International return predictability : a trade-based explanation?

Master Thesis

International Return Predictability: a Trade-Based Explanation?

Frank van Hoenselaar

6082882

July 6, 2015

Statement of Originality

This document is written by Student Frank van Hoenselaar, who declares to take full responsibility for the contents of this document. I declare that the text and the work

presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of

Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Contents

1 Introduction 4

2 Literature Review 6

3 Data and Methodology 9

3.1 Data . . . 9 3.2 Empirical Framework . . . 12 4 Results 15 4.1 Country-level Analysis . . . 15 4.2 Industry-Level Analysis . . . 18 4.3 Alternative Explanations . . . 27 4.4 Robustness Checks . . . 32 5 Conclusion 37

1

Introduction

The Efficient Market Hypothesis (EMH) states that in an efficient market, prices are immediately updated when new relevant information is revealed. However, in the re-cent empirical literature it has been found that several factors can impair the process of updating security prices. For instance, investors face constraints in their capacity to process information. The process of updating prices therefore strongly depends on investor attention and their ability to quickly draw conclusions about the implications of new information for a certain security.

Information in foreign markets often contains relevant information for domestic se-curities. Nguyen (2012) argues that for multinational firms, investors strongly focus on the domestic market and therefore information from foreign operations tends to slowly diffuse into the stock prices of multinational firms. In fact, this slow information dif-fusion creates a certain degree of return predictability, where past returns from foreign markets can serve as a predictor for this month’s returns. Such geographic momentum has not only been found at the firm level, but persists at the country level as well. Rapach et al. (2013) find that last month’s U.S. stock returns can predict the returns of many OECD country’s indexes. But their results show that not only the U.S. stock market has predictive power, in fact half of the countries in their sample seem to have significant predictive power for other country’s stock-indexes. However, Rapach et al. (2013) do not provide an explanation why these predictive relations exist.

Rizova (2010) provides a framework that aims to explain why stock returns in one country have predictive power for next month’s stock return in another country, namely: trade-relationships. His paper uses trade-weighted portfolios of trading part-ner’s country indexes to predict the return of a certain country’s index. These portfolios do have significant predictive power and can yield substantial abnormal returns when exploited. However, there is no convincing proof that this predictive power actually follows from trade-based relationships. Moreover, it does not explain many of the more ”random” patterns found bt Rapach et al. (2013). One might as well argue that the

observed return predictability is in fact due to regional shocks that are partially cap-tured by the portfolios constructed by Rizova (2010). Alternatively, it might be the case that some stock markets lead other stock markets, a phenomenon analogous to the findings of Hong et al. (2007) that certain industries lead other industries.

This thesis aims to explain whether international return predictability is a phe-nomenon based on trade relationship between countries or whether there are alterna-tive explanations. It will do so by comparing return predictability in 33 OECD and BRIICS economies in the period 1990-2013. The study will not only look at return predictability at a country level, but also at an industry level in these 33 countries. Studying return predictability at the industry has an additional advantage because in-dustries vary widely in their degree of trade-openness. If return predictability is really driven by trade, we would expect to see no effect in relatively closed industries.

Following the work by Rizova (2010), trade-weighted portfolios of trading part-ner’s country indexes are created, both at a country level and at an industry level. A long-short short strategy based on the returns of the lagged trading-partner portfolio returns yield alphas on a country level of 14.43% up to 19.84% on an industry level. Using Fama-Macbeth forecasting regressions it is shown that on a country level there is no evidence that more open countries have more return predictability. However, it is found that when a country has many trading partners and hence information is complex to process, return predictability is significantly lower.

At an industry level, there is again significant return predictability although there is no evidence that more openness or more complex trade-relations affect return pre-dictability significantly. However, when the eight industries are appended and sorted into quintiles according to openness or complexity, we find increased return predictabil-ity in more open and less complex quintiles. However, the latter differences are not significant.

At the end of this thesis we experiment with alternative trading partner portfolios. Some countries are import dependent, some are export dependent. Therefore we test whether a portfolio weighted by either imports or exports based on the trade-balance

can improve upon the fit of the model. It is found that this portfolio does not have a better fit, although a trading-partner portfolio that only uses exports has a somewhat better fit. Additionally, alternative portfolio such as the World Index, the U.S. stock market or other portfolios are also unable to improve the fit of the model.

Because of the limited evidence that return predictability is really a trade based phenomenon, other explanations are tested. It is found that return predictability is not driven by regional shocks. Since even for industries that have almost all trading partners outside their own region we observe return predictability based on the trading-partner portfolio. Furthermore it is found that return predictability is not driven by illiquidity in some of the countries studied, because we find that the effect is heavily significant in both countries with liquid and countries with illiquid stock markets. Re-turn predictability is also not a phenomenon of the past when markets were still more inefficient, because it is observed throughout the entire sample period. However, the effect seems to decline throughout time although not significantly.

The main finding of this thesis is that the predictive power of the trading-partner portfolio is still very strong. However, the found relationship between trade-openness and return predictability is somewhat weak. On the other hand, using alternative tests this thesis is not able to falsify the statement that international return predictability follows from trade-based relations between countries and industries.

2

Literature Review

This thesis can be placed in a strand of literature that studies the diffusion of relevant information into asset prices. The Efficient Market Hypothesis (EMH) is often cited as an important reference point. The models underpinning the EMH rely on the assump-tion of complete informaassump-tion sets and no market fricassump-tions (Merton, 1987). However, in recent decades more sophisticated theoretical models have been developed, that focus on incomplete information sets, investors’ limited capacity to process information and investors’ limited capacity to do arbitrage.

From the late 1990’s many empirical papers have confirmed the existence of such frictions in the financial market. An important branch of the literature focuses on investor (in)attention to certain information. DellaVigna and Pollet (2009) find for in-stance that when new information about a company’s performance enters the market on Friday (when attention is low), investors substantially underreact to this informa-tion. Among others Barber and Odean (2008) show that investors tend to limit their attention to stocks that are in the news a lot or that have very high trading volumes. Within Behavioral Finance, a large strand of literature has been focusing on in-vestor overconfidence. Daniel et al. (1998) argue for instance that inin-vestors have the tendency to underreact to public signals and overreact to private signals. The under-lying mechanism is overconfidence about own prior beliefs. A large body of literature seems to confirm that investors usually underreact to public information such as earn-ings revisions (Zhang, 2006).

More related to this thesis is the literature that describes the process of information diffusion into individual stock prices. Lo and MacKinlay (1990) show that the returns of small companies follow the returns of large companies. Their finding implies that information that affects a set of companies will enter the stock prices of small compa-nies (less attention/more complicated to monitor) with a lag. This effect is referred to as the lead-lag effect and can be exploited to yield substantial abnormal returns. More recently, Hong et al. (2007) show that lagged industry portfolio’s have predictive power for other industries. The reasoning is that because of the limited capacity to process information, investors do not immediately extract information from the prices of industries they are not active in.

The literature on the lead-lag effect is plentiful. Cohen and Frazzini (2008) show that the stock-return of customer firms have predictive power for the stock-return of their suppliers. A strategy that goes long in firms who’s customer did well and goes short in firms who’s customer did poorly yields monthly Alpha’s of 150 basis-points. A similar analysis on economically linked companies has been performed by Menzly and Ozbas (2010). Cohen and Lou (2012) provide an analysis and method that is

analogous to this thesis. They argue that firms that are active in multiple industries (conglomerates) are relatively complicated to judge when new information becomes available. They show that the lagged returns of portfolio’s of standalone firms that mimic the activities of a conglomerate firm, have substantial predictive power for the returns of the conglomerate firm. Their findings additionally show that when infor-mation is more complex (i.e. a conglomerate firm that is active in many industries), return predictability decreases. This finding is in line with models that predict that investors have limited capacity to process information.

On a country level Rapach et al. (2013) find that U.S. stock markets predict next month’s returns of many developed market indexes. Moreover, they find that not only the U.S. stock market has predictive power, all countries in their sample are able to predict the returns in at least 4 other countries. It therefore seems that return pre-dictability is a reciprocal phenomenon where country A can predict country B, but where country B can at the same time predict country A. Although their method and results are robust, the paper does not provide an analytic framework that explains why certain countries predict returns in other countries and others not. Rizova (2010) on the other hand provides a comprehensive framework that provides some insights how countries predict each other. Using a Lucas-tree, two countries two goods free trade model he argues that trade links are the main reason why also stock markets are interlinked. His model derives from the prediction that information in one country is will result in an immediate and full response of the stock market in the other country. Not surprisingly, the empirical section shows that news in trading partners’ markets is only incorporated in domestic prices with a lag.

