have shown that a VAR process ought to be stationary as long as the number of non-‐stationary variables is strictly smaller than the total number of variables and the non-‐stationary variables are co-‐integrated. The model set-‐up should fulfil their first requirement, because it features three stationary variables – 𝐸𝑝𝑢, the realised volatility indexes and industrial production – and two non-‐stationary variables. The second condition ought to be applicable, since appropriate Johansen tests – available in Appendix III – suggest the existence of co-‐integrating relationships. This model set-‐up should also lead to a correct identification of European euncertainty shocks, because it controls for the three factors that might be a source of bias: overall economic and credit conditions, monetary policy and the level of economic uncertainty. Economic policy uncertainty increases in the presence of adverse conditions in the financial system and negative economic performance, because policymakers are required to make extraordinary and possibly controversial decisions – as it is evident from Figure 1 above. The inclusion of the stock market should account for this, because the latter responds quite promptly to both (Christiano et al., 2010). For similar reasons, economic policy uncertainty has a negative correlation with monetary policy decisions and the inclusion of interest rates should control for this. Finally, economic policy uncertainty should be positively correlated with overall economic uncertainty. Following the footsteps of Broogard and Detzel (2012), the inclusion of the realised volatility indexes should account for this relationship. 3 𝐷! is equal to 1 from the country-‐specific date for the financial crisis to the date yielded by a Zinov-‐Andrews test on the post-‐crisis interval. For an explanation and the test results see Table 7 in Appendix III.


11

4.2 Diagnostic tests A successful specification of this model set-‐up ought to possess an adequate number of lags, uncorrelated residuals and normal residuals on the industrial production equation. The first useful step is to identify a sufficient number of lags. Ivanov and Kilian (2005) have shown that the Aikike Information Criterion performs best with VAR models constructed for IRF analysis. Following this insight, the tables below report the AIC value for specifications that include up to six lags.

Panel 1. Aikike Information Criterion*

𝒑 AIC 𝒑 AIC 𝒑 AIC 0 14.487 0 15.516 0 14.934 1 6.126 1 6.939 1 6.815 2 6.103 2 6.850 2 6.789 3 6.217 3 7.015 3 6.889 4 6.309 4 7.053 4 6.901 5 6.437 5 7.195 5 7.045 6 6.603 6 7.240 6 7.222

*The AIC was computed according to the standard formula 𝐴𝐼𝐶 = −2 (𝐿𝐿 𝑇) + (2𝑘 𝑇), where 𝐿𝐿 is the log-‐likelihood and 𝑘 is the number of parameters. A lower value indicates a better fit for model with 𝑝 lags.

Since for all countries the minimum AIC value occurs with the 𝑉𝐴𝑅(2) process, this will be considered as the baseline specification. It is fundamental to verify that these models actually fulfil the stationarity assumption, as hypothesised earlier. Lutkepohl (2005) has shown that a sufficient condition for the stationarity of a VAR process is the stability of the companion matrix. Taking the companion matrix for a VAR model with p lags

𝐴! 𝐴! ⋯ 𝐴!!! 𝐴!𝐼 0 ⋯ 0 00 𝐼 ⋯ 0 0⋮ ⋮ ⋱ 0 ⋮0 0 ⋯ 𝐼 0

(6)

this requires that

det (𝐼 − 𝐴!𝑧 −⋯− 𝐴!𝑧!) ≠ 0 (7) or alternatively that the solutions to this reverse characteristic polynomial lie outside the unit circle. The graphs on the following page plot these solutions.

France Germany Italy


12

Panel 2. Stability test, VAR(2) models

Since they lie outside the unit circle for all countries, VAR(2) models – and in general any 𝑉𝐴𝑅(𝑝) process derived from the model set-‐up – fulfil the stationarity assumption. It is essential to ensure that these models possess uncorrelated residuals, because this is a necessary condition for unbiased estimates. To check this property, I employ the Lagrange Multiplier test constructed by Johansen (1995). The results of this test for the 𝑉𝐴𝑅(2) models are provided below.

Panel 3. LMAR test, VAR(2) models*

𝒔 p-‐value 𝒔 p-‐value 𝒔 p-‐value 1 0.155 1 0.185 1 0.492 2 0.661 2 0.796 2 0.590 3 0.095 3 0.232 3 0.139 4 0.568 4 0.786 4 0.333 5 0.874 5 0.354 5 0.724 6 0.272 6 0.370 6 0.389

*The test statistic is computed for each lag 𝑠 as 𝐿𝑀! = 𝑇 − 𝑑 − 0.5 ln Σ Σ , where Σ is the ML estimate of the variance-‐covariance matrix of disturbances, Σ is the ML estimate of the variance-‐covariance matrix from a VAR augmented with lags of the residuals and 𝑑 is the number of coefficients in this VAR. I report the p-‐value from the 𝜒! distribution for a more concise representation. Under the null hypothesis, there is no serial correlation at lag 𝑠.

