CROSS-COUNTRY CORRELATION: CASE OF THE BALTIC STATES Arvydas Jadevičius Royal Agricultural University Stroud Road Cirencester GL7 6JS United Kingdom E-mail: a.jadevicius@gmail.com Ignas Goštautas Lietuvos bankas Gedimino pr. 6 01103 Vilnius E-mail: igostautas@lb.lt Introduction 95 *For more about the importance of geographical proximity for the economy and transmission of economically productive knowledge, see Rodríguez-Pose and Crescenzi (2008). There is a widespread belief among economic commentators that events in Estonia lead the other two Baltic economies by around six months. It is often considered that the economic growth prospects of Latvia and Lithuania should be seen against the backdrop of the observed economic activity in Estonia. Thus, observing the Estonian gross domestic product (GDP) growth by 1.1 per cent in Q1 2015, one can expect modest growth rates for Latvia and Lithuania by the end of the year, or so the popular thinking goes. The cultural, historical and economic proximity of the three countries indicates a high degree of resemblance among Estonian, Latvian, and Lithuanian economies. Given the degree of common factors among the three economies, it can be the case that any statistical lead/lag relationship is a product of the result of systemically faster and slower country responses to exogenous events. The response time to external economic shocks could be a consequence of differences in external and internal structural development factors among the Baltic States, which can point to Estonia s potential lead role. When it comes to external factors, one potential explanation could be linked to Estonia s geographical proximity and deeper interlinkages with more advanced Northern European economies (of course, assuming that these connections play a role in economic convergence and cyclical fluctuations).* In addition, Estonia also was the point of entrance into the Baltics for the international credit institutions, which contributed substantially to credit growth in the three countries. With regard to internal influences, Estonia is a more open economy as its exports constituted the largest part of GDP among the three Baltic States during the analysed period (on average it was 74.7%, 50.5% and 67.0% for Estonia, Latvia and Lithuania respectively), which could have made it more sensitive to outside shocks. In addition, Estonia is the smallest of all three nations (in terms of the economy, territory, and population), which naturally makes it less diversified and more likely to be promptly affected by external shocks. This study investigates the dynamics of GDP and housing data series of the three Baltic States over the period from Q1 2006 to Q4 2014 and tests the popular belief that Estonian economic indicators help foresee the future economic dynamics in the other two countries. The results indicate that Estonian GDP series to a certain extent lead the Arvydas Jadevičius is a Doctor of Social Sciences, Lecturer at the School of Real Estate and Land Management, Royal Agricultural University. Area of activity: real estate and land management. Ignas Goštautas is an Economist at the Macroeconomics and Forecasting Division of the Economics Department of the Economics and Financial Stability Service at the Bank of Lithuania and a doctoral student at Nottingham Trent University. Areas of activity: real estate, finance.
respective series in Lithuania and Latvia. In contrast, Latvia was shown to lead the dynamics of the Baltic housing markets. It should be noted, however, that the additional analysis of data series split into pre- and post-crisis episodes does not confirm the robustness of the results, ascribing the statistical lead/lag relationships mainly to the pre-crisis period. The paper is organised as follows: Section 1 reviews previous studies on the subject. Section 2 presents the data and discusses the statistical methodology and Section 3 presents modelling results. Pinigø studijos 2015/2 Kitos publikacijos 96 1. Previous studies A discussion on international linkages between economies can be traced back to Ricardo s (1821) work. A century later, commentators including Williams (1929), Samuelson (1938, 1939) and Flux (1933) analysed cross-country relationships with an emphasis on international trade. The subject gained traction over the last two decades following the seminal studies by Krugman (1993, 1997) and his work on international economics (including trade theory, economic geography, and international finance). With regard to cross-country business cycle correlations, Frankel and Rose (1998) suggested that countries with closer trade linkages tend to have more tightly correlated business cycles. The authors used the correlation analysis to examine annual series over the period from 1959 to 1993 for 21 industrial economies. All series were also smoothed using four different HP-filter variations. The authors estimated a regression equation, examining links between a pair of countries based on bilateral trade intensity and auxiliary factors affecting trade. Heathcote and Perri (2003) examined the interconnection between the United States (US) economy and its foreign trade partners. The authors employed quarterly series covering the period from Q1 1960 to Q2 2002 for the US and 15 countries of the European Union (EU), plus Japan. Variables under consideration were GDP, fixed capital formation, consumption and public sector employment. All series were then smoothed using HP and high-pass filters. Based on smoothed series, a general equilibrium model was developed. The model showed that US business cycles became less synchronised with the rest of the world over the analysed period. According to the authors, increased financial integration among the nations was the key factor behind expanding cross-border