Labor market and return predictability : cross-country analysis

University of Amsterdam

Amsterdam Business School

MSc Finance – Asset Management

Master Thesis

Labor Market and Return Predictability:

Cross-country Analysis

Author:

Hongyue Gong, 11374160

Supervisor:

Esther Eiling

July 2017, Amsterdam

Statement of Originality

This document is written by Student Hongyue Gong 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.

Acknowledgements

First and foremost, I would like to thank my supervisor Esther Eiling. She gives me a sustained and great help throughout the entire thesis process. She has rich knowledge, critical insight and patience. She is always willing to provide a qualified response for my questions. Thanks to her encouragement both for this paper and my future career.

Secondly, my gratitude goes to my family and friends who always give me support when I encountered difficult problems. Thank you for always being there for me.

Abstract

The first aim of this paper is to uncover the predictive relationship between labor adjustment cost and future stock market returns across countries. By employing 26 countries, including 21 developed and 5 developing markets, with more than 20000 monthly observations, I find evidence that labor adjustment cost that represented by the cross-sectional volatility of industry-specific stock returns (CSV) has varying predictive power across countries. Obviously, labor adjustment costs produced by human capital reallocation among industries fluctuate depending on labor conditions. This paper is the first to provide evidence of the link between labor market characteristics and both in-sample and out-of-sample forecast performance of the predictor. The empirical results reveal that the unemployment rate is statistically significant and positive with in-sample performance of CSV in long-term. But the growth of unemployment rate has no significant relationship with CSV in-sample performance. My results also show that the linkage between labor market flexibility and out-of-sample predictive performance of CSV is non-linear.

Keywords: return predictability, labor market conditions, labor adjustment costs,

Table of Content

1. Introduction... 1

2. Literature Review and Hypotheses Development ... 4

2.1. Labor Market and Stock Market Return Predictability ... 4

2.2. International Evidence of Stock Market Return Predictability ... 6

3. Methodology ... 8

3.1. Predictive Regression Model and Out-of-sample Statistical Analysis ... 8

3.2. Empirical Test for Labor Market Conditions and Forecast Performance ... 11

3.3. Robustness Tests ... 12

4. Data and Descriptive Statistics ... 13

5. Empirical Results of Predictability Test ... 17

5.1. Pooled Panel Regression ... 17

5.2. In-sample and Out-of-sample Return Predictability ... 17

6. Economic Channel for Different CSV Performance ... 20

6.1. INS CSV Performance and Unemployment Condition ... 20

6.2. OOS CSV Performance and Labor Market Flexibility ... 22

7. Robustness Tests ... 24

8. Conclusion ... 25

References ... 27

1

1. Introduction

Whether stock market returns have predictability has significant meaning for investors and academia. In light of powerful predictors, investors can make diversified and optimal asset allocation, while scholars can get a better understanding of asset pricing. Many proposed predictors are valuation variables since they are highly related with the stock price and reflect the price trend intuitively. For example, Campbell and Shiller (1998) extrapolate that there exists a linear relationship between dividend-price ratio and long-run stock market returns. Besides, some macro variables are also confirmed to have forecast power in previous literature. Wohar et al. (2005) confirm that short-term interest rates show statistically negative predictive ability across 1, 3, 12 and 24-month ahead stock market returns. However, there are only a few studies discuss the relationship between labor-related variables and future stock market returns, presumably because of the long time separation of research in terms of labor economics and asset pricing.

In neoclassical model, market value is directly determined by physical capital rather than human capital. Querying this benchmark, Merz and Yashiv (2007) first quantify the importance of labor market in explaining stock market valuation: Firms need to decide the optimal reallocation between hiring and investment. That means firms will face adjustment costs which affect their market value. So, financial market and labor market interact. Their results are consistent with the opinion of Arnott and Henriksson (1989) that the stock market is influenced by macroeconomic shocks and reflect the behavior of the macro economy. Boyd, Hu, and Jagannathan (2005) suggest a positive impact of rising unemployment on stocks returns during economic expansions and negative impact during contractions. Beyond all the attempts to find the linkage between labor market and future stock market returns, the latest research paper Eiling et al. (2016) shows that the cross-sectional volatility of industry-specific stock returns (hereafter CSV) as a proxy for labor adjustment cost has the strongest predictive power compared with other labor-related variables in U.S. market. Motivated by above evidence and follow Eiling et al. (2016), this paper continues the effort to understand the predictive relationship between the labor market and stock market rewards but instead explores the robustness of CSV for international stock markets. Because much of available literature is focused on U.S. market,

2

the non-U.S. market has received attention in terms of labor-return connection only recently. Studying the relationship between the labor market and return predictability in a global context has profound economic implications both for international labor economics and asset pricing.