Since we always fail to reject the null of no serial correlation at the 5% level, a VAR(2) model is sufficient to obtain uncorrelated residuals for all countries. Lastly, it is important to test whether these models possess normal residuals. Although the normality of residuals is not a necessary condition for the consistency of the estimates, it provides evidence that the model is well specified (Johansen, 1995). Moreover, the test is useful to clarify whether conventional asymptotic standard errors are applicable or bootstrapped standard errors are desirable. The results of a VAR-‐modified

Italy Germany France

Germany France Italy


13

Jarque-‐Bera test – for the residual vectors and the residuals from the industrial production equations – can be found in the tables below.

Panel 4. Jarque-‐Bera test

Equation 𝑽𝑨𝑹(𝟐) p-‐value


𝑽𝑨𝑹(𝟑) p-‐value


𝐴𝐿𝐿 0.000*** 𝐴𝐿𝐿 0.000*** 0.000*** 𝐴𝐿𝐿 0.000*** 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑) 0.546 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑) 0.003*** 0.079* 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑) 0.139 *The J-‐B procedure for VAR models computes the standard test statistic 𝜆! = 𝜆! + 𝜆!, where 𝜆! and 𝜆! are skewness and kurtosis, for the orthogonalised residuals vector and for the residuals of each separate equation. I report p-‐value from the 𝜒! distribution for a more concise representation. Under the null, residuals are normally distributed. ***, **, * denote significance at the 1%, 5%, 10% level.

For all countries we fail to reject the null of normality on the residual vectors. However, we fail to reject the null in the residuals on the Industrial Production equation in the 𝑉𝐴𝑅(2) model for France and Italy at the 10% level and the VAR(3) model for Germany at the 5% level. Given that we only intend to derive inferences on industrial production, this is result is sufficient to consider the model well specified for the present purpose. From here, a VAR(2) process for France and Italy and a VAR(3) process for Germany are satisfactory baseline specifications of the model set-‐up. 4

4 Granger Causality tables in Appendix III confirm the merit of these specifications for the present purpose. Clearly, the VAR(3) model for Germany satisfies also the two other fundamental properties.

Germany Italy France


14

5. Impulse Response Functions 5.1 IRF results IRF analysis from non-‐structural VAR models is sensitive to at least two factors: the choice of Cholesky ordering and the computation of standard errors. Before presenting the IRF results, it is essential to justify the choices adopted to address these two problems. The use of Cholesky decomposition is necessary to separate the shocks in the variable of interest from the other variables. However, it requires establishing an ordering, which only allows some variables – the ones that come “ahead” – to have a contemporaneous effect on the others. Baker et al. (2013) employ the baseline ordering economic policy uncertainty, stock market, interest rate, employment, industrial production, which assumes that EPU index has a contemporaneous impact on all other variables. Following their footsteps, my baseline formulations feature the ordering European policy uncertainty, stock market price index, realised volatility index, interest rate, industrial production. The choice of standard errors is crucial to evaluate the statistical significance of the IRFs. Section 3 has highlighted complications in this sense because it has shown that the baseline VAR models do not possess normal residuals. This implies that the conventional asymptotic standard errors are no longer valid (Johansen, 1995). In this case, reliable standard errors can be computed through bootstrapping. This involves building a dataset of replicated statistics, fitting the original model to derive IRFs and then calculating standard errors with the usual formula (Guan, 2003). Impulse Response functions for the baseline specifications possessing these two features are plotted below. They report percentage instead of proportionate changes to facilitate interpretation, as in Baker et al. (2012).

Figure 3. Baseline IRF, France


15

Figure 4. Baseline IRF, Germany

Figure 5. Baseline IRF, Italy A shock by one standard deviation in European economic policy uncertainty causes a similar decline in industrial production in all countries, which reaches a minimum during the second quarter and remains statistically significant for about 6 quarters. In France – as clarified by the tables in Appendix IV – the decline reaches a minimum of -‐0.51% in the fifth month and it becomes statistically insignificant at the 18th month after the shock. In Germany, industrial production falls up to a minimum of -‐0.76% in the sixth month and converges to a statistically insignificant value after 17 months. In Italy, the impact on industrial production reaches a minimum of -‐0.47% in the seventh month and becomes statistically insignificant after 18 months. 5

5 The effect of the shock seems to be stronger in Germany, possibly due the higher volatility of the German economy (Bachman et al., 2013). However, analysing differences and similarities across countries is not strictly necessary in here and should be regarded as a fruitful extension of this project.


