Measuring economic policy uncertainty for European emerging markets

Transcription

1 Measuring economic policy uncertainty for European emerging markets First draft - please do not cite without permission of the authors January 217 Alexander Kupfer a, Josef Zorn b a Department of Banking and Finance, University of Innsbruck (Austria). Tel.: , alexander.kupfer@uibk.ac.at. b Department of Banking and Finance, University of Innsbruck (Austria). Tel.: , josef.zorn@uibk.ac.at. Abstract We present a new index for economic policy uncertainty based on Google search volume. Search volume offers the advantage of being independent of language and newspapers availability. We construct the Google economic policy uncertainty (GEPU) index for 18 countries in Eastern Europe. In a vector autoregressive (VAR) model including government bond yields and squared returns as a proxy for volatility, we demonstrate that shocks of the GEPU index have an impact on both variables showing the expected sign. This implies that the GEPU index is able to capture economic policy uncertainty of the respective countries in Eastern Europe and can be a helpful tool for policy makers and investors. Keywords: Emerging markets, Eastern Europe, economic policy uncertainty, search volume

2 1. Introduction Political uncertainty matters. Its effect is unanimously reported in existing literature when analyzing the impact on economic variables, for instance, in the case stock markets (e.g., Pastor & Veronesi (212)), investment rates (e.g., Bloom (9)) or foreign direct investments (e.g., Julio & Yook (216)). A new field of research has recently been opened by Baker et al. (216) by focusing not only on political issues but also on economic uncertainty. The authors create their so-called economic policy uncertainty (EPU) index by counting the number of news articles with specific keywords combinations. For the United States their baseline EPU index is created by the amount of articles in leading newspapers including the words economic (or economy ), uncertain (or uncertainty ) plus one of the following keywords: congress, deficit, Federal Reserve, legislation, regulation, White House. The authors show that their index captures important events very well in US history such as presidential elections, Gulf Wars etc. 1 Up to now (November 216), the authors provide their EPU index also for the following countries on their website: Australia, Brazil, Canada, Chile, China, France, Germany, India, Ireland, Italy, Japan, Korea, the Netherlands, Russia, Spain and the United Kingdom. In the case of Germany, for instance, the authors have a list of 13 keywords including economy und uncertainty plus words related to (i) taxes, (ii) regulation and (iii) government debt (see the Appendix of Baker et al. (216) for a full list of keywords). In this case, relevant articles are gathered from two major newspapers (i.e., Handelsblatt and Frankfurter Allgemeine Zeitung). For the United States, the authors not only show that their EPU index works quite well on specific events (see above) but also to explain corporate investments rates, employment growth and stock market volatility. This indicates that the availability of an EPU index is of great interest for policy makers, investors and many others. An accurate index creation, however, requires specific knowledge of the country of interest, since not only the country s range of newspapers is important, 1 For the United States, Baker et al. (216) also extend their index to different perspectives of economic policy uncertainty. These perspectives comprise, for instance, taxes, government spending, monetary policy, health-care etc. For each perspective, the authors identify specific keywords for which they search again in newspapers. Note that, however, this is only available for the United States. 2

3 but especially also the country s language proficiency. While the latter requirement might be rather easy to meet for the Americas, it is much more difficult for Europe due to its great variety of different languages. The aim of this paper is to construct an index for economic policy uncertainty for Eastern Europe by drawing on a different data source, which does not require a language proficiency at all. More specifically, we use internet search volume, provided by Google Trends, to construct an index of economic policy uncertainty for Eastern Europe. Academic studies increasingly use Google Trends to explain asset returns (e.g., Da et al. (211a)), liquidity issues (e.g., Bank et al. (211)), volatility (Vlastakis & Markellos (212)) or firm revenue and earnings (e.g., Da et al. (211b)). The basic idea is always the same: Google Trends is a proxy for active attention and by the increase of search volume, there is a rise in people s attention on the issue as well. We make use of two features of Google Trends that allow us to extract the search volume for a specific country without knowing the country s language. 2 These features are by now not often used in academics but they are an elegant way to cope with issues such as lacking language proficiency or, more generally, a missing knowledge of the country itself (see Section 3 for a full description of index construction and the features used). We create our Google economic policy uncertainty (GEPU) index for 1 countries in Eastern Europe and show empirically its accuracy by estimating a vector autoregression including the country-specific GEPU indices, stock market volatility and government bond yields. Moreover and similar to Baker et al. s (216) EPU index, we can show that the index captures relevant national and international events. Especially for emerging markets, such as markets in Eastern Europe, an index that captures economic policy uncertainty is helpful for all stakeholders outside the respective country. Such stakeholders might often lack in understanding the country s language and conditions but the availability of our index gives them an idea of its economic setting. The remainder of this paper is organized as follows: The next section overviews existing literature on political and economic policy uncertainty. Section 3 outlines 2 Note that even though, we checked that categorization and assignment of search topics works properly with different languages. 3