Although the previous analysis is comprehensive, it does not guarantee that return predictability between countries is a phenomenon based on trade-links. For instance, the findings by Rapach et al. (2013) can very often not be explained by trade rela-tionships and therefore the analysis of Rizova (2010) seems not complete. This thesis will study on an industry level, rather than only at a country level whether return predictability is a trade-based phenomenon. Studying return predictability on the

industry level has the advantage over country level analysis because industries vary greatly in their degree of trade-dependence.

3

Data and Methodology

3.1

Data

The sample consists of all OECD countries including the so called BRIICS countries (Brazil, Russia, India, Indonesia, China and South Africa) for the period 1990-2013. This paper will focus for a small part on return predictability on the country level, but will mainly look at the industry level within these countries. In order to build a proper data set, many different data sources have to be employed. A main challenge in combining the different data sets is the different specification of industries between the OECD database (OECD classification system) and the Compustat database (SIC-codes). However, the SIC (Standard Industry Classification) system has such a level of detail that one can quiet easily match the OECD industry classifications to a certain range of SIC-codes. Based on the data availability in the Compustat Global database, eight industries are defined that can be matched with the OECD industry system:

Table 1: Conversion Table OECD to SIC Industries

Industry Name Abbreviation OECD-code SIC-Code

Agriculture Hunting and Forestry AHF C01T05 0-1000

Business Services BS C50T74 7000-8400

Chemicals CHM C23T26 2800-2900

Food and Tobacco FOT C15T16 2000-2200

Finance Real Estate Insurance FRI C65T67 6000-6700

Mining and Quarrying MQ C10T14 1000-1500

Transport Equipment TRA C34T35 3700-3800

Electrical Equipment EEQ C30T33 3600-3700

Following Rizova (2010) the OECD bilateral trade database will be employed to define trade-links between economies. However, the approach in this thesis will be to study return predictability at an industry, rather than only at a country level. The

OECD bilateral trade database provides detailed data about bilateral trade on the industry level. For every industry in every OECD country it recorded what part of total revenues was exported or imported to what country. For defining the total trade in an industry, the sum of imports and exports is used. To define shares of industry trade to a particular trading partner we use the following definition:

sijkt=

Importijkt+ Exportijkt

Importikt+ Exportikt

(1) Where subscript (i) defines a certain country, subscript (j) defines the industry, sub-script (k) defines the trading partner and subsub-script (t) defines in a certain year. When we study return predictability at the country level, we employ also the OECD bilateral trade database, but look at the total trade (sum of all industries) of that country. Data on imports and exports is only available at a yearly interval, the shares are determined at the start of every year and remain constant for 12 consecutive months. Since no trade-shares are available in the Business Services (BS) and Financial Services (FRI) are available, the shares of total trade of a particular country are used as a proxy. In order to create trading-partner portfolios broad country indexes are matched with the trade-shares. Country indexes are obtained via Compustat global indexes. The sample consists of over 70 potential trading partners and for most of them MSCI or HSBCI indexes are available. For a few countries alternative indexes are used, and if not available (for instance former USSR countries) broad regional indexes provided by MSCI are used. Following Rizova (2010) and Nguyen (2012) only trade shares larger

than 5% (sijkt> 0.05) are considered (it is assumed that even smaller trading partners

cannot exhibit any significant influence on returns). A trading partner portfolio for an industry is than defined as:

Rijkt =

n

X

i=1

sijkt× Ijt (2)

At the country level we use the same definition, but look at total trade between country (i) and trading partner (k) as recorded by the OECD. The correlation between the

trading-partner portfolios of the industries and the trading-partner portfolios of the countries range between 0.7662 and 0.9803. Furthermore there are an average of 4.41 trading-partners in every portfolio with a minimum of 1 and a maximum of 9.

The dependent variable at the country level (return) is then generated as the return of the broad country index (MSCI or HSBCI). Since no useful industry indexes are available for all the countries in the sample of this paper, an alternative strategy is employed. Compustat global securities database has data on over 35000 companies in over 40 countries. Therefore one can create their own indexes by calculating average returns (weighted by market capitalization) for every industry in every country using the following procedure:

1. Calculate the monthly returns for every company in the sample. 2. Generate industries based on the SIC system.

3. Calculate market capitalization for every company and also per industry. 4. Calculate returns weighted by market capitalization within industry 5. Collapse the data set on time and industry.

The industry returns are subsequently matched with the data set containing the trading-partner portfolios over every industry. After the merge the sample consists of 33 coun-tries each divided into eight induscoun-tries.

Trade openness (also referred to as trade dependence) is another important vari-able for this research. At the country level we follow the procedure by Rizova (2010) and define openness as the sum of imports and exports divided by GDP. This data is obtained from the Worldbank. Unfortunately, for the study at the industry level there is no data available on the total size of each OECD industry which implies that another definition of openness is used. In the TiVA (trade in value added) database of the OECD one can find data on what part of value added in every industry is earned where. Trade openness at the industry level is therefore defined as the part of value added of a certain industry that is earned outside the country. Both measures for trade-openness are only available at a yearly interval. Which means that the variable

openness is the same for every 12 monthly observations in one year.

Because return predictability is studied on the country and industry level, many of the control variables that would be included on a firm level (turnover, marketcap, market-to-book) are simply not available. This paper will therefore follow the proce-dure by Rizova (2010) of adding many lagged variables. The average return from two months until a year ago from both the trading-partner portfolio and also the industry return are calculated and included as control variables.

3.2

Empirical Framework

In this section we will outline an empirical framework that enables us to test several hypotheses leading to an answer of the research question: Does international return predictability arise from trade-links between countries? As outlined before, this re-search question will be tested at both the country and industry level. This section presents the regression models in the industry-level form, which in fact means that there is an additional subscript (j) in every specification. The research question will be tested using three main hypotheses:

Hypothesis 1 Gradual information diffusion induces return predictability in home stock markets using lagged exposure to foreign markets

Following the literature on momentum investing by Jegadeesh and Titman (2001), industries will be sorted by the performance of their corresponding lagged

trading-partner portfolio (equation 2). Using this procedure one can determine what the

monthly alpha’s of a long/short strategy might yield. To test the predictive power more formally a Fama-MacBeth forecast regression of the following form will be esti-mated:

T able 2: Summary Statistics and Correlations This table presen ts the summary statistics and the correlation co efficien ts of the v aria bles of in terest. R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th return and the a v erage return from 12 to 2 mon ths ago of that same industr y. P O R Tt is the return of a p ortfolio that consists of coun try index es w eigh ted b y the v alue of trade of an in dustry’s trading partners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return an d the a v era ge retu rn from 12 to 2 mon th s ago of the same p ortfolio. O peness is a v ariable that indicates the share of pro duction that is earned outside th e home coun try . The Herfindahl Index (H H I ) is a mea su re of complexit y of trade relations (lo w er is more complex). P anel A: S ummary Statistics M ean S D p 1 p 25 p 50 p 75 p 99 N R ett .00275 .10040 -.28838 -.03971 .00559 .05101 .25096 65950 R ett− 1 .00275 .10056 -.28927 -.03977 .00560 .05112 .25117 65695 R ett− 2: t− 12 .00279 .03452 -.10170 -.01337 .00592 .02221 .08166 62890 P O R Tt− 1 .00743 .05386 -.14929 -.02029 .01219 .03942 .13210 59092 P O R Tt .00740 .05366 -.14929 -.02016 .01208 .03928 .13180 59277 P O R Tt− 2: t− 12 .00696 .01918 -.05001 -.00286 .00997 .01818 .04768 57121 O penness .39368 .21180 .05568 .21938 .35664 .55658 .85398 57750 H H I .10596 .12033 0 .05231 .07278 .10558 .64192 65950 P anel B: Correlation Co efficien ts R ett R ett− 1 R ett− 2: t− 12 P O R Tt− 1 P O R Tt P O R Tt− 2:12 O penness H H I R ett 1 R ett− 1 0.0774 ∗∗ 1 R ett− 2: t− 12 0.0492 ∗∗ 0.0663 ∗∗ 1 P O R Tt− 1 0.102 ∗∗ 0.374 ∗∗ 0.0133 ∗∗ 1 P O R Tt 0.375 ∗∗ 0.0569 ∗∗ 0.00696 0.109 ∗∗ 1 P O R Tt− 2:12 0.000465 0.0266 ∗∗ 0.435 ∗∗ 0.0590 ∗∗ 0.0431 ∗∗ 1 O penness -0.00972 ∗ -0.0108 ∗ -0.0255 ∗∗ -0.0126 ∗∗ -0.0116 ∗ -0.0255 ∗∗ 1 H H I -0.00270 -0.00281 -0.0119 ∗∗ 0.00231 0.00202 0.000142 0 .0883 ∗∗ 1 + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

Where the coefficient on β1 measures the predictive power of the lagged trading-partner

portfolio for the corresponding indexes of a certain industry. If the hypothesis that

there indeed is predictive power is correct, we would expect a positive coefficient on β1.