16

Although these results suggest the presence of a strong link between European uncertainty and industrial production, they do not clarify the potential magnitude of this effect. For the present purpose, it is sufficient to note that the impact of European policy uncertainty on industrial production is moderately large. 6 Consider that European economic policy uncertainty after the financial crisis has increased by two standard deviations with respect to the pre-‐crisis period. This shock might have produced a decline in industrial production about twice as big as the figures identified by these simple IRFs. However, it is essential to demonstrate that these effects are qualitatively robust to two checks: modifying the Cholesky ordering and increasing the number of lags. This is crucial to ensure that the results do not depend on some possibly incorrect assumptions about the behaviour of the variables and their relationships (Stock and Watson, 2001). The figures from Appendix IV show that increasing the number of lags up to six does not change the shape of effect significantly and it even results in a lower minimum. Similarly, employing the reversed Cholesky ordering stock market, realised volatility index, interest rates, industrial production, European policy uncertainty does not significantly affect the shape or duration of the shock -‐ as the effects remain statistically significant at the 10% level. These last results are particularly meaningful, because they suggest that a shock in 𝐸𝑝𝑢 generates a statistically significant decline in industrial production, even when European policy uncertainty is considered to be endogenous with respect to the other variables. Hence, there is significant evidence that a shock in Economic policy uncertainty produces a negative decline in industrial production, which protracts for approximately six quarters and is followed by a mild increase in industrial production. 5.2 Discussion The results from IRF analysis suggest that European policy uncertainty has a persistent effect on the industrial production of Eurozone economies. This is broadly consistent with Baker et al (2013) findings for the United States, as Figure 12 in Appendix IV demonstrates. The dynamics of these effects seem odd when compared to the classical results from the literature about economic uncertainty. There is no evidence for the “wait-‐and-‐see” effect that Bloom (2009) has identified for the United States and Bachman et al. (2013) have confirmed for Germany. This difference becomes clear when comparing the response of output to overall uncertainty from Bloom (2009) – available in Appendix IV – with responses to economic policy uncertainty. Identifying the reason for this discrepancy is crucial, because it would enable to clarify in what ways European economic policy uncertainty affects the economy.

6 One could of course use Monte-‐Carlo simulations to exemplify this point. This is clearly beyond the “exploratory” scope of this study and constitutes a desirable extension.


17

One very plausible explanation lies in the special relationship between financial frictions and economic policy uncertainty. As discussed above, uncertainty shocks under financial frictions are likely to lead to more persistent effects than uncertainty shocks in “normal” times. Economic policy uncertainty has a strong and mutually reinforcing relationship with financial frictions. As observed earlier, economic policy uncertainty shocks especially occur in the presence of adverse circumstances for the financial system, such as the Eurozone debt crisis, when governments and central banks need to identify solutions to avoid its collapse. Moreover, the presence of economic policy uncertainty further fuels financial frictions, because it increases the level of perceived risk in the financial sector (Christiano et al., 2010). In the virtue of this special relationship, economic policy uncertainty ought to generate a persistent shock, consistent with the financial frictions hypothesis, as opposed to one consistent with a “wait-‐and-‐see” dynamic. One obvious problem with this interpretation is whether the shock in economic policy uncertainty is only picking up the effect of these financial frictions. This should not be the case, at least in here, because the stock market should be controlling effectively for the impact of these frictions on the economy, as suggested by Christiano et al. (2010). Thus, European economic policy uncertainty – and economic policy uncertainty in general – leads to a persistent decline in industrial production, because it features a special relationship with financial frictions.


18

6. Concluding remarks The present study has demonstrated that European economic policy uncertainty has a significant and persistent impact on industrial production in the Eurozone. The persistence of this impact has been shown to depend on the special relationship between economic policy uncertainty and financial frictions. These results ought to encourage policymakers and scholars interested in the European Union to value prompter policy decisions. They also confirm that indexes are a good proxy for economic policy uncertainty and that in their construction it is important to take into account super-‐national events. Clearly, this study only proves the existence of a link between European economic policy uncertainty and economic performance in the Eurozone. A substantial amount of research has to be done to understand the transmission mechanism of economic policy uncertainty and validate the relevance of this link. It would be important the see whether European economic policy uncertainty has an effect over and on top of idiosyncratic policy uncertainty or one encompasses the other, but this would require the construction of ad-‐hoc measures – since the variability in Baker et al (2012) idiosyncratic measures and the pitfalls in Broogard and Detzel (2012) variables make this currently unfeasible. It would be also crucial to apply the European policy uncertainty index to different countries, to a different time period and to different macroeconomic variables to see if there is a significant difference between Eurozone and non-‐Eurozone countries and if the effects are qualitatively robust. Finally, it would be essential to see whether including different measures of overall economic uncertainty – such as business surveys – renders the effect of European policy uncertainty insignificant. Only the answers to these questions will allow to make a definite judgement about the findings of this study. These might give them momentum or highlight some shortcomings. Nevertheless, as far as this study is concerned, there is evidence that European policy uncertainty matters for the Eurozone and that economic performance would benefit from blunter policymaking.