4 the GEPU index construction based on search volume and a validation of the index. Section 4 includes the empirical analysis for Eastern Europe and Section 5 concludes. 2. Existing literature In existing literature, the impact of uncertainty is studied both theoretically and empirically. Theoretical work starts with Bernanke (1983) by analyzing the effects of uncertainty on monetary policy and investment decisions. The basic idea is as follows: uncertainty in policy leads to uncertainty in other areas such as investment rates, spending, higher costs of finance etc. More recently, the paper by Pastor & Veronesi (212) proposes a general equilibrium model that features policy uncertainty and allows studying how policy shifts impact changes in stock prices. The authors conclusion based on their model is that policy changes lead to a decrease in stock prices (on average) and that policy changes lead to an increase in volatilities. In the next section, we will test whether this prediction also hold for our measure of policy uncertainty, namely, the GEPU index. Empirical work analyzes, for instance, effects of uncertainty on investment expenditures or foreign direct investments during election years (e.g., Julio & Yook (212) and Julio & Yook (216)). This paper, however, is very closely related to the study of Baker et al. (216) and their empirical analysis. Their EPU index explains variations in investment rates, employment growth and stock market volatility. Since data is very limited for Eastern Europe, we focus on stock market volatility and government bond yields. To the best of our knowledge, only the paper by Dzielinski (212) has combined Google search volume with economic policy uncertainty. The author, however, uses only one search term (i.e. economy ) in the entire analysis. Moreover, Carriere-Swallow & Cespedes (213) analyze the impact of uncertainty shocks in a worldwide sample differentiating between developed and emerging countries. In their empirical analysis, the authors show that uncertainty shocks have a stronger impact and are more long lasting for emerging countries than for developed countries. To sum up, existing literature focuses very much on developed countries (mostly on the US) in which one would expect economic policy uncertainty to be rather low. Google Trends has only been applied once in this context with only one single search term. This paper contributes in different ways to existing literature: First, we propose an index that 4

5 can be constructed for different countries without knowing its language. Second, we apply our GEPU index to emerging countries in Eastern Europe for which economic policy uncertainty is more severe than for developed countries. Finally, we propose an advanced application of Google Trends by using categories and search topics instead of single search terms. For this reason, the next section gives an overview on the construction of the GEPU index. 3. GEPU Index based on Search Volume 3.1. Index construction As already outlined in the introduction, the existing EPU index by Baker et al. (216) relies on the number of news articles. Other indices, such as the Political Instability Index by the Economic Intelligence Unit 3 or the International Country Risk Guide by Political Risk Services 4, are based on expert opinions or on changes in investments, exports or foreign direct investments. We, on the other hand, create our GEPU index by using only Google search volume. Existing studies that employ Google search volume apply throughout a specific search term (e.g., stock ticker, company names or, more generally, words related to finance). Since we are interested in search volume of different countries with different languages, we make use of Google s automatic categorization and its new feature called search topics. First, Google divides all search queries into specific categories. For instance, if someone searches for Apple and clicks on Apple Store, this search query is attached to the category Shopping. However, if someone searches again for Apple but clicks on recipes with Apples, this search query will be added to the category Food & Drink. There also exist subcategories that are more precise than the main category - e.g. the category Food & Drink has a subcategory called Cooking & Recipes. 5 Second, Google provides a new feature called search topics. With this feature, Google automatically assigns search terms to specific search topics. Let s use again the search query Apple. If someone clicks on the Apple store, this specific search query is assigned to the search topic Apple - Technology 3 See for further information on the Economic Intelligence Unit. 4 See for more information on Political Risk Services. 5 See for a full list of categories as well as a detailed description of Google Trends. 5