I contribute to existing studies in several dimensions. Firstly, although there are some papers that study cross-country stock returns predictability, there is no paper study the different explanatory power of a predictor in different countries. This paper is the first to uncover the underlying reasons driving the cross-country differences in the performance of the predictor--CSV. Eiling et al. (2016) find the negative forecast power of CSV is much more stronger in higher performing industries that have a higher demand for skilled workers. That is because top performing industries should wait for a longer time to hire talents thus face more labor adjustment costs. Adopting the same logical, I link the cross-country differences in CSV performance to labor market conditions. More specifically, if one country has higher unemployment rate and workers need longer time to be re-employed, then the aggregate industries in this country will face more labor adjustment costs. Besides, another motivation to link with labor market conditions is that policies in different countries affect the size of costs associated with worker dismissal (Liosa et al., 2012) so that different labor market conditions can be the most important drivers for forecasting performance. In the main empirical tests, I define the labor conditions from two aspects, starting with the unemployment rate and changes in the unemployment rate. Then, I use proxies for labor market flexibility. In the robustness test, I use an indicator to proxy for the demand of high-skilled workers. Second, I investigate whether CSV has forecast ability for a set of markets for which there is essentially no previous proof. My sample includes both developed and emerging economies. Thirdly, compared with prior cross-country studies, I re-examine return predictability by using 7 alternative variables and various time horizons as a robustness check. In other words, this paper uses a broader sample and a more comprehensive time series data to research the role of labor adjustment cost in return predictability. The statistical benchmark for country-by-country analysis is Newey-West standard errors to deal with overlapping observations and serial correlation in the disturbance term. (Giot and Petitjean, 2006)

3

The following are my 4 research questions. Answering those questions not only fills the research gap between labor-predictability across countries, but also helps both practical investors to make a better international or domestic time-varying investment decisions. Firstly, can labor adjustment costs forecast stock returns by pooling all countries? I provide initial evidence on the link between the labor market and the aggregate international stock markets. If the pooled result of this regression is significant, we can expect a natural tension between country-level returns and labor adjustment cost. I start by running pooled panel regressions with and without control for country fixed effects to see whether the labor markets and future stock market returns are related in global context. The answer is yes. Both with and without country fixed effects regression results show that CSV has predictive power for future returns in short-term and long-term. Secondly, does the predictive power of CSV emerge across countries? I run the country-by-country predictive regression to see the forecasting power of CSV in international markets. My result shows that there exist large variation of predictive power for CSV in different countries in in-sample predictive regressions. Turn to my next empirical test, since Campbell and Thompson (2008) propose the predictive regressions and out-of-sample regressions without parameter restriction will cause underprediction of future returns in US market, I calculate both unrestrictive and restrictive out-of-sample statistics (𝑂𝑂𝑆 𝑅2). However, 𝑂𝑂𝑆 𝑅2𝑠 with restriction or not are almost the same. That means that restrictive model does not improve OOS predictability, at least outside of US market. This finding also consistent with Jordan and Vivian (2011). So I only represent 𝑂𝑂𝑆 𝑅2 without restriction in this paper. Thirdly, compared with forecast ability of alternative predictors, which level of predictive power of CSV is? Eiling et al. (2016) compare the performance of CSV to 11 alternative predictive variables in in-sample and out-of-sample test. They show CSV has the strongest power in US market. My paper uses a different set of alternative variables, including fundamental variables and macro variables. My results show CSV is still a powerful predictor in US market, but other markets tell a different story: CSV is not always the most powerful variable to forecast future market returns. Last and the most innovative, What economic mechanism drives forecast performance of CSV to countries with different characteristics? Then, I begin to examine the economic factors that drive the difference for CSV predictive performance by using pooled panel predictive regression model. The results show unemployment rate is a significant driver for cross-country differences in long-term. More specifically, the higher the

4

unemployment rate, the stronger forecast ability of CSV thus the negative the future stock market returns. In contrast, the growth of unemployment rate does not have the significant impact on in-sample predictive performance of CSV. Thus, I can still doubt that the significant impact of the unemployment rate results from the time when the unemployment rate is high. Besides, whether differences in out-of-sample performance of CSV across countries depend on labor market flexibility is unclear in my results.