19

7. Bibliography 7.1 Cited works Arellano, Cristina, Bai, Yan and Kehoe, Patrick (2012): “Financial Markets and Fluctuations in Uncertainty”, Federal Reserve Bank of Minneapolis Research Department Staff Report. Bachmann, Rüdiger, Elstener, Steffen and Sims, Eric (2013): “Uncertainty and Economic Activity: Evidence from Business Survey Data”, forthcoming American Economics Journal: Macroeconomics. Baker, Scott, Bloom, Nicholas and Davis, Steve (2012): “Measuring Economic Policy Uncertainty”, Stanford mimeo. Baum, Cristopher F. (2005): “ZANDREWS: Stata module to calculate Zivot-‐Andrews unit root test in presence of structural break”, available at: http://econpapers.repec.org/software/ bocbocode/s437301.htm. Bernanke, Ben (1983): “Irreversibility, Uncertainty and Cyclical Investment”, Quarterly Journal of Economics, 98 (1): 85–106. Bloom, Nicholas (2009): “The Impact of Uncertainty Shocks”, Econometrica, 77 (3): 623-‐685. Brogaard, Jonathan and Detzel, Andrew (2012): “The Asset Pricing Implications of Government Economic Policy Uncertainty”, University of Washington mimeo. Christiano, Lawrence, Motto, Roberto and Rostagno, Massimo (2010): “Financial Factors in Economic Fluctuations”, ECB Working Paper 1192. European Central Bank (2012): “Annual report, 2011”, available at: http://www.ecb.int/pub/ pdf/annrep/ar2011en.pdf. Guan, Weihua (2003): “Bootstrapped Standard Errors”, Stata Journal, 3 (1): 71-‐80. Ivanov, Ventzislav and Kilian, Lutz (2005): “A Practitioner's Guide to Lag Order Selection for VAR Impulse Response Analysis”, Studies in Nonlinear Dynamics & Econometrics, 9 (1): 1558-‐3708, ISSN (online). Johansen, Soren (1995): “Likelihood-‐Based Inference in Cointegrated Vector Autoregressive Models”, Oxford: Oxford University Press. Julio, Brandon and Yook, Youngsung (2012): “Political Uncertainty and Corporate Investment Cycles”, Journal of Finance, 67 (1): 45-‐83.


20

Lutketpohl, Helmut (2005): “New Introduction to Multiple Time Series Analysis”, New York: Springer. Perron, Pierre (1989): “The Great Crash, the Oil-‐Price Shock, and the Unit-‐Root Hypothesis”, Econometrica, 57 (6): 1361-‐1401. Pisani-‐Ferry, Jean (2012): “The Euro-‐crisis and the new impossible trinity”, Bruegel Policy Contribution, January. Sim, Cristopher, Stock, James and Watson, Mark (1990): “Inference in Linear Models with Some Unit-‐Roots”, Econometrica, 58 (1): 113-‐144. Stock, James and Watson, Mark (2001): “Vector Autoregressions”, Journal of Economic perspectives, 15 (4): 101-‐115. Zivot, Eric and Andrews, Donald (1992): “Further Evidence on the Great Crash, the Oil-‐Price Shock, and the Unit-‐Root hypothesis”, Journal of Business & Economic Statistics, 10 (3): 251-‐270. 7.2 Data sources and other references All financial and macroeconomic data were taken from DataStream. The European policy uncertainty index is publicly available at: http://www.policyuncertainty.com/europe_mon thly.html. For an introduction to VAR models and Impulse Response Functions, I used the notes for the University of Warwick MSc in Economics, which are available at: http://www2.warwick.ac.uk /fac/soc/economics/pg/modules/ec910/clements/. The cartoon in the cover page was first published in the weekly magazine “The Economist” on 7th December 2011 and is available at: http://www.economist.com/node/21541470.


21

8. Appendix 8.1 Appendix I: dataset description

Table 1. Full dataset description Variable Name Description

𝐹𝑟_𝐼𝑛𝑑 Real Industrial Production Index for France, excluding construction, seasonally adjusted (2005=100), taken from DataStream.

𝐺𝑒𝑟_𝐼𝑛𝑑 Real Industrial Production Index for Germany, excluding construction, seasonally adjusted (2005=100), taken from DataStream.

𝐼𝑡_𝐼𝑛𝑑 Real Industrial Production Index for Italy, excluding construction, seasonally adjusted (2005=100), taken from DataStream.

𝐶𝑎𝑐40 Monthly average for the Cac40 Price Index, computed with daily values taken from DataStream, standardised by its mean.

𝐷𝑎𝑥30 Monthly average for the Dax30 Price Index, computed with daily values taken from DataStream, standardised by its mean.

𝐹𝑡𝑠𝑒𝐼𝑡𝑎𝑙𝑖𝑎 Monthly average of FtseItalia Price Index, computed with daily values taken from DataStream, standardised by the mean.

𝑉𝐶𝑎𝑐40 Monthly standard deviation of daily values for the Cac40 price index index, standardised by the mean. Daily values taken from DataStream.

𝑉𝐷𝑎𝑥30 Monthly standard deviation of daily values for the Dax30 Price index, standardised by the mean. Daily values taken from DataStream.

𝑉𝐹𝑡𝑠𝑒𝐼𝑡 Monthly standard deviation of daily values for the FtseItalia Price index, standardised by the mean. Daily values taken from DataStream.