The vector X consists of standard control variables time and other lagged regressors. The coefficient β2 captures standard momentum effects.

Hypothesis 2 International return predictability will be highest in trade intensive in-dustries/countries.

Industries will be sorted in different quintiles based on their trade- openness (Note that trade-openness is differently defined between the country level and industry level anal-ysis, see data section for explanation.). Following an approach similar to DellaVigna and Pollet (2009), a new variable called TopQuintile (TopQ) will be created that is interacted with the lagged return of the trading-partner portfolio. More formally the following regression model will be estimated:

RETijt = α+β1×T opQijt+β2×Rkt−1+β3×T opQijtRkt−1+β4×RETijt−1+λ×Xijt+εijt

(4)

In order to confirm hypothesis 2, we would have a positive and significant β3.

Alter-natively, we will estimate equation 3 for the different quintiles of openness.

Hypothesis 3 More complicated trade relationships slow down the information diffu-sion and decrease international return predictability.

Cohen and Lou (2012) and Nguyen (2012) show on a firm level that return predictability decreases with the complexity of the information. More particularly, Nguyen (2012) shows that return predictability is lower for multinationals that are operating in many different countries. The complexity of information is proxied by the Herfindahl Index:

Herf indahlIndex(HHI) =

n

X

j=1

s2_ijkt (5)

In other words, the Herfindahl Index in this thesis is the sum of squared trade shares of industry (j) in country (i) with country (k) on time (t). To formally test hypothesis 3, the following regression model is employed:

RETijt = α+β1×LowHHIijt+β2×Rkt−1+β3×lowHHIijtRkt−1+β4×RETijt−1+λ×Xijt+εijt

(6) The variable lowHHI is an indicator variable that indicates when an industry belongs to the top 20% of complicated trade-relationships. Hypothesis 3 would be confirmed

when β3 is negative and significant.

4

Results

This section will outline the main results of this thesis. It will do so by first analyzing the results found on a country level and subsequently analyze the results on an industry level. It will end by doing multiple robustness tests and alternative specifications.

4.1

Country-level Analysis

Table 3 presents the outcome of a momentum strategy involving a long-short strategy based on the past performance of the trading-partner portfolio. Every month the portfolio is rebalanced, by shorting and longing the indexes which trading partner portfolios performed respectively the worst and best. The CAPM model is estimated for different quintiles of last month’s trading-partner portfolio performance. The α in column 1 implies that investing in country indexes for which the last month’s trading-partner portfolio performed the worst (bottom 20%), will yield a monthly loss relative to the World Index of 0.51%. On the other hand, investing in the top quintile will result in a monthly excess return of 0.78%. Therefore a long-short strategy yields on

average excess returns of 1.13% monthly (14.43% yearly). Due to the high transaction costs of such a strategy of monthly rebalancing, the effective yield will be substantially lower.

Table 3: Portfolio Sorting Country Level

This table presents the output of an estimation of the CAPM model. The MSCI World index has been used as the market portfolio and the 3-month U.S. treasury rate (transformed into monthly rates) serves as the risk-free rate. The observations are sorted based on the previous month’s performance of the trading-partner portfolio. The last column represents a portfolio that goes long in the top quintile of last month’s performance of the trading partner portfolio and goes short in the bottom quintile. The α represents the monthly excess return of every portfolio.

1 (Low) 2 3 4 5 (High) High-Low

α -0.0051∗∗ 0.0005 0.0015 0.0063∗∗ 0.0078∗∗ 0.0113∗∗

(-3.02) (0.33) (1.08) (4.94) (5.39) (8.34)

β 0.917∗∗ 0.945∗∗ 0.966∗∗ 0.939∗∗ 0.954∗∗ -0.299∗∗

(31.26) (28.81) (29.36) (25.59) (26.78) (-11.07)

t statistics in parentheses+ _{p < 0.1,}∗ _{p < 0.05,}∗∗ _{p < 0.01}

Table 4 presents the results of three different regression that represent the three hypothesis presented in the methodology section. The variable of interest is the lagged

trading-partner portfolio P ORTt−1. Hypothesis 1 states that this regressor has

pre-dictive power for this month’s return and in column 1 the hypothesis is confirmed. The coefficient seems somewhat lower than the one found by Rizova (2010) where the

coefficient was on average around 0.1. The coefficients on Rett−2:t−12 and Rett−1 are

all positive but insignificant (at 5%), signaling that domestic public information is in-corporated well in prices (i.e. all information on past return is already inin-corporated in prices). Furthermore we observe a strong linearity between the returns of the trading-partner portfolio and the returns of the country index itself, the coefficient is close to 1. This coefficient is much lower in the analysis done by Rizova (2010) where it is on average only 0.5. However, omitting the variable drastically reduces the fit of the model and makes many of the other coefficients insignificant. Moreover, the coefficient is also included in the analysis by Rizova (2010).

Table 4: Country Level Regressions

This table presents the output of a regression of monthly Country Index returns on a set

of regressors for the period 1990-2013. Rett is the monthly return of a certain industry.

Rett−1and Rett−2:t−12are the previous month return and the average return from 12 to

2 months ago of that same industry. P ORTt is the return of a portfolio that consists

of country indexes weighted by the value of trade of a country’s trading partners.

P ORTt−1 and P ORTt−2:t−12 are the previous month’s return and the average return

from 12 to 2 months ago of the same portfolio. Openess is a variable that indicates the share of production that is earned outside the home country. The Herfindahl Index (HHI) is a measure of complexity of trade relations (lower is more complex). lowHHI is an indicator variable equal to one if a country belongs to the top 20% of complex trade relationships. highOpeness is an indicator equal to one if a country belongs to the top 20% of trade-openness.Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust to heteroskedasticity.

(1) (2) (3)

Rett Rett Rett

Rett−2:t−12 0.122+ 0.120+ 0.122+ (1.74) (1.71) (1.74) Rett−1 0.0162 0.0162 0.0177 (0.72) (0.72) (0.78) P ORTt−2:t−12 -0.168∗ -0.166∗ -0.166∗ (-2.26) (-2.24) (-2.23) P ORTt 0.860∗∗ 0.860∗∗ 0.861∗∗ (48.11) (48.12) (48.08) P ORTt−1 0.0518∗ 0.0495∗ 0.0651∗∗ (2.20) (2.22) (2.71) highOpeness -0.00153 (-1.09) highOpeness × P ORTt−1 0.0103 (0.32) lowHHI 0.00107 (0.73) lowHHI × P ORTt−1 -0.0810∗ (-2.31) constant 0.00504+ _0.00533+ _0.00483+ (1.76) (1.84) (1.67) R2 _{0.396 0.396 0.397} N 7812 7812 7812 t statistics in parentheses. + _{p < 0.1,}∗_{p < 0.05,} ∗∗ _{p < 0.01}

In Column 2, the second hypothesis is tested. The results reveal that return pre-dictability seems not more of an issue in more open economies. However, as can be seen in in column 3 there seems to be less return predictability in countries that have more complex trade relationships (low HHI). The latter finding confirms hypothesis 3. Altogether, the country level analysis does not provide strong support that the predictive power of trading-partner portfolios seems really be following from trade. However, complexity in trade based relationships does seem to hamper return pre-dictability. Although, the latter finding might have to do with other characteristics of countries that have complex trade patterns.

4.2

Industry-Level Analysis

Table 5 shows the results of a long-short strategy analogous to the one presented in previous section. Instead of studying this strategy at the country level, the data is now broken down into different industries (a description of the industries can be found back in table 1). Although all industries vary much in their trade dependency, the strategy seems to yield substantial excess returns for all the industries. The least profitable is the food and tobacco industry 0.715% monthly (8.89% yearly) and the most profitable the electrical equipment industry 1.52% (19.84% yearly).