The structure of this paper is as follows. Section 2 introduces main theories in the existing literature for the linkage of the labor market and stock market return predictability as well as the evidence of stock return forecasting in international stock markets. Moreover, this section provides hypotheses and comprehensive contributions of this paper. Section 3 describes the methodology and Section 4 provides research data as well as summary statistics. Section 5 is the empirical result of predictive regressions and out-of-sample analysis. Section 6 detailed discuss the economic channel for different CSV performance across countries. Section 7 describes robustness check. The conclusion is in Section 8.

2.

Literature Review and Hypotheses Development

2.1. Labor Market and Stock Market Return Predictability

Actually, the labor market and the capital market are inseparable. The labor market is the driving force of the capital market, and the capital market creates space for the workforce. Labor can also be understood as "hiring" capital. The nature of labor relations is a kind of contract established by the company in this form. The company is actually the middleman of the labor force and the capitalist side, which plays the role of reducing the transaction cost. In other words, labor market can be seen as exogenous shocks which affect the productive abilities of an economy. Thus labor markets frictions have essential implications for financial markets, such as returns for stock. However, prior works linking labor markets with future stock returns seems unbelievable scarce. Chen and Zhang (2011) investigate how time-varying stock market excess return affect the labor market as well as whether labor market variables have predictive power for the stock market. They focus on three labor-related variables, saying, payroll growth, net hiring rate and net job creation in manufacturing. Their results show that high aggregate risk premium has a different impact on above labor-related

5

variables but predictive power practically exists. This means that labor demand and the stock market is inextricably linked. More importantly, payroll growth and net job creation rate in manufacturing can forecast risk premium. Boyd, Hu, and Jagannathan (2005) mainly focused on the impact of the unemployment rate on short-term stock return. They find that increasing unemployment rate has a positive impact on daily stock returns while the economic condition is in the expansion. A possible explanation is that interest rates can be lower in the future if the unemployment rate is rising. Their findings also show the negative impact appears while an economy is in the recession because profit for firms and shareholders tends to devolve with poor labor condition. Other literature is in cross-sectional return case: Belo, Lin, and Bazdresch (2014) suggest that labor market frictions have a significant impact on cross-sectional stock return predictability. When hiring rate is higher, future stock returns for companies are more likely to decrease.

Arnott and Henriksson (1989) is the first study to consider the impact of unemployment level and unit labor costs on international capital markets. Their results indicate that both stock and bond markets react statistically significant stronger to rising unemployment rate. Besides, the higher the unit labor costs, the lower the stock and bond market excess returns in all sample countries (except Canada). Arnott, Li, and Liu (2016) extend the work of Arnott and Henriksson (1989) by using a broader sample and longer-term series. They find that 24- month forward continuously compounded excess returns are higher when the level of unemployment is higher in 22 developed countries. The most enhancement of their study is the attempts to explore the impact of government policy on unemployment statistics. For example, labor laws in some sample countries like Belgium, Ireland, and Spain effectively protect employees from being terminated and offer minimum wage for them, which obviously reduce unemployment and labor cost. Their paper gives me implication to explore the reason for the difference of predictive power of CSV across countries.

Different from Arnott, Li, and Liu (2016) that use unemployment rate to test the international labor conditions and future capital market return, Eiling et al. (2016) pay attention to labor adjustment costs and future stock returns in US market. Based on in-sample and out-of-in-sample predictive regression, this paper shows that labor adjustment costs produced by sectoral reallocation shocks, compared with other alternative predictors including labor-related variables, has strongest predictive power for future stock returns in

6

US. Moreover, similar to Chen and Zhang (2011), they employ production–based asset pricing model (Cochrane, 1991) to enhance empirical results that labor adjustment cost negatively predicts future stock returns. The mechanism behind the predictive power of CSV is that higher dispersion in industry stock returns means a larger difference between the performance of top and bottom industries and an increasing need for labor reallocation across industries. Workers need to spend time and resources to move from low to high performing industries so that one dollar or one euro that firms invest in hiring will generate less workforce. Thus returns on hiring are declining and expected returns on stocks will reduce.