𝐼𝑟 ECB average cost of funds for banks, taken from DataStream. 𝐸𝑝𝑢 European Policy Uncertainty Index, computed from the weighted

average: 0.5𝑁𝑒𝑤𝑠 + 0.25𝐶𝑃𝐼𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦 + 0.25𝐵𝐷𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦. Mean is standardised to equal 100 prior to 2011. Taken from: http://www.po licyuncertainty.com/europe_monthly.html

𝑁𝑒𝑤𝑠 European Google News Index, measuring the number of search results for various combinations of “uncertainty”, “economy” and relevant expressions – such as “policy”, “tax”, “spending”, “regulation”, “central bank”, “budget”, and “deficit” – for 10 major European newspapers (El Pais, El Mundo, Corriere della Sera, La Repubblica, Le Monde, Le Figaro, The Financial Times, The Times of London, Handelsblatt, FAZ).

𝐶𝑃𝐼_𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦 Monetary policy uncertainty measure, reporting the average of the interquartile range of monthly inflation forecasts for the following calendar year for France, Germany, Italy, Spain and the UK. Forecasts originally taken from Consensus Economics.

𝐵𝐷_𝑈𝑛𝑐𝑒𝑟𝑡𝑎𝑖𝑛𝑡𝑦 Fiscal policy uncertainty measure, reporting the average of the interquartile range of monthly budget deficit forecasts for the following calendar year for France, Germany, Italy, Spain and the UK. Forecasts originally taken from Consensus Economics.

𝐶𝑟𝑖𝑠𝑖𝑠_𝐹𝑟𝑎𝑛𝑐𝑒 Dummy variable accounting for the financial crisis in France equal to one in in the interval 11/2008-‐01/2010 and zero otherwise.

𝐶𝑟𝑖𝑠𝑖𝑠_𝐺𝑒𝑟𝑚𝑎𝑛𝑦 Dummy variable accounting for the financial crisis in Germany equal to one in in the interval 10/2008-‐03/2010 and zero otherwise.

𝐶𝑟𝑖𝑠𝑖𝑠_𝐼𝑡𝑎𝑙𝑦 Dummy variable accounting for the financial crisis in Italy equal to one in in the interval 09/2008-‐01/2010 and zero otherwise.


22

8.2 Appendix II: descriptive statistics

Table 2. Summary Statistic for continuous variables* Variable Mean Mean in

01/1999-‐08/2008

Mean in 09/2008-‐ 12/2012

Standard Deviation

Correlation with Epu

𝐹𝑟_𝐼𝑛𝑑 96.880 99.857 90.240 4.951 -‐0.732 𝐺𝑒𝑟_𝐼𝑛𝑑 101.269 99.402 105.435 7.764 0.231 𝐼𝑡_𝐼𝑛𝑑 97.635 102.617 86.519 8.052 -‐0.782 𝐶𝑎𝑐40 100 108.037 82.072 23.198 -‐0.677 𝐷𝑎𝑥_30 100 96.078 108.749 23.929 -‐0.046 𝐹𝑡𝑠𝑒𝐼𝑡 100 115.183 66.129 29.566 -‐0.839 𝑉𝐶𝑎𝑐40 100 99.938 100.139 56.250 0.186 𝑉𝐷𝑎𝑥30 100 93.687 114.082 57.826 0.294 𝑉𝐹𝑡𝑠𝑒𝑖𝑡 100 102.095 95.326 60.986 0.088 𝐼𝑟 2.524 3.086 1.269 1.192 -‐0.546 𝐸𝑝𝑢 109.597 90.688 151.778 35.634 1.00

*The mean of the variables is generally different in the pre-‐crisis subsample as opposed to the post-‐crisis subsample and this supports the structural change hypothesis – standardising variables by the mean renders this evident for stock market variables. Appropriate Zivot-‐Andrews testing later will clarify the significance of these differences. 𝐸𝑝𝑢 has well-‐behaved correlation coefficients with all variables, positive with volatility indexes and negative with the other pro-‐cyclical variables. The positive correlation coefficient with the German industrial production index should not be of major concern, because it just reflects the fact that the two variables both trend upwards after 12/2010. The correlation coefficient in the precedent subsample is in fact a more reasonable -‐0.12.

Figure 1. 𝑪𝒂𝒄𝟒𝟎,𝑫𝒂𝒙𝟑𝟎,𝑭𝒕𝒔𝒆𝑰𝒕𝒂𝒍𝒊𝒂

Figure 2. 𝑰𝒓


23

Figure 3. 𝑽𝑪𝒂𝒄𝟒𝟎*

Figure 4. 𝑽𝑫𝒂𝒙𝟑𝟎*

Figure 5. 𝑽𝑭𝒕𝒔𝒆𝑰𝒕𝒂𝒍𝒊𝒂*

*The realised volatility indexes react more decisively than the 𝐸𝑝𝑢 index to shocks of financial nature, such as the change of policy by the FED and country-‐specific political events – elections and political difficulties. This confirms that the 𝐸𝑝𝑢 index is a specific measures of economic policy uncertainty at the European level.