6 Panel A. List of relevant categories Financial markets (subcategory of news) Economy News (subcategory of business news) Politics (subcategory of news) Panel B. List of relevant search topics Government spending Dissolution of parliament Coup d état Insolvency Elections & Campaigns (subcategory of politics) World news (subcategory of news) Coalition government Government debt Government budget balance European debt crisis Table 1. List of relevant categories and search topic for the GEPU index construction. company. However, if someone clicks on recipes instead of the Apple store, the search query is matched with the search topic Apple - Fruit. For the paper s purpose, there are at least two important advantages for the use of categories and search topics instead of simple search terms: The first advantage is that the category search volume and the topic search volume include more than one single search term which is potentially positive concerning data availability. The second benefit is that no language proficiency is required to select relevant categories or search topics since Google provides them in English language and they are the same for all countries. While the problem of missing language proficiency might not be a problem in the US, it might be quite relevant in Eastern Europe due to its various different languages. So we make use of both features provided by Google for the construction of our GEPU index. Google not only allows downloading search volume for individual search terms but also for categories and search topics from the Google Trends website. Data availability starts in 4 (up to today) and can be obtained as worldwide search volume as well as for a specific country. We reviewed all existing search categories (including subcategories) on their economic relevance. This means that changes in search volume on these categories imply an increase or decrease in economic policy uncertainty. Table 1 (Panel A) shows the list of relevant categories and subcategories which are used in the following. Moreover, we similarly examined available search topics related to economic policy uncertainty (see Table 1, Panel B). For all countries of interest (see Section 4 for a full list), we individually download time series of search volume for the relevant categories and search topics. Note that Google automatically normalizes and scales all series of search volume. This 6

7 is important to be aware of since this implies that maximum search volume is normalized to (no or low search volume is ) and that it is relative to the entire search volume. Google Trends does not provide an absolute search volume but the interpretation of the downloadable search volume series can be interpreted as a kind of interest in this search term/category/topic relative to all search queries. In order to construct our GEPU index for country A, we simply add up all search volume series for categories and search topics of the country A Index validation For a first validation of our GEPU index based on search volume, we create the index for Germany and compare it with the EPU index by Baker et al. (216). For this purpose, we download all categories and search topics presented in Table 1 for Germany. Figure 1 compares both indices with each other over time. At first glance, general events are captured quite well by both indices (see, e.g., financial crisis and sovereign debt crisis) and also rather specific political events, such as the motions of no confidence in summer 5 by the former Chancellor Gerhard Schröder, are gathered in both cases. However, there is a main difference between two indices: The EPU index by Baker et al. (216) includes all news articles of the respective country (here Germany) no matter whether this uncertainty has its source in the country or not. Whereas our GEPU index shows whether people search more for relevant categories or not which can be interpreted as a kind of expected uncertainty by the German population. [Figure 1 about here.] Finally, since the construction of the EPU index by Baker et al. (216) is solely based on keywords closely related to government debt issues (see Section 1 or the Appendix of Baker et al. (216)), we also construct a reduced GEPU index that only consists of debt related categories and search topics which fits even more the movements of the EPU index by Baker et al. (216) (see Figure 2). 6 [Figure 2 about here.] 6 Note that the different scaling is due to the summation of the single search volume indices to the GEPU index. 7

8 For the analysis in the next section, we use the full sample GEPU index including all 13 categories and search topics since it describes a broader view on economic policy uncertainty than only focusing on government debt issues. Up to now, we conclude that the GEPU index based on search volume is able to capture important events but focuses more on country-specific issues and a wider definition of economic policy uncertainty than its newspaper s alternative. 4. Empirical test 4.1. Data Since the aim of this paper is to show the creation of an EPU index for countries with a foreign language, we focus on Eastern Europe, more specifically, on the following countries: 7 Albania (AL), Belarus (BY), Bosnia (BA), Bulgaria (BG), Croatia (HR), the Czech Republic (CZ), Estonia (EE), Georgia (GE), Hungary (HU), Latvia (LV), Lithuania (LT), Macedonia (MK), Moldova (MD), Poland (PL), Romania (RO), Serbia (RS), Slovakia (SK), Slovenia (SI), Ukraine (UA). Figure 3 shows the GEPU index for these countries over time. [Figure 3 about here.] In the empirical analysis, we test whether our EPU index (based on search volume) tracks some fraction of variation in country-specific variables. Data availability as well as frequency issues force us to concentrate on major national stock market indices as well as government bond yields and industrial production. In particular, we use the respective MSCI stock market indices in local currencies, the long-term government bond yields (to maturity) and (changes) in industrial production. We obtain data from Thomson Reuters DataStream, ECB Statistical Data Warehouse and the OECD Estimation model To further verify the relevance of our GEPU indices we run vector autoregressions both for individual countries and for the whole pooled sample of our dataset. We 7 Of course, the index can also be constructed for the US or Germany (as shown above), however, we restrict our analysis to the Eastern Europe due the inflationary emergence of indices for western countries. 8