Hypothesis 1 is again formally tested in table 6. The industries are grouped based on trade-dependence (trade-openness). The difference in trade-openness between these groups is substantial. However, there does not seems to be a large difference between

the coefficients of these subgroups. The lagged trading-partner portfolio (P ORTt−1) is

significant for all industries except Mining and Quarrying. The magnitude of the coeffi-cient does vary across industries (high for Electrical Equipment and low for Chemicals), although these differences do not seem related to the trade dependency of the

indus-try. The coefficient on Rett−2:t−12 does appear to be significant in all industries. This

implies that on an industry level, public information is not reflected in prices properly (i.e. one can still predict today’s industry returns based on the average performance of

last year). Additionally in the Mining and Quarrying industry last month’s return do have explanatory power for this month’s return as well. The coefficient on last year’s average performance of the trading partner portfolio is significantly negative. The most plausible explanation for this is a mechanism of mean reversion. Where the effect of the trading-partner portfolio tends to overshoot and is corrected in later periods.

In table 7 hypothesis 2 is tested more formally, now on an industry level. This implies that when data is sorted (according to openness and HHI), this is also done at an industry level. The interaction term on lagged trading-partner portfolio return and the indicator for high openness is only significant for the Mining and Quarrying indus-try. This result is remarkable because this industry is the only industry where there is no general effect of the lagged trading-partner portfolio. Such a result implies that the return predictability within this industry is conditional on being open. In other words, return predictability is in this case only present in countries where Mining and Quarrying is an open industry. Furthermore the results show that in the Agriculture, Hunting and Forestry (AHF) industry, returns are generally lower in the top quintiles of trade-openness. This finding is important for the later part in this thesis, when the data is sorted according openness, it implies that we should use standardized betas when comparing the coefficients.

Hypothesis 3 is tested in table 8. One might notice that the Business Services and Financial Services industry are omitted from the regression. This is due to the fact that, not industry level data on bilateral trade is available for these industries, hence the Herfindahl Index would reflect the fragmentation of the overall trade in the economy. Despite the fact that on a country level we found that more complex trade relationships hamper return predictability, we do not observe this finding on an

indus-try level. Moreover, the overall fit of the model (R2_{) is substantially lower compared}

to the specification in table 6 and 7.

At this point the three hypotheses outlined in the methodology section have been tested properly. However, the results seem to indicate that the return predictability that is observed in these different industries is unrelated to trade. There is an alternative

T able 5: P ortfolio Sorting Industry Lev el This table presen ts the output of an estimatio n of the CAPM mo del. The MSCI W orld index has b een used as the mark et p ortf olio and the 3-mon th U.S. trea su ry rate (transformed in to mon thly rates) serv es as the risk-free rate. The observ a tions are so rted based on the previous mon th’s p erformance of the tradin g-partner p ortfolio. The last colu mn represen ts a p ortfolio that go es long in the top quin tile of last mon th’s p erformance of the tra ding partner p ortfolio and go es shor t in the b ottom quin tile. The α represen ts the mon thly excess return of ev ery p ortfo lio. 1 (Lo w) 2 3 4 5 (High) High-Lo w αAH F -0.0129 ∗∗ -0.00637 ∗ -0.00569 + 0.00288 0.00261 0.00975 ∗∗ (-4.54) (-2.28) (-1.83) (0.92) (0.87) (4.61) αB S -0.0102 ∗∗ -0.00995 ∗∗ -0.00303 0.00499 ∗ 0.00839 ∗∗ 0.0131 ∗∗ (-4.49) (-4.66) (-1.34) (2.08) (3.61) (7.44) αC H M -0.00592 ∗∗ -0.00214 0.00260 0 .00441 + 0.00690 ∗∗ 0.00857 ∗∗ (-2.79) (-1.04) (1.15) (1.87) (3.27) (5.42) αF O T -0.00269 -0.00395 + 0.000872 0.00473 + 0.0105 ∗∗ 0.00715 ∗∗ (-1.36) (-1.89) (0.47) (2.12) (5.56) (5.05) αF R I -0.00909 ∗∗ -0.00513 ∗ -0.00258 0.00713 ∗∗ 0.00593 ∗∗ 0.0115 ∗∗ (-4.40) (-2.55) (-1.13) (3.51) (2.96) (7.07) αM Q -0.0157 ∗∗ -0.00230 -0.00165 0.00539 ∗ 0.00696 ∗∗ 0.0123 ∗∗ (-4.80) (-0.86) (-0.59) (2.05) (2.64) (5.56) αT R A -0.0131 ∗∗ -0.00931 ∗∗ -0.00136 0.00602 ∗ 0.00695 ∗ 0.0132 ∗∗ (-4.91) (-3.64) (-0.49) (2.11) (2.57) (6.64) αE E Q -0.0131 ∗∗ -0.0139 ∗∗ -0.000882 0.00272 0.00829 ∗∗ 0.0152 ∗∗ (-4.98) (-5.65) (-0.33) (0.99) (3.28) (7.78) t statistics in paren theses + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

T able 6: Hyp othesis 1 Industry Lev el This table presen ts the output of a re gression of mon thly Industr y Index returns on a set of regressors for the p erio d 1990-2013. R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th return and the a v era ge return from 12 to 2 mon ths ago of that same industry . P O R Tt is the retu rn of a p ortfolio that consists of coun try indexes w eigh te d b y the v alue of trade of an in dustry’s tr ading pa rtners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return and the a v erage return from 1 2 to 2 mon ths ago of the same p ortfolio. Excluded are the co efficien ts of mon th fixed effects. Standard errors (not rep orted) are robust for heter osk edasticit y. R elatively T rade Dep endent R elatively T rade Indep endent TRA EEQ CHM MQ BS F OT FRI AHF R ett− 2: t− 12 0.190 ∗∗ 0.133 ∗∗ 0.162 ∗∗ 0.179 ∗∗ 0.141 ∗∗ 0.124 ∗ 0.211 ∗∗ 0.241 ∗∗ (4.05) (2.87) (3.20) (3.76) (2.93) (2.36) (3.71) (4.70) R ett− 1 0.0158 0.0217 0.0270 + 0.0552 ∗∗ 0.0164 0.0168 0.0313 0.0128 (1.05) (1.56) (1.77) (3.56) (0.97) (1.18) (1.42) (0.67) P O R Tt− 2: t− 12 -0.313 ∗∗ -0.224 ∗∗ -0.309 ∗∗ -0.0605 -0.293 ∗∗ -0.201 ∗∗ -0.333 ∗∗ -0.179 ∗ (-3.89) (-2.86) (-5.10) (-0.85) (-4.44) (-3.97 ) (-4.98) (-2.19) P O R Tt 0.785 ∗∗ 0.810 ∗∗ 0.614 ∗∗ 0.464 ∗∗ 0.749 ∗∗ 0.459 ∗∗ 0.789 ∗∗ 0.483 ∗∗ (28.32) (33.21) (29.64) (17.17) (35.08) (24.31 ) (38.11) (18.65) P O R Tt− 1 0.152 ∗∗ 0.135 ∗∗ 0.0530 ∗ 0.0460 + 0.105 ∗∗ 0.0684 ∗∗ 0.112 ∗∗ 0.105 ∗∗ (5.37) (5.43) (2.46) (1.91) (4.57) (3.57) (4.42) (3.49) constant 0.000816 0.00556 -0.00224 -0.0 0323 0.00624 -0.00658 -0.00854 0.00663 (0.13) (0.96) (-0.41) (-0.49) (1.07) (-1.23) (-1.61) (0.94) R 2 0.143 0.172 0.139 0.101 0.190 0.099 0.239 0.082 N 7197 7333 7649 653 8 7196 7705 7509 5054 t statistics in paren theses. + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