24

8.3 Appendix III: Tests

Table 3. Zinov-‐Andrews test* Variable p Break point Test statistic Conclusion ln (𝐶𝑎𝑐40) 3 06/2008 -‐3.421 I(1) ln (𝐷𝑎𝑥30) 3 06/2008 -‐3.236 I(1) ln (𝐹𝑡𝑠𝑒𝐼𝑡) 3 06/2008 -‐3.338 I(1) 𝑉_𝐶𝑎𝑐40 2 06/2007 -‐5.629*** I(0) with break 𝑉_𝐷𝑎𝑥30 1 03/2007 -‐6.675*** I(0) with break 𝑉_𝐹𝑡𝑠𝑒𝐼𝑡 2 07/2007 -‐5.816*** I(0) with break ln (𝐼𝑟) 3 10/2008 -‐4.488 I(1) 𝐸𝑝𝑢 0 08/2008 -‐5.483*** I(0) with break

*p is as usual the number of augmenting lags in the ADF regression and was selected using sequential t-‐Tests. 5% and 1% critical values are -‐4.80 and -‐5.43, respectively. All variables exhibit a break due to the financial crisis apart from the volatility indexes, which feature an earlier break, possibly induced by the bailout of Northern Rock. These results do not modify in any significant way the conclusions from the ADF tests.

Table 4. Johansen tests for cointegration, France Maximum Rank T-‐Statistic1 5% Critical Value 𝑯𝟎

0 181.200 68.52 𝑅𝑎𝑛𝑘 = 0 1 98.031 47.21 𝑅𝑎𝑛𝑘 ≤ 1 2 32.205 29.68 𝑅𝑎𝑛𝑘 ≤ 2 3 6.970* 15.41 𝑅𝑎𝑛𝑘 ≤ 3

*There are three co-‐integrating relationships. This suggests that Sim et al. (1990) conditions could be satisfied, because the Unit-‐Root variables ought to be co-‐integrated. The fact that these tests do not yield a full-‐rank result is not concerning for the stationarity of the process, because they do not allow a priori for the possibility of a break – something that the stability condition clearly allows for.

Table 5. Johansen tests for cointegration, Germany Maximum Rank T-‐Statistic 5% Critical Value 𝑯𝟎

0 156.996 68.52 𝑅𝑎𝑛𝑘 = 0 1 82.526 47.21 𝑅𝑎𝑛𝑘 ≤ 1 2 19.052* 29.68 𝑅𝑎𝑛𝑘 ≤ 2 3 3.872 15.41 𝑅𝑎𝑛𝑘 ≤ 3

*There are two co-‐integrating relationships. This suggests that Sim et al. (1990) conditions could be satisfied, because the Unit-‐Root variables ought to be co-‐integrated.

Table 6. Johansen tests for cointegration, Italy Maximum Rank T-‐Statistic 5% Critical Value 𝑯𝟎

0 182.027 68.52 𝑅𝑎𝑛𝑘 = 0 1 111.732 47.21 𝑅𝑎𝑛𝑘 ≤ 1 2 54.343 29.68 𝑅𝑎𝑛𝑘 ≤ 2 3 8.474* 15.41 𝑅𝑎𝑛𝑘 ≤ 3

*There are three co-‐integrating relationships. This suggests that Sim et al (1990) conditions could be satisfied, because the Unit-‐Root variables ought to be co-‐integrated.


25

Table 7. Zinov-‐Andrews test in the country-‐specific post-‐crisis interval* Variable p Break point Test statistic Conclusion ln(Fr_Ind) 3 01/2010 -‐5.094** I(0) with break ln(Ger_Ind) 3 03/2010 -‐4.889** I(0) with break ln(It_Ind) 3 01/2010 -‐4.981** I(0) with break

*p is as usual the number of augmenting lags in the ADF regression and was selected using sequential t-‐Tests. 5% and 1% critical values are -‐4.80 and -‐5.43 respectively. Test is estimated in the interval 11/2008-‐12/2012 for France, 10/2008-‐12/2012 for Germany, 09/2008-‐12/2012 for Italy. Dummy variable takes the value one from the starting date of this sub-‐sample to the date of the break. Together with a time trend, this allows accounting for the break caused by the financial crisis, with a more synthetic representation with respect to the one with a dummy variable taking the value one for the entire post-‐crisis sample.