9 choose the VAR approach particularly due to endogeneity effects between our variables and potential spurious relationships arising from autocorrelated regressors. Consider a VAR(p) model of the form X t = A 1 X t A p X t p + ε t, (1) where X t is a stacked vector of our variables in time t, X t = [GEP U t, y t, r 2 t, d ip t ]. GEPU denotes our country specific economic policy index, y t is the yield to maturity, r 2 t the squared stock index returns and d ip t the log difference of industrial production. A 1 to A p denote the coefficient matrices including the constant. ε t denotes the error term. Based on lag length criteria following Schwarz and Hannan-Quinn information we choose a lag length of p = 2. Critically, we check the stability of the VAR by making sure all roots lie outside the unit circle Results Figure 4 shows impulse response functions for the countries Bulgaria, Croatia, Czech Republic, Estonia, Hungary, Lithuania, Poland, Romania and Slovenia. Functions are based on a VAR(2) and x t = [GEP U t, y t, rt 2 ]. We use the Cholesky factorization for the covariance matrix. 8 The graphs on the left hand side depict the response of bond yields to a one standard deviation innovation in the GEPU index. Dotted lines emphasize 2 standard deviation errors. For all countries but Romania and Bulgaria we detect positive responses of bond yields. Intuitively, this should be expected since the higher economic policy uncertainty, the higher potential risk for default of the bonds. On the right hand side we plot graphs summarizing the response of squared returns to shocks to the GEPU index. For 6 out of 9 countries squared returns as a proxy for variance rises after a shock to the GEPU index. Again, this is as expected since economic policy uncertainty may correspond to higher variance in stock returns. This is also in line with findings in Pastor & Veronesi (212) who detect a positive relationship between policy uncertainty and valoatility. Both effects from bond yields and squared returns highlight the relevance of our GEPU index and potential usefulness for policy makers. 8 Using different Cholesky orderings does not change our results nor does the use of generalized impulse response functions. 9

10 [Figure 4 about here.] [Figure 5 about here.] Figure 5 shows impulse response functions for pooled data comprising the countries from the previous exercise. Is there a common pattern across countries? The left graph highlights the (aggregated) positive response to a one standard deviation shock in the GEPU. Results are even more pronounced than in the single country case. Standard error bounds (± 2 std. err.) are above zero suggesting a significant positive variation in bond yields after a shock in GEPU. Squared returns, on the right hand side, immediately rise following a GEPU shock but turn negative after two periods and then approach zero. In a nutshell, the GEPU index clearly tracks some variation in the analyzed variables. [Figure 6 about here.] We additionally incorporate changes in industrial production in our model for the countries Czech Republic, Estonia, Hungary, Lithuania, Poland and Slovenia (due to data availability). Figure 6 shows results for impulse response functions. The response of bond yields stays virtually the same as in the previous test. Squared returns (middle graph) respond more mildly, only slightly turning negatively after 2 periods. Interestingly though, on the very right graph, changes in industrial production are positive in the first period. This might suggest that change in industrial production are not immediately (negatively) influenced by economic policy uncertainty contrary to bond yields and stock return variance. [Figure 7 about here.] 5. Conclusion In this paper, we propose an index of economic policy uncertainty based on Google search volume for 18 Eastern European countries. The index is country-specific and based on a wide range of search categories and search topics. We validate the construction of the GEPU index by comparing it with the EPU index for Germany. We observe co-movements during specific dates of increased policy uncertainty (e.g., sovereign debt crisis). In addition to this first validation, we show in a vector 1

11 autoregressive (VAR) model, including government bond yields and squared returns as a proxy for volatility, that shocks of the GEPU index have an impact on both variables with the expected sign. We conclude that the GEPU index is able to capture economic policy uncertainty in Eastern Europe. Due to its availability and flexibility, we expect that the GEPU index will be of great interest for policy makers and investors - inside as well as outside the respective country. 11