correlation. Crosby and Bodman (2005) assessed linkages between Australian and US business cycles. The analysis of 130-year time series over the period from 1870 to 2000 led Crosby and Bodman to conclude that a significant post-war correlation between US and Australian GDP was driven by the recessions in the 1980s and 1990s and subsequent hikes in interest rates. The authors also tested their hypothesis against 8 other economies, including Canada, France, Germany, Italy, Japan, New Zealand and United Kingdom (UK). The time series smoothing (HP and BP filtering) and correlation analysis demonstrated an increased synchronisation between Australia and the US especially after the 1980s a pattern that is in contrast with the other OECD (Organization for Economic Cooperation and Development) economies investigated in Heathcote and Perri s (2003) study. The above-discussed studies provide some important insights about the subject of cross-border correlation. A greater (lesser) integration between countries implies lower (greater) interconnection of economies. An increased level of intra-border trade and financial integration is likely to result in a higher level of business cycle correlations. The three Baltic States were mostly analysed in the broader regional context of Central and Eastern Europe (CEE). Maneschiöld (2006) investigated co-integration and causality among Baltic stock markets and major international stock markets from 1996 to 2004. Among other results, Maneschiöld found that the Estonian stock market influences the Lithuanian stock market, but not vice versa. Gardo and Martin (2010) assessed economic developments in Central, Eastern and Southeastern European countries, including Estonia and Lithuania. The authors focused on the strengths and weaknesses of these economies, as well as the impact of the global financial crisis on their development. Their econometric model showed that the global
financial crisis hit the Baltic nations harder in comparison to their CEE peers. The authors claimed that a higher level of financial interconnection was one of the key reasons why the Baltic States were among those worst affected by the crisis. Huynh-Olesen et al. (2013) investigated housing markets in the CEE countries. Their cointegration analysis suggested that demand-side fundamentals and factors relating to the state of economic transition played a central role in explaining house price variations. Variables, such as disposable income, population, interest rates, acquisition of housing financed by foreign remittances and credit growth were significant across the region. Demian (2011) examined financial market linkages among CEE countries and found a common stochastic trend for the region economies*. Nikkinen et al. (2012) looked into the stock market integration of the three Baltic States and concluded that the markets of the three countries were integrated. The authors noted that interconnectedness was growing over time, though the markets took somewhat diverging paths prior to the 2008 2009 crisis. A number of studies showed that the economies of the Baltic States are tightly interlinked and have many common factors that may influence their quite synchronised economic fluctuations. Notably, the legislative framework in all three countries was harmonised in the process of European integration. Also, many international corporations see the Baltic States as one integrated region various banking, media, telecommunication and retail groups established their subsidiaries in all three countries. It should also be noted that for more than a decade the Baltic States had their currencies pegged to the euro and currently are members of the euro area. 97 2. Data and methodological framework To examine the interdependence among Estonian, Latvian, and Lithuanian GDP series and house price indices, we apply two well-known methods. One of them is crosscorrelation, which allows for estimating a degree of co-movement between selected series for a range of time lags. In addition to that, Granger causality of the series is analysed. Cross-correlation is a useful measure for examining the degree of covariation of selected series. The problem with the correlation analysis is that it cannot be straightforwardly applied to trending data series. Therefore, in the actual analysis we employ a three-stage procedure to deal with these problems. First, common trends are removed from the examined series. This is done by determining the ARIMA model of the first data series and applying the model for both series. The best fitting ARIMA model is chosen based on Akaike Information Criterion. Then, residuals from both series are extracted. Finally, the residuals from both data series are cross-correlated as follows: r d = N Y Y X X i ( i ) i d = 1 ( + ) ( Y Y) ( X X) N 2 N i = 1 i i = 1 i+ d 2, where X i and Y i are residual values of the first and second series respectively at a time i and I + d, where d is spacing (lag or lead) between the data series. X and Y are their means. Granger causality is estimated using the methodology of Toda and Yamamoto (1995). Using the VAR framework, the subsequent set of equations is computed: Y = α + θy + + θ Y + β X + + β X + e t 1 t 1 k+ d t ( k+ d) 1 t 1 k+ d t ( k+ d) t, X = α + θ X + + θ X + βy + + β Y + e t 1 t 1 k+ d t ( k+ d) 1 t 1 k+ d t ( k+ d) t, *Lithuania was not included in the sample. where θ q are regression coefficients, a is a slope, e t is an error term at the time t, k is an optimal number of lags, and d is a maximum order of integration of the series. X i and Y i are dependent and explanatory variables accordingly. In addition to that, a VAR Granger Causality/Block Erogeneity Wald test was performed for the VAR model excluding additional lags (d) to test whether X does not Granger cause Y and vice versa.