Table 8. Granger Causality tests, VAR(2) France* Dependent Variable

Excluded: 𝑙𝑛(𝑐𝑎𝑐40)

Excluded: 𝑉𝑐𝑎𝑐40

Excluded: 𝑙𝑛(𝑖𝑟)

Excluded: 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑)

Excluded: 𝐸𝑝𝑢

Excluded: 𝐴𝐿𝐿

𝑙𝑛(𝐶𝑎𝑐40)! – 0.028** 0.000*** 0.063* 0.765 0.003*** 𝑉𝐶𝑎𝑐40! 0.244 – 0.002*** 0.055* 0.037** 0.000*** 𝑙𝑛(𝑖𝑟)! 0.000*** 0.107 – 0.064* 0.062* 0.000***

𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑)! 0.481 0.288 0.892 – 0.000*** 0.000*** 𝐸𝑝𝑢! 0.003*** 0.602 0.020** 0.001*** – 0.002***

*The table reports Granger Causality statistics computed according to the standard formulation. ***, **, * denote significance at the 1%, 5% and 10% level respectively. Under the null hypothesis, the dependent variable is not “Granger-‐caused” by the “excluded” variable. The tests show that every variable is “Granger-‐Caused” at the 1% level by the linear combination of the other variables and that 𝐸𝑝𝑢 “Granger-‐causes” 𝑙𝑛 (𝐹𝑟_𝐼𝑛𝑑) at the 1% level. This suggests that there are strong links among all the variables in the system and that past values of 𝐸𝑝𝑢 have a significant influence on 𝑙𝑛 (𝐹𝑟_𝐼𝑛𝑑). From here, the model is suitable to the present purpose.

Table 9. Granger Causality tests, VAR(2) Germany Dependent Variable

Excluded: 𝑙𝑛(𝐷𝑎𝑥30)

Excluded: 𝑉𝐷𝑎𝑥30

Excluded: 𝑙𝑛(𝑖𝑟)

Excluded: 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑)

Excluded: 𝐸𝑝𝑢


𝑙𝑛(𝐷𝑎𝑥30)! – 0.056* 0.014** 0.351 0.135 0.028** 𝑉𝐷𝑎𝑥30! 0.025** – 0.091* 0.286 0.000*** 0.000*** 𝑙𝑛(𝑖𝑟)! 0.028** 0.308 – 0.475 0.000*** 0.000***

𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑)! 0.000*** 0.074* 0.028** – 0.012** 0.000*** 𝐸𝑝𝑢! 0.006*** 0.418 0.919 0.362 – 0.033**

*The table reports Granger Causality statistics computed according to the standard formulation. ***, **, * denote significance at the 1%, 5% and 10% level respectively. Under the null hypothesis, the dependent variable is not “Granger-‐caused” by the “excluded” variable. A 𝑉𝐴𝑅(2) model was employed, instead of the baseline 𝑉𝐴𝑅(3) process, because Granger Causality tests are particularly sensitive to potentially insignificant lags. The tests show that every variable is “Granger-‐caused” at the 5% level by the linear combination of the other variables and that 𝐸𝑝𝑢 “Granger-‐causes” 𝑙𝑛 (𝐺𝑒𝑟_𝐼𝑛𝑑) at the 5% level. This confirms that the model is well suited to the present purpose.

Table 10. Granger Causality tests, VAR(2) Italy Dependent Variable

Excluded: 𝑙𝑛(𝐹𝑡𝑠𝑒𝐼𝑡)

Excluded: 𝑉𝐹𝑡𝑠𝑒𝐼𝑡!

Excluded: 𝑙𝑛(𝑖𝑟)!

Excluded: 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑)!

Excluded: 𝐸𝑝𝑢!


𝑙𝑛(𝐹𝑡𝑠𝑒𝐼𝑡)! – 0.025** 0.001*** 0.076* 0.712 0.006*** 𝑉𝐹𝑡𝑠𝑒𝐼𝑡! 0.852 – 0.001*** 0.001*** 0.090* 0.001*** 𝑙𝑛(𝑖𝑟)! 0.004*** 0.320 – 0.167 0.400 0.000***

𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑)! 0.001*** 0.828 0.318 – 0.088* 0.000*** 𝐸𝑝𝑢! 0.000*** 0.712 0.000*** 0.001*** – 0.000***

*The table reports Granger Causality statistics computed according to the standard formulation. ***, **, * denote significance at the 1%, 5% and 10% level respectively. Under the null hypothesis, the dependent variable is not “Granger-‐caused” by the “excluded” variable. The tests show that every variable is “Granger Caused” at the 1% level by the linear combination of the other variables and that 𝐸𝑝𝑢 “Granger-‐causes” 𝑙𝑛 (𝐼𝑡_𝐼𝑛𝑑) at the 10% level. This confirms that the model is well suited to derive meaningful IRFs for 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑).


26

8.4 Appendix IV: Impulse Response Functions

Table 11. IRF, France, VAR (2), Baseline Ordering* Months after

shock Orthogonalised IRF for

𝒍𝒏(𝑭𝒓_𝑰𝒏𝒅) 90% Boostrapped Confidence

Interval 2 -‐0.399% [-‐0.596%, -‐0.203%] 4 -‐0.504% [-‐0. 700%, -‐0.308%] 5 -‐0.511% [-‐0.707%, -‐0.315%] 6 -‐0.504% [-‐0.701%, -‐0.307%] 8 -‐0.470% [-‐0.667%, -‐0.272%] 12 -‐0.355% [-‐0.546%, -‐0.164%] 16 -‐0.216% [-‐0.393%, -‐0.039%] 18 -‐0.146% [-‐0.315%, 0.022%] 20 -‐0.081% [-‐0.242%, 0.080%] 24 0.029% [-‐0.120%, 0.177%] 36 0.132% [0.005%, 0.259%]

*Outlines the percentage change in 𝐹𝑟_𝐼𝑛𝑑 for a shock by one standard deviation in 𝐸𝑝𝑢. I report in here the confidence interval from the final log file but I have simulated the errors multiple times to ensure that conclusions did not depend on an “anomalous” simulation. The only case where the simulation outcomes affected inferences was the 36th month in the IRF for France, because confidence intervals included the value zero in some instances. This suggests that the annexed coefficient is best considered as statistically insignificant.