T able 7: Hyp othesis 2 Industry Lev el This table presen ts the output of a regression of mon thly Coun try Index returns on a set of regressors for th e p erio d 1990-2013. R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th return and the a v erage return from 12 to 2 mon ths a go of that same industry . P O R Tt is the return of a p ortfolio that consists of coun try indexes w eigh ted b y the v alue of trade of a n industry’ s trading partners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return and the a v erage return from 12 to 2 mon ths ago of the same p ortfolio. O peness is a v ariable that indicates the share of pro duction that is earned outside the h ome coun try . hig hO peness is an indicator equal to one if a coun try b elongs to the top 20% of trade -op enness.Excl uded are the co efficien ts of mon th fixed effects. Standard errors (not rep orted) are robust to heterosk edasticit y. R elatively T rade Dep endent R elatively T rade Indep endent TRA EEQ CHM MQ BS F OT FRI AHF R ett− 2: t− 12 0.190 ∗∗ 0.130 ∗∗ 0.162 ∗∗ 0.178 ∗∗ 0.141 ∗∗ 0.124 ∗ 0.209 ∗∗ 0.235 ∗∗ (4.05) (2.80) (3.19) (3.74) (2.93) (2.36) (3.75) (4.58) R ett− 1 0.0161 0.0216 0.027 2 + 0.0534 ∗∗ 0.0167 0.0167 0.0305 0.0116 (1.07) (1.56) (1.78) (3.43) (0.98) (1.17) (1.39) (0.60) P O R Tt− 2: t− 12 -0.314 ∗∗ -0.221 ∗∗ -0.310 ∗∗ -0.0601 -0.293 ∗∗ -0.201 ∗∗ -0.334 ∗∗ -0.173 ∗ (-3.91) (-2.83) (-5.12) (-0.84) (-4.44) (-3.97) (-4.99) (-2.12) P O R Tt 0.785 ∗∗ 0.810 ∗∗ 0.614 ∗∗ 0.464 ∗∗ 0.749 ∗∗ 0.459 ∗∗ 0.788 ∗∗ 0.483 ∗∗ (28.33) (33.20) (29.6 2) (17.24) (35.11) (24.30) (38.11) (18.68) P O R Tt− 1 0.140 ∗∗ 0.144 ∗∗ 0.0572 ∗ 0.0318 0.108 ∗∗ 0.0650 ∗∗ 0.0954 ∗∗ 0.0964 ∗∗ (4.54) (5.46) (2.48) (1.24) (4.31) (3.11) (3.68) (3.11) hig hO penness 0.00199 -0.00535 -0.00101 0.000837 0.00019 7 0.0000209 -0.00156 -0.00794 ∗ (0.61) (-1.61) (-0.41) (0.26) (0.08) (0.01) (-0.66) (-2.04) hig hO penness × P O R Tt− 1 0.0683 -0.0473 -0.0241 0.121 ∗ -0.0168 0.0180 0.0967 0.0391 (0.99) (-0.76) (-0.44) (2.16) (-0.39) (0.37) (1.58) (0.50) constant 0.000475 0.006 44 -0.00208 -0.003 60 0.00619 -0.00657 -0.00820 0.00798 (0.08) (1.10) (-0.38) (-0.54) (1.06) (-1.22) (-1.53) (1.12) R 2 0.143 0.173 0.139 0.102 0.190 0.099 0.240 0.083 N 7197 7333 7649 6538 7196 7705 7509 5054 t statistics in paren theses. + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

T able 8: h yp othesis 3 Industry Lev el his ta ble presen ts the output of a regression of mon thly Industry Index returns on a set of re gressors for the p erio d 1990-2013. R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th return and the a v erage return from 12 to 2 mon ths a go of that same industry . P O R Tt is the return of a p ortfolio that consists of coun try indexes w eigh ted b y the v alue of trade of a n industry’ s trading partners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return and the a v erage return from 12 to 2 mon ths ago of the sa me p ortfolio. The Herfindahl Index (H H I ) is a measure of complex it y of trade relations (lo w er is more complex). low H H I is an indicator v ariable equal to one if a coun try b elongs to the top 20% of comple x trade relationship s. Exc luded are the co efficien ts of mon th fixed effects. Standard errors (not rep orted) are robust to heterosk edasticit y. AHF CHM F OT MQ TRA EEQ R ett− 2: t− 12 0.240 ∗∗ 0.162 ∗∗ 0.121 ∗ 0.175 ∗∗ 0.187 ∗∗ 0.133 ∗∗ (4.68) (3.21 ) (2.32) (3.66) (3.97) (2.86) R ett− 1 0.0126 0.0270 + 0.0167 0.0533 ∗∗ 0.0157 0.0222 (0.66) (1.77 ) (1.17) (3.41) (1.04) (1.60) P O R Tt− 2: t− 12 -0.177 ∗ -0.310 ∗∗ -0.201 ∗∗ -0.0509 -0.31 0 ∗∗ -0.221 ∗∗ (-2.17) (-5.11) (-3.97) (-0.71) (-3.86) (-2.82) P O R Tt 0.483 ∗∗ 0.614 ∗∗ 0.459 ∗∗ 0.465 ∗∗ 0.785 ∗∗ 0.811 ∗∗ (18.63) (29.62) (24.30) (17.28) (28.32) (33.24) P O R Tt− 1 0.0993 ∗∗ 0.0587 ∗ 0.0711 ∗∗ 0.0705 ∗ 0.150 ∗∗ 0.147 ∗∗ (3.01) (2.44 ) (3.39) (2.32) (5.08) (5.65) low H H I 0.000550 -0.00101 0.00228 -0.00397 0.00534 0.00170 (0.15) (-0.4 3) (0.99) (-1.39) (1.37) (0.42) low H H I × P O R Tt− 1 0.0336 -0.0276 -0.0 0969 -0 .0430 0.0271 -0.112 (0.50) (-0.6 1) (-0.22) (-1.01) (0.31) (-1.63) constant 0.00658 -0.00202 -0.00710 -0.00207 0.000417 0.00536 (0.93) (-0.3 7) (-1.32) (-0.31) (0.07) (0.92) R 2 0.083 0.139 0.100 0.102 0.143 0.173 N 5 054 7649 7705 6538 7197 7333 t statistics in paren theses + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

way of studying openness in the entire data set (8 industries together). Instead of sorting on Openness and HHI within every industry, one might as well sort the entire data set into different quintiles of openness and complexity. It is important to realize that characteristics among these different quintiles might differ substantially. In other words, industries (in certain countries) that are relatively open for trade, are not a proper ”control group” for those in relatively closed industries. In order to overcome this problem, standardized Betas can be applied. Standardized Betas are expressed in standard deviations of the dependent variable, hence one can safely compare the magnitude of the coefficients between different quintiles.

In table 9 the data is sorted according to trade openness and divided in quintiles.

We are mainly interested in the predictive power of P ORTt−1 and see a somewhat

increasing effect throughout the quintiles. It seems to provide some evidence of a rela-tionship between trade and return predictability, however when we test for statistical significance between column 1 and 5, we find a p-value of 0.6225, indicating no signifi-cant difference. An alternative specification where the data is sorted into 10 quantiles

reveals that in the lowest quantile (very little trade) P ORTt−1 becomes insignificant.

This procedure has also been applied by sorting the data according to a its complex-ity of trade relationships. Business Services and Financial Services are again excluded from the sample because of the reasons stated earlier. Note that a low Herfindahl Index implies that a certain industry has many trading partners and is therefore complex.

Table 10 presents the results of the regression for different quintiles. P ORTt−1 is more

or less increasing throughout the quintiles, which supports hypothesis 3. However, when we test for significance of the difference between column 1 and 5 we find that is not statistically significant (p=0.1503).

Altogether the analysis on the industry level provides mixed results. Analysis within industries, show significant predictive power of trading-partner portfolio returns. How-ever, within industries we can find hardly any evidence that the effect is stronger in more open industries or lower in more complex industries. Yet, the quintile regressions find weak support for hypothesis 2 and 3.

Table 9: Regressions for different quintiles of trade openness

This table presents the output of a regression of monthly Industry Index returns on a set of regressors for the period 1990-2013. The data is sorted in quintiles according to

trade-openness.Rett is the monthly return of a certain industry. Rett−1 and Rett−2:t−12 are

the previous month return and the average return from 12 to 2 months ago of that same

industry. P ORTt is the return of a portfolio that consists of country indexes weighted

by the value of trade of an industry’s trading partners. P ORTt−1 and P ORTt−2:t−12

are the previous month’s return and the average return from 12 to 2 months ago of the same portfolio. Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust for heteroskedasticity. The reported Betas are standardized i.e. how many standard deviations change in Industry Return does a given regressor yield. 1 (Low) 2 3 4 5 (High) Rett−2:t−12 0.0000793 0.000562 -0.00331 0.0459∗∗ 0.0134+ (0.16) (0.49) (-1.17) (3.17) (1.65) Rett−1 0.0465∗∗ 0.0399∗∗ 0.0474∗ 0.0363∗ 0.0215 (2.63) (2.90) (2.76) (2.95) (1.79) P ORTt−2:t−12 -0.142∗∗ -0.113∗ -0.102∗ -0.0822 -0.102+ (-3.13) (-2.37) (-2.15) (-1.58) (-1.76) P ORTt 0.494∗∗ 0.613∗∗ 0.617∗∗ 0.704∗∗ 0.797∗∗ (29.20) (35.68) (29.28) (35.55) (37.67) P ORTt−1 0.0777∗∗ 0.0770∗∗ 0.0678∗∗ 0.0935∗∗ 0.0938∗∗ (3.88) (4.00) (3.18) (4.87) (4.41) constant -0.00522 -0.000469 0.00325 0.00599 0.00517 (-1.21) (-0.10) (0.61) (1.31) (1.08) R2 _{0.108 0.142 0.155 0.164 0.179} N 9828 9907 9592 9782 9898 t statistics in parentheses+ _{p < 0.05,}∗ _{p < 0.01,}∗∗ _{p < 0.001}

Table 10: Regressions for different quintiles of Herfindahl Index

This table presents the output of a regression of monthly Industry Index returns on a set of regressors for the period 1990-2013. The data is sorted in quintiles according to

Trade complexity (Herfindahl Index).Rett is the monthly return of a certain industry.