Summing up, the existing studies suggest that labor market and stock market are correlated in some extent, while labor-related variables can forecast future stock returns. This is the case since a healthy economy not only bring reward to stock but also to the labor market. My contribution is exploiting the forecast performance of labor adjustment costs both in developed and developing countries, which can get a full understanding of labor-stock relationship in general and help firms to maximize the profit by managing hiring and investment decisions. The above is just a part of what I want to study, and what I focus on is to explore what causing different predictability of stock markets while using CSV as the predictor. Obviously, labor adjustment costs have a large variation in different countries so that the forecast power of CSV may be very different in different countries. For example, Liosa et al. (2012) document that large differences across OECD countries in labor adjustment cost appear to be related to policies that impede the dismissal of workers. Banker et al. (2013) proposed that cross-country variation in strictness employment protection legislation (EPL) is a reliable exogenous source of variation in labor adjustment costs since EPL implicitly set the limitation for firing cost. So I include EPL as a proxy for labor market flexibility to see how EPL affect CSV performance. This intention can help investors use labor variables to make better asset allocation decisions.

2.2. International Evidence of Stock Market Return Predictability

Much of existing studies in terms of return forecasting focus on the U.S. stock market. Charles et al. (2016) give a clear review of stock return predictability in the non-U.S. market. They use financial ratios like earnings-price ratio and dividend-payout ratio, technical

7

indicators like the change in volume and short-term interest. By including 37 countries, they find that the predictive power of financial ratios is weak, especially in out-of-sample tests, but technical indicators and interest rate have strong predictive power. In my empirical analysis, I mainly focus on the labor-related variables, specifically CSV, as the independent variable to predict future excess stock market while picking financial ratios and technical indicators as alternative predictive variables. However, they use monthly data from 2000 to 2014 which I think is not enough and valid to examine the predictive power of the out-of-sample approach. My out-of-sample period for the out-of-out-of-sample test has minimum 10 years estimation window and 10 years forecast window for 1-month excess return. On the other hand, their work lacks economic explanations among forecast performance of predictors, but the economic channel is always my research area.

Wohar et al. (2005) inspect the predictability of stock returns using macroeconomics variables in twelve industrialized countries and also find that the interest rate stands out both in-sample and out-of-sample across countries. They leave some significant and interesting questions to solve: for example, to uncover whether some macro variables have different forecasting power if the effects of macro variables on stock rewards vary across the stage of the economic cycle. The question can relate to my research for CSV: CSV has strongest predictive power for both unemployment growth and stock market returns in crises time. (Eiling et al. 2016). The effect of reallocation shocks is expected to be weaker during times when labor mobility is high or when the economy is in a good state. (Davis, 1987). Thus, I would like to test whether the predictive power changes as labor market conditions change in each country, especially developed countries with less data issue. Jordan et al. (2014) is the first study to examine the link between country characteristics and out-of-sample predictability performance. They provide evidence that macro and technical predictors have statistically forecasting ability and can produce economical gains to investors. They also find evidence that market development, size, and liquidity relate to the forecast performance.

Above all, by considering alternative predictive variables, I will compare the performance of all variables in terms of both statistical and economic significance. My most innovative point is connecting the labor condition across countries into the context of future stock market returns. My hypotheses are:

8

Hypothesis 1: In general, CSV significantly and negatively predicts aggregate future stock

market excess returns. The reason why I expect the negative relationship is that higher CSV

means higher labor reallocation cost. Stock returns that investors expected to acquired are reduced as the cost of hiring increases.

Hypothesis 2: The predictive power for CSV varies country to country and the differences in

CSV performance depend on labor market conditions. More specifically, on the one hand, I

want to explore how the unemployment condition affects CSV performance. Since labor adjustment cost is a factor behind the forecast power of CSV and unemployment rate may increase with higher labor adjustment costs, I can expect poor labor condition that represented by unemployment condition may have a positive effect on CSV performance. On the other hand, labor market flexibility is also a factor that cannot be ignored to explain differences in labor adjustment costs across countries. In an inflexible labor market, there are many limitations to dismissing and hiring. When the demand for labor redistribution is high, workers will not be easy to leave their jobs, and even if they leave, they will not be easily re-employed by highly profitable industries. So the cost of adjustment of the labor force becomes higher. Thus, I would expect lower future stock returns predicted by sectoral reallocation shocks in inflexible labor markets.

3.