To test the robustness of the results, the sample was split into two equal periods (from Q1 2006 to Q2 2010 and from Q3 2010 to Q4 2014). The cross-correlation and VAR Granger causality analysis was performed for the two separate periods. The first period includes pre-crisis and crisis periods, while the second represents relatively moderate growth in the post-crisis environment. Fig. 1. Estonian, Latvian, and Lithuania GDP growth compared to the previous quarter Pinigø studijos 2015/2 Kitos publikacijos 98 Source: Eurostat; authors calculations. Two economic variables are selected for the analysis, namely, the GDP and the House Price Index (HPI) series. In the case of GDP, as Stevenson and McGarth (2003) noted, it is quite representative of the overall economic activity. The importance of the housing market to the national economy and banking stability was well documented by Koetter and Poghosyan (2009). The GDP series are expressed in real terms, while the HPI series are deflated using country-specific GDP deflators. All data series are indexed starting from 2006 and are unit free. Figures 1 2 show dynamics of the selected quarterly series for the period from Q1 2006 to Q4 2014 (Table 1 of Appendix reports series summary statistics). Natural logarithms of the selected variables are used in of the econometric analysis. Fig. 2. Growth in housing prices in Estonia, Latvia and Lithuania compared to the previous quarter Source: Eurostat; authors calculations.
An initial visual inspection suggests a strong co-movement of the series. All three nations experienced growth in GDP and housing markets in 2006 2007, which was followed by a significant correction in 2008 2009 and a rebound in 2010. 3. Estimation results The cross-correlation results suggest that the Estonian GDP series leads the Lithuanian GDP series by one quarter, whereas the Latvian economy follows the Estonian and Lithuanian economies with a two-quarter lag (see Figure 3). With regard to the housing market, the Latvian series leads the other two Baltic housing price indices. Cross-correlations also indicate that house prices fluctuations in Estonia lag the Latvian series by one quarter. Both Latvian and Lithuanian house price series exhibit a bi-directional cross-correlation. In other words, changes in Latvian and Lithuanian house prices take place simultaneously. Fig. 3. Cross-correlation of filtered series Estonia GDP and Lithuania GDP Estonia GDP and Latvia GDP 99 Lithuania GDP and Latvia GDP Estonia HPI and Lithuania HPI Estonia HPI and Latvia HPI Lithuania HPI and Latvia HPI Source: formed by the authors.