Rett−1 and Rett−2:t−12 are the previous month return and the average return from 12

to 2 months ago of that same industry. P ORTtis the return of a portfolio that consists

of country indexes weighted by the value of trade of an industry’s trading partners.

P ORTt−1 and P ORTt−2:t−12 are the previous month’s return and the average return

from 12 to 2 months ago of the same portfolio. Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust for heteroskedasticity. The reported Betas are standardized i.e. how many standard deviations change in Industry Return does a given regressor yield.

1 (Low) 2 3 4 5 (High) Rett−2:t−12 0.00246 0.0000560 0.00127 0.0530∗∗ 0.0132+ (0.80) (0.17) (0.15) (3.00) (1.92) Rett−1 0.0397∗∗ 0.0428∗∗ 0.00428 0.0154 0.0550∗∗ (2.77) (3.07) (0.31) (1.11) (3.73) P ORTt−2:t−12 -0.0185 -0.108∗ -0.132∗ -0.120+ -0.153∗ (-0.32) (-2.22) (-2.34) (-1.88) (-2.19) P ORTt 0.466∗∗ 0.578∗∗ 0.632∗∗ 0.647∗∗ 0.682∗∗ (19.53) (30.49) (27.62) (29.67) (25.29) P ORTt−1 0.0444∗ 0.0657∗∗ 0.103∗∗ 0.117∗∗ 0.0920∗∗ (2.01) (3.34) (4.65) (4.91) (3.71) constant 0.000151 0.000889 -0.00254 0.00195 -0.00487 (0.03) (0.21) (-0.39) (0.33) (-0.96) R2 _{0.097 0.139 0.118 0.121 0.127} N 7852 8552 8409 8421 8242 t statistics in parentheses+ _{p < 0.1,}∗ _{p < 0.05,}∗∗ _{p < 0.01}

4.3

Alternative Explanations

So far there is no convincing evidence that the observed return predictability of the trading-partner portfolio is really following from trade-relations. This section tests some extensions and alternative explanations for the observed predictive power.

Some industries are dependent on exports, some are more dependent on imports. In fact, one might argue that trading-partner portfolios should be weighted by either imports or exports, depending on this dependency. In table 13 we test whether a trading partner portfolio that is import weighted when an industry is import depen-dent (import > export) and export weighted when an industry is export dependepen-dent (export > import) has an improved model fit. This is tested at the country level and the result can be found in column 3. We observe that both the coefficient on the

lagged portfolio and the R2 _{are almost equal to the specification in table 4 (where}

the portfolio is weighted with both imports and exports). In column 1 of table 13 the trading-partner portfolio is weighted according to only exports, which results in a somewhat better fit of the model. Using only imports (column 2) is an inferior model with respect to model fit.

One might argue that the observed explanatory power of the trading-partner port-folio is simply due to region-wide shocks. In other words, it is maybe not trade that induces return predictability, but rather region-wide shocks that are partly captured by the trading-partner portfolio. To test this hypothesis, industry in countries are divided into 5 broad regions. Europe North America, South America, Asia + Oceania and Others. Subsequently, a new variable is generated that represents the share of trade outside of one’s own region. If return predictability is really a regional effect, we should not observe it for industries that trade a lot outside their own region. Because of the analysis on the industry level, we find a wide distribution of the new defined variable (5% of the sample has more than 80% outside region). The data is then sorted according to the share of outside region trade. Table 11 presents the results of the regression for the different quintiles. We see that the lagged return of the

trading-partner portfolio is significant for all the quintiles. We even find that it is the highest for industries that have relatively the most outside region trade. Even when the data is sorted into 10 deciles, the return predictability remains present in all deciles. Based on this regression we conclude that predictive power of the trading-partner portfolio is not caused by region wide shocks.

A last flaw in the sample might be the inclusion of relatively illiquid stock mar-kets. Since these markets are less efficient in updating prices to new information, the results of this thesis might as well be driven by these illiquid stock markets. To test for this possibility, I make the strong assumption that the relatively poor countries in the sample have the most illiquid stock-markets (China, India, Indonesia, Brazil, Argentina and South Africa). Table 12 reports the results of two regressions based on either liquid or illiquid markets. The table reports standardized Betas and finds support for the claim that return predictability is more of an issue in illiquid markets (higher coefficient on lagged trading-partner portfolio return). However, the results for the liquid markets are strongly significant as well, implying that the findings are not solely driven by the illiquidity of stock-markets.

It might as well be the case that some country’s are in fact leading the stock mar-kets of other countries. This finding is similar to the finding by Hong et al. (2007), who finds that some industries lead other industries. In order to test this hypothesis, alter-native portfolios have been put to the test (not reported). The U.S. stock market is often considered to be leading other stock markets (Rapach et al., 2013). However, the U.S. stock market has much less predictive power than the trading-partner portfolios. Alternatively, a portfolio of the 5 leading (largest) economies is build, consisting of the United States, China, Japan, U.K. and Germany. This portfolio also performs far worse than the trading-partner portfolio. Finally, the World Index is employed, which also fails to improve upon the existing model. All the alternative explanations tested in this section are unable to falsify the statement that return predictability follows from trade-based relationships between countries or industries.

Table 11: Regressions for different quintiles of Outside Region Trade

This table presents the output of a regression of monthly Industry Index returns on a set of regressors for the period 1990-2013. The data is sorted in quintiles according to

trade shares to regions outside ones own region.Rett is the monthly return of a certain

industry. Rett−1 and Rett−2:t−12 are the previous month return and the average return

from 12 to 2 months ago of that same industry. P ORTt is the return of a portfolio

that consists of country indexes weighted by the value of trade of an industry’s trading

partners. P ORTt−1and P ORTt−2:t−12are the previous month’s return and the average

return from 12 to 2 months ago of the same portfolio. Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust for heteroskedasticity. The reported Betas are standardized i.e. how many standard deviations change in Industry Return does a given regressor yield.

1 (Low) 2 3 4 5 (High) Rett−2:t−12 0.0603+ 0.00939 0.0201+ -0.000785 0.00148 (1.81) (1.04) (1.75) (-0.90) (0.16) Rett−1 0.0218+ 0.0176 0.0482∗∗ 0.0465∗∗ 0.0275+ (1.84) (1.57) (4.13) (3.05) (1.91) P ORTt−2:t−12 -0.150∗ -0.219∗∗ -0.0320 -0.0306 -0.176∗∗ (-2.14) (-4.53) (-0.69) (-0.63) (-3.42) P ORTt 0.557∗∗ 0.775∗∗ 0.616∗∗ 0.612∗∗ 0.622∗∗ (22.81) (40.20) (33.12) (32.78) (31.45) P ORTt−1 0.0830∗∗ 0.0793∗∗ 0.0663∗∗ 0.0783∗∗ 0.117∗∗ (3.90) (3.93) (3.71) (4.02) (5.64) constant -0.0155∗ -0.00293 0.000216 0.00635+ _0.00529 (-2.43) (-0.55) (0.05) (1.76) (1.38) R2 _{0.105 0.182 0.158 0.136 0.110} N 10207 10977 12112 11516 11369 t statistics in parentheses+ _{p < 0.1,}∗ _{p < 0.05,}∗∗ _{p < 0.01}

Table 12: Regressions for liquid and illiquid Markets

This table presents the output of a regression of monthly Industry Index returns on a set of regressors for the period 1990-2013. The Data is divided in liquid vs illiquid

stock-markets. Rett is the monthly return of a certain industry. Rett−1 and Rett−2:t−12 are

the previous month return and the average return from 12 to 2 months ago of that same

industry. P ORTt is the return of a portfolio that consists of country indexes weighted

by the value of trade of an industry’s trading partners. P ORTt−1 and P ORTt−2:t−12

are the previous month’s return and the average return from 12 to 2 months ago of the same portfolio. Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust for heteroskedasticity. The reported Betas are standardized i.e. how many standard deviations change in Industry Return does a given regressor yield. (1) (2) Illiquid Liquid Rett−2:t−12 0.231∗∗ 0.157∗∗ (5.49) (8.13) Rett−1 0.0531∗∗ 0.0229∗∗ (3.79) (3.58) P ORTt−2:t−12 -0.309∗∗ -0.207∗∗ (-5.90) (-6.91) P ORTt 0.524∗∗ 0.663∗∗ (26.99) (64.67) P ORTt−1 0.107∗∗ 0.0828∗∗ (5.01) (8.49) constant 0.0104∗∗ -0.00306 (2.99) (-1.26) R2 _{0.114 0.147} N 10205 45976 t statistics in parentheses. + p < 0.05,∗ p < 0.01,∗∗ p < 0.001