Methodology

3.1. Predictive Regression Model and Out-of-sample Statistical Analysis

In the research of Arnott, Li and Liu (2016), unemployment rate statistically insignificant predict stock returns for individual countries, but the result for pooled panel regressions shows the significant positive impact of the unemployment rate on future stock returns. In this paper, I start by confirming my hypothesis 1 that whether the labor adjustment costs have the negative predictive relationship with the stock market in the whole countries, I perform pool panel regressions with and without the country--fixed effects:

𝑟_{𝑖,𝑡+1:𝑡+𝑘} = 𝛼_𝑖 + 𝛽_𝑖𝐶𝑆𝑉_𝑖,𝑡+ 𝐹𝐸 + 𝜀_{𝑖,𝑡+1:𝑡+𝑘} (1)

Where 𝑟𝑖,𝑡+1 is the continuously compounded excess return form period t to t+1 for all

forward-9

looking returns which equals to 𝑟𝑡+1+ ⋯ + 𝑟𝑡+𝑘 (k=1,3,12,24,and 36 respectively in the

paper). 𝛼𝑖 represents individual effects. When the pool regression without fixed effects,

𝛼_𝑖 = 𝛼 for all countries. 𝜀𝑖,𝑡+1:𝑡+𝑘 is the error term. FE is the country-fixed effects. 𝐶𝑆𝑉𝑖,𝑡 is

the one-month lagged cross-sectional volatility of industry-specific stock returns for country i.

I use monthly industry return in excess risk-free rate as well as monthly excess market return to construct industry-specific CSV. Industry returns are calculated as one-month percentage change of level 3 industry return indices ( Details in the next section ) from Datastream. Firstly, I use past 36 months data to get abnormal returns for each industry by equation as below:

𝑅_𝑖,𝑞 = 𝛼_𝑖+ 𝛽_𝑖𝑅_𝑀,𝑞+ 𝜀_𝑖,𝑞 (2)

Where q=t-35,…,t and 𝑅𝑖,𝑞 is industry excess return for industry i at month q and 𝑅𝑀,𝑞 is

market excess return. According to CAPM, abnormal return for industry i at month s is defined as:

𝐴𝑅𝑖,𝑠= 𝛼̂𝑖+ 𝜀̂𝑖,𝑞 (3)

Where 𝛼̂_𝑖 is the estimate constant term and 𝜀̂_𝑖,𝑞 is fitted residual respectively. Then, CSV as the cross-sectional standard deviation of industry returns at month t is measured by:

𝐶𝑆𝑉_𝑡= [1 𝑁∑ (𝐴𝑅𝑖,𝑡−11:𝑡− 𝐴𝑅̅̅̅̅𝑡−11:𝑡) 2 𝑁−1 𝑛=1 ] 1 2 (4) Where N is the total number of sample industries for each countries and

𝐴𝑅_{𝑖,𝑡−11:𝑡} = ∏𝑡 (1 + 𝐴𝑅_𝑖,𝑠) − 1 𝑞=𝑡−11 (5) 𝐴𝑅 ̅̅̅̅_{𝑡−11:𝑡 =} 1 𝑁∑ 𝐴𝑅𝑖,𝑡−11:𝑡 𝑁 𝑖=1 (6)

It is general that model prediction should be different across countries. In other words, the slope coefficient 𝛽𝑖 cannot be identical across i since equity premium is different for each

country. Besides, Ferson and Harvey (1994) also point out that international CAPM betas vary from 0.4 to 1.3, which means there has a sizeable amount of variation in international stock returns. The pooling can get rid of endogeneity problem and a second order bias of

10

estimator. On the other hand, one may argue the economic meaning for the joint slope coefficient if it is not the common evidence between the labor market and the stock market. In this case, pooled estimation can be seen as a statement of the average predictive relationship in the panel so that it is still helpful to get an overview empirical result based on diversified individual results.

After getting the general and average trend of predictability, I begin to use the univariate regressions to test CSV and seven predictive variables for country-by-country analysis. Specifically, I use traditional regression model and take all the data within a country to perform equation (1), saying, using one month lagged 𝑧𝑡 (Seven predictive variables

introduced in section 4) to predict 𝑟𝑖,𝑡+1:𝑡+𝑘. Vast of prior studies have used gross return and

the reason why I use excess return is that stock market can be influenced by various economic and monetary regimes and using real return over short-term interest rate achieve equal comparison across countries. Besides, lots of studies such as Richardson and Stock (1989) show overlapping observations appear when using multi-period data. In order to deal with this problem, I adjust standard errors by using Newey-West standard errors with k-1 lags other than OLS standard errors. Then, evaluate in-sample performance by looking at the significance of 𝛽𝑖 together with their 𝑅2. Besides, Welch and Goyal (2008) point out that

plenty economic variables which have in-sample predictive power for stock excess returns do not have out-of-sample forecasting ability. Actually, if both in-sample and out-of-sample test are significant, these two tests can be seen as robustness test for the other. Thus, it is necessary to generate the out-of-sample forecasts of the market returns. I use expanding estimation window to produce forecasts of the equity premium. In order to get the most effective analysis, I will perform the out-of-sample test for CSV at each horizon and pay large attention to one horizon where CSV has the most predictive power. Out-of-sample test for alternative predictors will only be implemented in that horizon. We start by constructing the forecast of the k-month excess market return from month t to month t+k:

𝑟̂𝑖,𝑡+1:𝑡+𝑘 = 𝛼̂𝑡+ 𝛽̂𝑡𝑧𝑡 (7)

Where 𝛼̂𝑡 and 𝛽̂𝑡 are the ordinary least squares (OLS) estimations. Campbell and Thompson

(2008) proposed that a reasonable investor could not use the model with a negative sign of as well as negative equity premium. That is to say, while implement restriction on the

out-11

of-sample model, the OOS forecast performance can be improved. Jordan et al. (2014) proved that parameter restriction can use to solve poor performance problem in US market but cannot improve predictability in their European samples. So in order to get more accurate results, I use models with and without restriction. Then, we can use the 𝑂𝑂𝑆 𝑅2 to evaluate out-of-sample performance, while using the historical average excess market return as a benchmark and use data up to month t. Note that I allow for the minimum 10-year training period, 𝑂𝑂𝑆 𝑅2 can be computed as:

𝑅_𝑂𝑂𝑆2 = 1 −∑ (𝑟𝑖,𝑡+1:𝑡+𝑘−𝑟̂𝑖,𝑡+1:𝑡+𝑘) 𝑇−𝑘

𝑡=120

∑𝑇−𝑘𝑡=120(𝑟𝑖,𝑡+1:𝑡+𝑘−𝑟̅𝑖,𝑡+1:𝑡+𝑘 ) (8)

Where 𝑟̅𝑖,𝑡+1:𝑡+𝑘 is the mean of excess market returns, T is the length of the return series.

When 𝑅_𝑂𝑂𝑆2 > 0, the Mean Squared Forecast Error (MSPE) of the predictive regression

model is lower than that of benchmark model. That is to say, variable 𝑧𝑡 on average beats

the historical mean over the sample period. If 𝑅𝑂𝑂𝑆2 < 0, the historical mean is a better

predictor than 𝑧𝑡. For example, if 𝑅𝑂𝑂𝑆2 = −0.2, the MSPE of the regression model is 20%

higher than that of historical average prediction and lead to an underperformance of 20% over the sample period.

3.2. Empirical Test for Labor Market Conditions and Forecast Performance

In order to investigate whether the difference in labor market conditions can explain the differences in CSV predictability across countries, I plan to run fixed-effects, unbalanced panel regression between labor market conditions and predictive performance. My first measurement set of labor market condition is unemployment rate and unemployment growth. As I mentioned before, I will mainly focus on one time horizon where CSV has strongest predictive power so that the first set of two regression models are as following:

r𝑡+1:𝑡+𝑘 = 𝛼 + 𝛽1𝐶𝑆𝑉𝑡+ 𝛽2𝐶𝑆𝑉𝑡∗ 𝑈𝑁𝑡+ 𝛽3𝑈𝑁𝑡+ 𝐹𝐸 + 𝜀 (9)

r𝑡+1:𝑡+𝑘 = 𝛼 + 𝛽1𝐶𝑆𝑉𝑡+ 𝛽2𝐶𝑆𝑉𝑡∗ ∆𝑈𝑁𝑡+ 𝛽3∆𝑈𝑁𝑡+ 𝐹𝐸 + 𝜀 (10)

Where FE is the fixed effect. I include an interaction term between CSV and unemployment rate or unemployment growth. Beta 2 measures how the interaction term affects the left-hand-side variable. Since I expect CSV to forecast future return with a negative sign, hence,

12

if unemployment rate or unemployment growth is associated with more strong return predictability, we would expect beta 2 to be negative.