Pinigø studijos 2015/2 Kitos publikacijos 100 The split sample analysis (see Table 2 and Figure 1 of the Appendix) reveals that in the subsample covering the pre-crisis and crisis episodes, the results are very similar to the fullsample results. The Estonian GDP series was leading the Lithuanian and Latvian economies by one and two quarters, respectively. The two latter GDP series exhibit bi-directional cross-correlation. However, cross-correlations in the post-crisis subsample appeared to be unsystematic and do not lend themselves to an intuitive explanation. Cross-correlations of the housing price series do not show interdependence in both subsamples. Before formally testing the Granger causality, the stationarity of all series is examined by using the Augmented Dickey-Fuller and Phillips-Perron tests. All data series are integrated of order 1 at the 5 per cent significance level, except the Latvian GDP, which is integrated of order 1 at the 10 per cent significance level. The optimal number of lags in Granger causality tests is determined by autoregressive lag order selection criteria, with the maximum number lags equal to six. The lag length selection is based on the Augmented Dickey-Fuller test, Schwarz information criterion, Hannan-Quinn information criterion, Final Prediction Error, and sequential modified LR test statistics. To ensure that the models are well specified, a set of additional tests was performed. Residuals were tested using the LM test, which rejected serial interdependence between variables. Residuals were then examined for normality, using Cholesky factorisation. It is shown that GDP EE GDP LT, GDP LV GDP EE and HPI EE HPI LT models exhibit an element of excess skewness, while GDP LV GDP EE and HPI EE HPI LV models show non-normal kurtosis in the residuals. The Jarque-Bera test fails to reject the hypothesis that skewness and excess kurtosis are different from zero, but the test may lack the statistical power due to a small sample size. We apply the White test to test for the heteroscedasticity of residuals. All except the HPILV-HPILT model indicated constant variance of the residuals, which warns that the model s estimates may be somewhat biased. Having performed the testing procedures, we estimate the VAR Granger causalities (VAR parameter estimates are available upon request). The Granger causality results (see Table 1) suggest that the Estonian GDP and housing market has predictive power for the Lithuanian series. The Estonian GDP series also Granger causes the Latvian GDP series, while the Lithuanian and Latvian GDP series have bi-directional Granger causality. Notably, the Latvian housing prices series has Granger causes Estonian and Lithuanian housing prices series. Table 1 Results of VAR Granger causality Series pair From first to second From second to first First Second p-value Significance p-value Significance GDP EE GDP LT 0.00 GDP LV GDP LT 0.00 0.63 0.00 GDP LV GDP EE 0.16 0.00 HPI EE HPI LT 0.00 0.25 HPI EE HPI LV 0.54 0.00 HPI LV HPI LT 0.00 0.72 Note: significant at 1% level. Source: formed by the authors. Similar to the case of cross-correlation analysis, the robustness check using split data series shows that the Granger causality results mainly apply to the pre-crisis and crisis period, whereas they are not generally corroborated for the post-crisis period. The dynamics of Estonian GDP series seemed to Granger cause the other two series in the first subsample, but not in the second subsample. In a similar vein, Latvian housing price series appears to lead the other two housing price series in the first, but not the second, subsample.
Conclusions Economic commentators occasionally assert that the level of economic activity in Estonia leads the other two Baltics economies by around six months. Though there could, in fact, be such interdependence among the Baltic economies, it might also be a subjective interpretation of selective events as there was no prior econometric assessment of this claim. The current study therefore examined whether fluctuations of economic activity and housing prices in Estonia lead fluctuations in similar Latvian and Lithuanian indicators. The study examined quarterly GDP and house price series, covering the period from Q1 2006 to Q4 2014. Cross-correlations and Granger causality tests were used to dissect the relationship. As the empirical results suggest, in the analysed period Estonia leads the Lithuanian and Latvian economy and the Lithuanian housing market. At the same time, Latvian and Lithuanian economies share bi-directional Granger causality. In contrast, both Estonian and Lithuanian housing price fluctuations are Granger caused by Latvian housing price movements. However, the results should be treated with caution, as they were shown not be robust when the data was split into two subsamples. The evidence that the Estonian economic activity Granger causes economic fluctuations in Latvia and Lithuania is based on the pre-crisis and crisis period data, but there is no Granger causality in the post-crisis episode. 101
Appendix Table 1 Growth series summary statistics (q/q) (per cent) Pinigø studijos 2015/2 Kitos publikacijos Estonia Latvia Lithuania GDP EE HPI EE GDP LV HPI LV GDP LT HPI LT Mean 0.31 0.21 0.35 0.58 0.56 0.48 Median 0.78 1.07 0.60 0.70 0.81 0.14 Standard deviation 2.57 6.24 2.38 5.35 2.66 4.91 Kurtosis 9.65 3.49 1.15 0.01 21.40 10.28 Skew. 2.61 1.46 0.90 0.87 4.16 2.14 Range 14.57 30.97 10.62 21.42 16.46 32.68 Min. 10.81 19.28-6,15 13.77 13.13 21.69 Max. 3.76 11.69 4.46 7.65 3.33 10.98 Count 35 35 35 35 35 35 Source: formed by the authors. 102 Table 2 Results of VAR Granger causality for the split period + Series pair Q1 2006 Q2 2010 Q3 2010 Q4 2014 From first to second From second to first From first to second From second to first First Second p-value Statistical p-value Statistical p-value Statistical p-value Statistical significance significance significance significance GDP EE GDP LT 0.00 0.37 0.12 0.19 GDP LV GDP LT 0.03 ** 0.77 0.00 0.34 GDP LV GDP EE 0.02 ** 0.04 ** 0.40 0.53 HPI EE HPI LT 0.00 0.97 0.48 0.95 HPI EE HPI LV 0.29 0.00 0.67 0.84 HPI LV HPI LT 0.00 0.07 * 0.49 0.28 Notes: + VAR Granger causality analysis of the split data was performed with two (plus order of integration) lags, instead of optimal number (plus order of integration) of lags due to data length limitations; *, **, denotes significance at the 10%, 5% and 1% levels respectively. Source: formed by the authors.