Table 13: Country Level Regressions with different TTPs

This table presents the output of a regression of monthly Country Index returns on

a set of regressors for the period 1990-2013. Rett is the monthly return of a certain

industry. Rett−1 and Rett−2:t−12 are the previous month return and the average return

from 12 to 2 months ago of that same industry. P ORTt is the return of a portfolio that

consists of country indexes weighted by the value of either exports, imports or imports

or exports depending on trade balance of a country’s trading partners. P ORTt−1 and

P ORTt−2:t−12 are the previous month’s return and the average return from 12 to 2

months ago of the same portfolio. Openess is a variable that indicates the share of production that is earned outside the home country. The Herfindahl Index (HHI) is a measure of complexity of trade relations (lower is more complex). lowHHI is an indicator variable equal to one if a country belongs to the top 20% of complex trade relationships. highOpeness is an indicator equal to one if a country belongs to the top 20% of trade-openness.Excluded are the coefficients of month fixed effects. Standard errors (not reported) are robust to heteroskedasticity.

(1) (2) (3)

Export Import M ixed

Rett−2:t−12 0.149∗ 0.145 0.147∗ (2.17) (1.64) (2.15) Rett−1 0.0146 0.0205 0.0194 (0.63) (0.93) (0.86) P ORTt−2:t−12 -0.205∗∗ -0.161∗ -0.202∗∗ (-2.80) (-2.21) (-2.85) P ORTt 0.879∗∗ 0.831∗∗ 0.855∗∗ (51.56) (44.53) (45.83) P ORTt−1 0.0498∗ 0.0491∗ 0.0505+ (2.04) (2.19) (1.88) constant 0.00545+ _0.00499+ _0.00529+ (1.91) (1.77) (1.88) R2 _{0.402 0.387 0.395} N 7872 8004 7992 t statistics in parentheses. + p < 0.1,∗p < 0.05, ∗∗ p < 0.01

4.4

Robustness Checks

This section will put the findings of this thesis to some additional tests, that test both the internal and external validity of the results. An important issue to deal with in panel data is clustering. Since the panel data is in fact estimated cross-sectionally (Fama-MacBeth Regression), one often observes that residuals are in fact clustered among entities. Table 14 columns 1-3 test on a country level whether the results re-main significant with clustered standard errors. In fact we find that the significance of the specification in column 1 now drops to 10%. The results in Column 2 and 3 remain unchanged. Table 15 now tests the effect of clustering on an industry level. All these results remain significant. Hence, we conclude that the results are robust to clustering. Rizova (2010) alternatively uses clustering per month. The results in this thesis are also robust to clustering per month.

Throughout the paper the data has often been sorted into top and bottom quintiles. On a country level, we did not find evidence that openness is related to more return predictability. Table 11 now tests an alternative specification where the top or bottom tertile (=33%) is used. However,in column 3 we still do not find evidence that return predictability is related to openness. Even when the middle tertile is omitted (column 7), we do not find a positive relation between openness and return predictability. For the Herfindahl Index we still find significantly less return predictability in complex trade relationships (column 4). Yet, if we only look at the extremes (column 6) this effects seems to disappear.

One might be concerned that one group of countries is very concentrated among the open industries and another group among the more closed industries. By looking at how the countries are represented among the five quintiles of trade-openness, we see that in all the quintiles there are at least 23 different countries represented for which at least 100 observations are available. This result seems surprising, but it is likely due to the fact that industries themselves vary widely in openness in almost every country. The latter finding confirms that the results are not only driven by a few countries.

One might argue that return predictability is more of an issue of the past. Markets have become much more efficient throughout time and therefore the results in this paper might be driven by certain early time-periods in the data. To test this hypothe-sis, the regression specification of hypothesis 1 has been ran for different time-periods. Table 16 presents the results of the regressions of six 4-year periods. Due to limited data availability and manipulations of the data, only few observations are available in the first period. We do not find significant return predictability here. The periods thereafter we do find significant return predictability and moreover we find a clearly decreasing pattern in the magnitude of the coefficients. Because the Betas have been standardized, we find clear support that the relationship between the trading-partner portfolio and industry returns is diminishing. However, the difference between the

co-efficient on P ORTt−1 in period 1990-1993 and 2010-2013 is not statistically significant

(p=0.2377). If we look at the explanatory power of the model, we observe an opposite pattern. The increase in the fit of the model might as well be driven by the other explanatory variables.

T able 14: Coun try-lev el Analysis Robustness Chec ks This regression table sho ws an alternativ e sp ecification of ta ble 4. low H H I (< 33%) are hig hopenness (> 33%) in this sp ec-ification the b ottom and top tertile of the resp ectiv e v aria ble. Column 1 to 3 uses clustering robust standard deviations (on the coun try le v el). Column 4-5 chec ks the new sp ecification of low H H I and hig hopeness with normal heterosk edasticit y robust standard err ors. Column 6-7 do es the same but omits the middle tert ile (33%) of the data. (1) (2) (3) (4) (5) (6 ) (7) R ett R ett R ett R ett R ett R ett R ett R ett− 2: t− 12 0.122 ∗∗ 0.123 ∗∗ 0.121 ∗∗ 0.123 ∗∗ 0.121 ∗∗ 0.0963 ∗ 0.0741 + (3.07) (3.05) (3.00) (3.39) (3.36) (2. 09) (1.67) R ett− 1 0.0162 0.0176 0.0163 0.0176 0.0163 0.0 0897 0.0267 + (0.79) (0.84) (0.79) (1.55) (1.44) (0. 64) (1.89) P O R Tt− 2: t− 12 -0.168 ∗∗ -0.165 ∗∗ -0.168 ∗∗ -0.165 ∗∗ -0.168 ∗∗ -0.106 + -0.106 + (-3.36) (-3.27) (-3.38) (-3.29) (-3.35) (-1 .80) (-1.73) P O R Tt 0.860 ∗∗ 0.861 ∗∗ 0.860 ∗∗ 0.861 ∗∗ 0.860 ∗∗ 0.839 ∗∗ 0.819 ∗∗ (16.90) (16.88) (16.92) (68.26) (68.15) (62 .05) (53.08) P O R Tt− 1 0.0518 + 0.0706 ∗∗ 0.0437 0.0706 ∗∗ 0.0437 ∗ 0.0616 ∗∗ 0.0244 (2.04) (2.97) (1.40) (3 .96) (2.38) (2.9 0) (0.99) low H H I 0.000969 0.000969 0.00258 + (0.96) (0.72) (1.87) low H H I × P O R Tt− 1 -0.0614 ∗ -0.0614 ∗ -0.0408 (-2.10) (-2.34) (-1.54) hig hO penness -0.00123 -0.00123 -0.00158 (-1.19) (-0.90) (-0.99) hig hO penness × P O R Tt− 1 0.0223 0.0223 0.0368 (0.86) (0.87) (1.23) constant 0.00504 + 0.00471 0.00545 + 0.00471 ∗ 0.00545 ∗ 0.00125 0.00297 (1.76) (1.65) (1.90) (2 .10) (2.43) (0.5 1) (1.03) R 2 0.396 0.397 0.396 0 .397 0.396 0.4 51 0.388 N 7812 7812 7812 7812 7812 5136 4968 Clustered SD X X X T ertiles All All All All All 1&3 1&3 t statistics in paren theses. + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