The second measurement set is labor market flexibility.As Eiling et al. (2016) suggests, low labor mobility across industries add reallocation cost, thus firms need to compensate this cost and their market value may decrease. So I explore whether 𝑂𝑂𝑆 𝑅2 for each country related to labor market flexibility. I use two proxies for labor market flexibility. The first one is OCED’s Employment Protection Legislation ( EPL ) index that used in Simintzi et al. (2014). This measure consists of three aspects: individual dismissal of workers with regular contracts (EPR), additional costs for collective dismissals (EPC), and regulation of temporary contracts (EPT). EPL as a proxy for employment protection is negatively related to labor mobility. The second one is Fraser Institute’s Economic Freedom of the World index (EFW) that used in Freeman et al. (2008). This index also comprises of 6 components: Hiring regulations and minimum wage (category 5Bi), Hiring and firing regulations (category 5Bii), Centralized collective bargaining (category 5Biii), Hours regulations (category 5Biv), Mandated cost of worker dismissal (category 5Bv), and Conscription (category 5Bvi). All of these components rank between 0-10 scale, thus higher value means higher flexibility. Besides, it is necessary to control for other countries’ specific characteristics to examine the relationship because existing studies provide evidence that forecasting power primarily exists in countries with large financial markets or developed markets, such as Rangvid et al. (2011) for the dividend-price ratio. I include proxies for countries’ market development, market size and liquidity. The three proxy are Stock Market Capitalization to Total GDP ratio, Stock Market Capitalization and Turnover ratio (Total Volume to Stock Market Capitalization) respectively. The dependent variable is out-of-sample forecast accuracy of CSV:

𝑅_{𝑂𝑂𝑆,𝑖𝑘}2 = 𝛼 + 𝛽₁𝐸𝑃𝐿_𝑖 (𝐸𝐹𝑊_𝑖) + 𝛽2𝐶𝑜𝑢𝑛𝑡𝑟𝑦𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 + 𝜀 (11)

3.3. Robustness Tests

The most important empirical analysis of this paper is to uncover what drives the cross-country stock return predictability and the main predictor is CSV that proxy for labor adjustment cost. The first step to see why labor adjustment costs can predict future stock

13

market returns in some countries while cannot in other countries is understand the mechanism behind the forecasting power of CSV. As I mentioned before, when workers need to improve working skills to move to higher performing industries, higher performing industries also need time to wait for employees who have enough capability to deal with their works. That means hiring cost of firms produced by labor reallocation will increase, especially for those firms rely on higher skilled workers. Because the higher skill the firm request, the longer time the workers should spend on learning. Then in the robustness tests, I use a measure for the demand of high-skilled workers for each country. It is Educational attainment of employed persons (EDU). It described the demand for the educational level in employment. The methodology is the cross-sectional regressions between OOS performance of CSV and the measurement for the demand of high-skilled workers:

𝑅_{𝑂𝑂𝑆,𝑖𝑘}2 = 𝛼 + 𝛽1𝐸𝐷𝑈𝑖+ 𝐶𝑜𝑢𝑛𝑡𝑟𝑦𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑠 + 𝜀 (12)

4. Data and Descriptive Statistics

My research is mainly focused on cross-country analysis and I intend to study well-diversified stock markets, thus country selection is the first step to start data collection. According to country size, economic development and market flexibility, Morgan Stanley Capital International (MSCI) divided international markets into three types: developed markets, emerging markets and frontier markets. Based on MSCI’s classification, I choose countries if all of the variables are available as well as the time period is enough to get a convincing result of the out-of-sample test. ( e.g., I exclude all countries with less than 20 years of data; this allows for a 10-year estimation period and a minimum of a 10-year forecasting period.) My sample include 21 developed markets: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Hong Kong, Italy, Japan, The Netherland, New Zealand, Norway, Portugal, Singapore, Spain, Sweden, Switzerland, the UK, the United States; 5 developing markets: India, South Korea, Malaysia, Philippines, Thailand. Data in this paper are monthly time series and mainly come from Datastream. For return predictability tests, the dependent variable is always the stock market excess returns over various horizons (1,3,12,24 and 36 month-ahead). It is the return on stocks in local currency over the local risk-free rate for each country. Stock return is computed in continuously compounded form, which is the natural log difference in the stock return index (RI), also the

14

level 1 equity index in Datastream. I use short-term interest rate as risk-free rate and consider the 3-month Treasury bill rate firstly. While the 3-month Treasury bill rate data are not available, I will use money market rate or 3-month interbank rate for such markets. As I mentioned in section 3, CSV is calculated by using level 3 industries returns for each country. There are totally 20 industries in level 3 industry according to the definition of Datastream: Oil & Gas, Chemicals, Basic Resources, Construction & Materials, Industrial Goods & Services, Automobiles & Parts, Food & Beverage, Personal & Household Goods, Healthcare, Retail, Media, Travel & Leisure, Telecommunications, Utilities, Banks, Insurance, Real Estate, Financial Services, Equity Investment Instruments, and Technology.