Table 12. IRF, Germany, VAR (3), Baseline Ordering* Months after


𝒍𝒏(𝑮𝒆𝒓_𝑰𝒏𝒅) 90% Boostrapped Confidence

Interval 2 -‐0.267% [-‐0.473%, -‐0.060%] 4 -‐0.656% [-‐0.875%, -‐0.441%] 6 -‐0.759% [-‐1.012%, -‐0.506%] 8 -‐0.726% [-‐1.007%, -‐0.446%] 12 -‐0.534% [-‐0.840%, -‐0.229%] 16 -‐0.317% [-‐0.608%, -‐0.026%] 17 -‐0.265% [-‐0.549%, 0.019%] 20 -‐0.123% [-‐0.379%, 0.133%] 24 0.025% [-‐0.191%, 0.242%] 36 0. 175% [-‐0.008%, 0.359%]

* Reports the percentage change in 𝐺𝑒𝑟_𝐼𝑛𝑑 for a shock by one standard deviation in 𝐸𝑝𝑢.

Table 13. IRF, Italy, VAR (2), Baseline Ordering* Months after


𝒍𝒏(𝑰𝒕_𝑰𝒏𝒅) 90% Boostrapped Confidence

Interval 2 -‐0.359% [-‐0.594%, -‐0.123%] 4 -‐0.448% [-‐0.692%, -‐0.203%] 6 -‐0.466 % [-‐0.726%, -‐0.207%] 7 -‐0.467% [-‐0.735%, -‐0.200%] 8 -‐0.465% [-‐0.740%, -‐0.189%] 12 -‐0.417% [-‐0.709%, -‐0.126%] 16 -‐0.323% [-‐0.604%, -‐0.042%] 18 -‐0.265% [-‐0.533%, 0.004%] 20 -‐0.204 % [-‐0.457%, 0.049%] 24 -‐0.083% [-‐0.303%, 0.139%] 36 0.154% [-‐0.050%, 0.360%]

* Reports the percentage change in 𝐼𝑡_𝐼𝑛𝑑 for a shock by one standard deviation in 𝐸𝑝𝑢.


27

Figure 6. France, 𝑽𝑨𝑹(𝟔), Baseline Ordering

Figure 7. Germany, 𝑽𝑨𝑹(𝟔), Baseline Ordering

Figure 8. Italy, 𝑽𝑨𝑹(𝟔), Baseline Ordering


28

Figure 9. France, Reversed Cholesky Ordering*

*Reversed Ordering is 𝑙𝑛(𝐶𝑎𝑐40),𝑉𝐶𝑎𝑐40, 𝑙𝑛(𝑖𝑟), 𝑙𝑛(𝐹𝑟_𝐼𝑛𝑑), 𝑒𝑝𝑢

Figure 10. Germany, Reversed Cholesky Ordering*

*Reversed Ordering is 𝑙𝑛(𝐷𝑎𝑥30),𝑉𝐷𝑎𝑥30, 𝑙𝑛(𝑖𝑟), 𝑙𝑛(𝐺𝑒𝑟_𝐼𝑛𝑑), 𝑒𝑝𝑢

Figure 11. Italy, Reversed Cholesky Ordering*

*Reversed Ordering is 𝑙𝑛 𝐹𝑡𝑠𝑒𝐼𝑡 ,𝑉𝐹𝑡𝑠𝑒𝐼𝑡, 𝑙𝑛(𝑖𝑟), 𝑙𝑛(𝐼𝑡_𝐼𝑛𝑑), 𝑒𝑝𝑢. Note that the presence of a positive effect in the first lag ought not to be of major concern, given that it is largely statistically insignificant.


29

Figure 12. Estimated Industrial Production response to an EPU shock in the US *

**Taken from Baker, Bloom and Davis (2012). The magnitude of the Epu shock is the difference between the post-‐crisis level and the pre-‐crisis level. The response was computed using Monte-‐Carlo simulation techniques.

Figure 13. Estimated Output response to Uncertainty shock in the US*

*Taken from Bloom(2009). Uncertainty is measured with a time-‐varying component based on the VIX index of stock market volatility. Bloom (2009) names this effect volatility-‐overshoot.

· pdf file · 2013-06-05title: microsoft word - s. piano dissertation.docx

Documents