T able 15: Industry Lev el Regression Robust for Clustering This table presen ts the output of a regression of mon thly Coun try Index returns on a set of regressors for th e p erio d 1990-2013. R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th return and the a v era ge return from 12 to 2 mon ths ago of that same industry . P O R Tt is the retu rn of a p ortfolio that consists of coun try indexes w eigh te d b y the v alue of trade of an in dustry’s tr ading pa rtners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return and the a v erage return from 12 to 2 mon ths ago o f the same p ortfolio .Excluded are the co efficien ts of mon th fixed effects. Standard errors (not rep orted) are robust to clusteri ng (coun tr y lev el). R elatively T rade Dep endent R elatively T rade Indep endent TRA EEQ CHM MQ BS F OT FRI AHF R ett− 2: t− 12 0.190 ∗∗ 0.133 ∗∗ 0.162 ∗∗ 0.179 ∗∗ 0.141 ∗∗ 0.124 + 0.211 ∗∗ 0.241 ∗∗ (4.57) (3.20) (2.79) (4.01) (4.07) (2.02) (4.30) (6.23) R ett− 1 0.0158 0.0217 0.0270 + 0.0552 ∗∗ 0.0164 0.0168 0.0313 0.0128 (1.11) (1.46) (2.03) (3.69) (0.81) (1.09) (1.04) (0.69) P O R Tt− 2: t− 12 -0.313 ∗∗ -0.224 ∗∗ -0.309 ∗∗ -0.0605 -0.293 ∗∗ -0.201 ∗∗ -0.333 ∗∗ -0.179 ∗ (-3.20) (-3.50) (-4.31) (-0.82) (-5.97) (-4.33 ) (-5.17) (-2.39) P O R Tt 0.785 ∗∗ 0.810 ∗∗ 0.614 ∗∗ 0.464 ∗∗ 0.749 ∗∗ 0.459 ∗∗ 0.789 ∗∗ 0.483 ∗∗ (14.92) (13.21) (13.99) (8.24) (13.90) (18.16) (18.99) (13.00) P O R Tt− 1 0.152 ∗∗ 0.135 ∗∗ 0.0530 ∗ 0.0460 0.105 ∗∗ 0.0684 ∗∗ 0.112 ∗∗ 0.105 ∗∗ (4.79) (5.14) (2.28) (1.57) (4.09) (3.50) (4.20) (3.77) constant 0.000816 0.00556 -0.00224 -0.0 0323 0.00624 -0.00658 -0.00854 ∗ 0.00663 (0.24) (1.21) (-0.57) (-0.47) (1.28) (-1.30) (-2.14) (0.94) R 2 0.143 0.172 0.139 0.101 0.190 0.099 0.239 0.082 N 7197 7333 7649 6538 7196 7705 7509 5054 Clustered SD X X X X X X X X t statistics in paren theses. + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

T able 16: Regressions All Industries o v er Time This table presen ts the output of a re gression of mon thly Industr y Index returns on a set of regressors for the p erio d 1990-2013. The data is sorted in quin tiles according to T rade complex it y (Herfindahl Index). R ett is the mon thly return of a certain industry . R ett− 1 and R ett− 2: t− 12 are the previous mon th retu rn and the a v erage return from 12 to 2 mon ths ago of that same industry . P O R Tt is the return of a p ortfolio that consists of coun try indexes w eigh ted b y the v alue of trade of an industry’s trading partners. P O R Tt− 1 and P O R Tt− 2: t− 12 are the previous mon th’s return and the a v erage return from 12 to 2 mon ths ago of the same p ortfolio. Excluded are the co efficien ts of mon th fixed effects. Standard errors (not rep orted) are robust for heterosk edasticit y. The rep orte d Betas are standardized i.e. ho w ma n y stand ard deviations change in Industry Return do es a giv en regressor yield. 1990-1993 1994-1997 1998 -2001 2002-2005 2006-2009 2010-2013 R ett− 2: t− 12 0.278 ∗ 0.149 ∗∗ 0.0775 ∗∗ 0.137 ∗∗ 0.103 ∗∗ 0.342 ∗∗ (2.43) (5.21) (2.61) (5.11) (3.63) (12.96) R ett− 1 0.170 ∗∗ 0.00938 0.0206 ∗ 0.0161 0.0 567 ∗∗ 0.0111 (3.71) (0.95) (2.19) (1.73) (6.05) (1.20) P O R Tt− 2: t− 12 0.0752 -0.0738 -0.19 2 ∗∗ -0.169 ∗∗ -0.148 ∗∗ -0.253 ∗∗ (0.11) (-1.06) (-2.93) (-3.69) (-3.81) (-5.02) P O R Tt -0.347 + 0.605 ∗∗ 0.520 ∗∗ 0.635 ∗∗ 0.694 ∗∗ 0.592 ∗∗ (-1.82) (27.66) (24.32) (39.62) (50.11) (37.94) P O R Tt− 1 -0.107 0.106 ∗∗ 0.112 ∗∗ 0.120 ∗∗ 0.0748 ∗∗ 0.0655 ∗∗ (-0.54) (4.69) (5.11) (7.07) (4.81) (3.97) constant 0.0555 ∗ 0.0147 ∗∗ -0.0573 ∗∗ 0.0151 ∗∗ 0.0126 ∗∗ 0.0146 ∗∗ (3.27) (4.61) (-13.58) (5.71) (4.63) (6.42) R 2 0.155 0.101 0.098 0.140 0.269 0.187 N 534 10255 11176 11363 11303 11550 t statistics in paren theses + p < 0 .1, ∗ p < 0 .05, ∗∗ p < 0 .01

5

Conclusion

Information in foreign markets often contains relevant information for domestic mar-kets. Due to the difficulty to analyze this information, this information only enters domestic markets with a lag. Rizova (2010) argues that return predictability follows from trade-based relationships between countries. However, his findings are in many instances conflicting with the more ”random” patterns found by Rapach et al. (2013). This thesis has aimed to explore whether international return predictability is really a phenomenon between countries that trade and finds no evidence that can falsify this statement.

The findings show very strong predictive power of trading-partner portfolios on both the country and industry level. A long short strategy can yield yearly alphas of around 14.43%. However, evidence that the predictive power is related to trade is mixed. For instance, there is no evidence that the effects are stronger in countries or within industries that are more open. However, when appending the eight industries and sorting the data based on openness, higher return predictability is found in higher openness quintiles, although not significant.

Complexity of trade relationships on the other hand seems to play an important role in diminishing the magnitude of return predictability. The Herfindahl Index ap-pears negatively significant in all specifications except on the separate industry level. Additionally when sorting the entire dataset in different quintiles according to com-plexity, we see lower return predictability in the most complex quintiles (although not statistically significant).

Several robustness checks show that the data is relatively robust to clustering. Furthermore it is showed that illiquidity does not drive the return predictability. Al-ternative explanations such as region wide shocks do also not drive the found return predictability, because return predictability is also very strong for industries who’s trading-partners are mostly outside their own region. Additionally, alternative indexes do not seem to be able to achieve a better fit. Altogether, the findings in this thesis

are unable to reject nor confirm the statement that international return predictability follows from trade-based relationships.

Future research should aim at exploring alternative portfolios that might exhibit superior predictive power (based on countries with same culture, language, political ties etc.). Furthermore it would be beneficial to find a solution for the inclusion of the so called Fama-French factors at the country or industry level. The latter would clearly improve the accuracy of the estimates of the Alphas of a long-short strategy. Additionally, one might be able to quantify the transaction costs of the long-short strategy to find what such a strategy would effectively yield.

References

Brad M. Barber and Terrance Odean. All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. Review of Financial Studies, 21(2):785–818, 2008.

Lauren Cohen and Andrea Frazzini. Economic links and predictable returns. The Journal of Finance, 63(4):1977–2011, 2008.

Lauren Cohen and Dong Lou. Complicated firms. Journal of Financial Economics, 104(2):383–400, 2012.

Kent Daniel, David Hirshleifer, and Avanidhar Subrahmanyam. Investor psychology and security market undereactions and overreactions. the Journal of Finance, 53(6): 1839–1885, 1998.

Stefano DellaVigna and Joshua M. Pollet. Investor inattention and friday earnings announcements. The Journal of Finance, 64(2):709–749, 2009.

Harrison Hong, Walter Torous, and Rossen Valkanov. Do industries lead stock markets? Journal of Financial Economics, 83(2):367–396, 2007.

Narasimhan Jegadeesh and Sheridan Titman. Profitability of momentum strategies: An evaluation of alternative explanations. The Journal of Finance, 56(2):699–720, 2001.

Andrew W. Lo and Archie Craig MacKinlay. When are contrarian profits due to stock market overreaction? Review of Financial Studies, 3(2):175–205, 1990.

Lior Menzly and Oguzhan Ozbas. Market segmentation and cross-sectional predictabil-ity of returns. The Journal of Finance, 65(4):1555–1580, 2010.

Robert C. Merton. A simple model of capital market equilibrium with incomplete information. The journal of finance, 42(3):483–510, 1987.

Quoc H Nguyen. Geographic momentum. Available at SSRN 1921537, 2012.

David E Rapach, Jack K Strauss, and Guofu Zhou. International stock return pre-dictability: what is the role of the united states? The Journal of Finance, 68(4): 1633–1662, 2013.

Savina Rizova. Predictable trade flows and returns of trade-linked countries. In AFA 2011 Denver Meetings Paper, 2010.

X. Zhang. Information uncertainty and stock returns. The Journal of Finance, 61(1): 105–137, 2006.