12 References Baker, S. R., Bloom, N. & Davis, S. J. (216), Measuring economic policy uncertainty, The Quarterly Journal of Economics. Bank, M., Larch, M. & Peter, G. (211), Google search volume and its influence on liquidity and returns of german stocks, Financial Markets and Portfolio Management 25(3), Bernanke, B. S. (1983), Irreversibility, uncertainty, and cyclical investment, The Quarterly Journal of Economics 98(1), Bloom, N. (9), The impact of uncertainty shocks, Econometrica 77(3), Carriere-Swallow, Y. & Cespedes, L. F. (213), The impact of uncertainty shocks in emerging economies, Journal of International Economics 9(2), Da, Z., Engelberg, J. & Gao, P. (211a), In search of attention, Journal of Finance 66(5), Da, Z., Engelberg, J. & Gao, P. (211b), In search of fundamentals, AFA 212 Chicago Meetings Paper.. Dzielinski, M. (212), Measuring economic uncertainty and its impact on the stock market, Finance Research Letters 9(3), Julio, B. & Yook, Y. (212), Political uncertainty and corporate investment cycles, The Journal of Finance 67(1), Julio, B. & Yook, Y. (216), Policy uncertainty, irreversibility, and cross-border flows of capital, Journal of International Economics 13, Pastor, L. & Veronesi, P. (212), Uncertainty about government policy and stock prices, The Journal of Finance 67(4), Vlastakis, N. & Markellos, R. N. (212), Information demand and stock market volatility, Journal of Banking & Finance 36(6),

13 List of Figures 1 Comparison of EPU index by Baker et al. (216) and the GEPU index for Germany Comparison of EPU index by Baker et al. (216) and the reduced GEPU index for Germany GEPU index over time Impulse-response functions for individual countries Impulse-response functions - pooled data Impulse-response functions - pooled including industrial production. 2 13

14 GEPU index for Germany EPU index for Germany by Baker et al. (216) Fig. 1. Comparison of EPU index by Baker et al. (216) and the GEPU index for Germany. 14

15 reduced GEPU index for Germany EPU index for Germany by Baker et al. (216) Fig. 2. Comparison of EPU index by Baker et al. (216) and the reduced GEPU index for Germany. 15

16 Albania Bosnia Bulgaria Belarus Estonia Georgia Croatia Hungary Lithuania Latvia Moldova Macedonia 5 Poland Romania Serbia Slovenia Slovakia Ukraine Fig. 3. GEPU index over time. 16

17 Response of Y_BG to GEPU_BG Response of R_BG^2 to GEPU_BG Response of Y_CZ to GEPU_CZ Response of R_CZ^2 to GEPU_CZ Response of Y_EE to GEPU_EE Response of R_EE^2 to GEPU_EE Response of Y_HR to GEPU_HR Response of R_HR^2 to GEPU_HR Response of Y_HU to GEPU_HU Response of R_HU^2 to GEPU_HU Fig. 4. Impulse-response functions for individual countries These graphs show responses to Cholesky one standard deviation innovations to the GEPU index following a VAR(2) from Equation (1). 1 month 17horizon. Vertical axes are in units of the response variable.

18 Response of Y_LT to GEPU_LT Response of R_LT^2 to GEPU_LT Response of Y_PL to GEPU_PL Response of R_PL^2 to GEPU_PL Response of Y_RO to GEPU_RO Response of R_RO^2 to GEPU_RO Response of Y_SI to GEPU_SI Response of R_SI^2 to GEPU_SI Fig. 4 (cont.). Impulse-response functions for individual countries These graphs show responses to Cholesky one standard deviation innovations to the GEPU index following a VAR(2) from Equation (1). 1 month horizon. Vertical axes are in units of the response variable. 18

19 Response of Y_ to GEPU_ Response of r^2 to GEPU_ Fig. 5. Impulse-response functions - pooled data These graphs show responses to Cholesky one standard deviation innovations to the GEPU index following a VAR(2) from Equation (1). 1 month horizon. Vertical axes are in units of the response variable. 19

20 .12 Response of Y_ to GEPU_.2 Response ofr^2 to GEPU_.3 Response of IP_DIFF_ to GEPU_ Fig. 6. Impulse-response functions - pooled including industrial production These graphs show responses to Cholesky one standard deviation innovations to the GEPU index following a VAR(2) from Equation (1). 1 month horizon. Vertical axes are in units of the response variable. 2