Fig. 1. Cross-correlation of filtered series Estonia GDP and Lithuania GDP, period 1 Estonia GDP and Lithuania GDP, period 2 Estonia GDP and Latvia GDP, period 1 Estonia GDP and Latvia GDP, period 2 103 Lithuania GDP and Latvia GDP, period 1 Lithuania GDP and Latvia GDP, period 2 Estonia HPI and Lithuania HPI, period 1 Estonia HPI and Lithuania HPI, period 2
Estonia HPI and Latvia HPI, period 1 Estonia HPI and Latvia HPI, period 2 Pinigø studijos 2015/2 Kitos publikacijos Lithuania HPI and Latvia HPI, period 1 Lithuania HPI and Latvia HPI, period 2 104 Source: formed by the authors.
References Crosby M., Bodman P. 2005: When the US Sneezes, Do We Need to Catch a Cold? Historical and Future Linkages between the Australian and US Business Cycles. The Changing Nature of the Business Cycle, 226 254. Demian C. V. 2011: Cointegration in Central and East European Markets in Light of EU Accession. Journal of International Financial Markets, Institutions and Money 21(1), 144 155. Flux A. W. 1933: Industrial Productivity in Great Britain and the United States. The Quarterly Journal of Economics 48(1), 1 38. Frankel J. A., Rose A. K. 1998: The Endogeneity of the Optimum Currency Area Criteria. Economic Journal 108(449), 1009 1025. Gardo S., Martin R. 2010: The Impact of the Global Economic and Financial Crisis on Central, Eastern and South-eastern Europe: A Stock-taking Exercise. ECB Occasional Paper No. 114. Heathcote J., Perri F. 2003: Why Has the U.S. Economy Become Less Correlated with the Rest of the World? The American Economic Review 93(2), 63 69. Huynh-Olesen D., Steiner K., Hildebrandt A. Wagner K. 2013: Residential Property Prices in Central, Eastern and Southeastern European Countries: The Role of Fundamentals and Transition- Specific Factors. Focus on European Economic Integration 2, 52 76. Koetter M., Poghosyan T. 2009: Real Estate Prices and Bank Stability. Journal of Banking & Finance 34(6), 1129 1138. Krugman P. 1993: Geography and Trade (Gaston Eyskens Lectures). MIT Press. Krugman P. 1997: Development, Geography and Economic Theory (Ohlin Lectures). MIT Press. Maneschiöld P. O. 2006: Integration Between the Baltic and International Stock Markets. Emerging Markets Finance & Trade 42(6), 25 45. Nikkinen J., Piljak V., Äijö J. 2012: Baltic Stock Markets and the Financial Crisis of 2008 2009. Research in International Business and Finance 26(3), 398 409. Ricardo D. 1821: On the Principles of Political Economy and Taxation. London: John Murray. Rodríguez-Pose A., Crescenzi R. 2008: Research and Development, Spillovers, Innovation Systems, and the Genesis of Regional Growth in Europe. Regional Studies 42(1), 51 67. Rose A. K., Engel C. 2000: Currency Unions and International Integration. NBER Working Paper No. 7872. Samuelson P. A. 1938: Welfare Economics and International Trade. The American Economic Review 28(2), 261 266. Samuelson P. A. 1939: The Gains from International Trade. The Canadian Journal of Economics and Political Science (Revue canadienne d Economique et de Science politique) 5(2), 195 205. Stevenson S., McGrath O. 2003: A Comparison of Alternative Rental Forecasting Models: Empirical Tests on the London Office Market. Journal of Property Research 20(3), 235 260. Toda H. Y., Yamamoto T. 1995: Statistical Inference in Vector Autoregressions with Possibly Integrated Processes. Journal of Econometrics 66(1 2), 225 250. Williams J. H. 1929: The Theory of International Trade Reconsidered. The Economic Journal 39(154), 195 209. 105