Seven alternative forecasting variables which are believed to have predictive power are used, including valuation variables, technical variables and a macro variable. The four frequently used valuation variables are log dividend-price ratio, log dividend-yield ratio, log earnings-price ratio, dividend payout ratio.1 _{The two technical variables are price pressure}

and change in volume.2 _{(Sullivan et al., 1999). Price pressure described the extent of market}

aggregate momentum and the sentiment of investors to stock. Change in volume is a proxy for price trend. The short-term interest rate is a macro variable I considered. Those data can be successfully obtained from Datastream.

Table 1 reports the sample time period as well as the mean and the standard deviation for eight predictors in each country. In order to achieve apple-to-apple comparison, I choose sample period which depends on the longest time range of excess stock market returns. For example, stock return and all valuation ratios are available from 1973.2 in Austria but 3-month interbank rate which is a proxy for risk-free rate can only be acquired from 1991.6 so that 1991.6 is the start date for acquiring excess returns to explore Austria market. India has the least time span but 24 years are enough for out–of–sample test. It means that excludes 48 months (4 years) to calculate CSV, there still have totally 20 years to test. Interesting

1 _{The codes of row data in Datastream in parentheses. All raw ratio are reported in percent. Log dividend-price}

ratio is the total dividend amount divided by the market value and I transfer row data (DY) by using ln(DY/100); Log dividend-yield ratio is calculated as dividend-price ratio multiplying price index (PI) at time t dividing by price index (PI) at time t-1, then use natural logarithm; Log Earnings-price ratio is generated by using the natural log of reciprocal of price-earnings ratio (PE); Payout ratio equals to the difference between log dividend-price ratio and log earnings-ratio.

2 _{Price pressure is the ratio of risers to fallers. Numbers of risers (RS) is the number of stocks whose closing}

price is higher than closing price at previous date, while numbers of falls (FS) is the number of stocks whose closing price is lower than before; Change in trading volume is computed as monthly change of turnover by volume (VO).

15

differences for variables in Table 1 appear. It is well known that stock returns are significantly related to the extent of financial development ( Dellas and Hess, 2005). Among all the countries, Hong Kong and Sweden have the highest excess returns which equal to 0.007, in the meanwhile, the standard deviations of excess return for Hong Kong and Sweden are obviously at a higher level. It is logical since investors require the risk premium to compensate risk. In contrast, Thailand experiences the highest stock market volatility since financial development, such as the quality of banking system is at a lower level compared with developed countries. The same trend can also be seen in another developing country – India where the stock market volatility is 0.08. The average of CSV varies substantially across countries from 0.15 (U.S.) to 0.36 (Thailand), which means sectoral reallocation shocks are lower in the US and higher in Thailand. Sectoral shift induces optimal allocation of human capital or resources, while workers need to improve skills if they want to move in high-performing industries. Thus differences in mean of CSV may suggest labor reallocation cost is different across countries. Since time span varies from country to country and labor adjustment cost as a driver for unemployment tends to change over time (Eiling et al., 2016), there should exist economic reasons for the difference, such as different labor policies in the different country and different time period which have the impact on labor mobility. India experienced the higher labor adjustment cost (the second highest CSV) during the sample period while Philippines, as a developing country as India do not see such higher average of CSV. This is consistent with Munshi and Rosenzweig (2016) who show that labor mobility in India is surprisingly low compared with other large developing countries because of a combination of well-functioning rural insurance networks and the absence of formal insurance. Besides, Belgium, during the time span of our study, has prolonged unemployment compensation and minimum wage set which means labor market can be less mobile thus reallocation takes time and resources. This is the case since we can see CSV is much higher in countries like Denmark with the generous social welfare system. After all, this kind of difference in the international labor market and stock market motivates me to uncover the relationship between these two markets into the context of cross-country.

[ Insert Table 1 here ]

Moreover, the monthly unemployment rates (UN) are download from Federal Reserve Economic Data (FRED). Follow Eiling et al.(2016), I transfer unemployment rate into log form,

16

where UN = log ( UN

1−UN). The unemployment growth is calculated as the first difference of

UN. Turn to the two index of labor mobility, EPL is available in OECD. Stat for 20 OCED countries in my sample from 1998-2013. As I mentioned in section 3, EPR and EPC are two measures that describe the ease of dismissal. If constraints on firing are fewer, lower performing industries are more likely fire employees when facing economic shocks. Workers then have a large motivation to move to better performing industries, thus labor adjustment costs increase. The third measure of EPL is EPT which captures constraints on hiring. When hiring becomes difficult, a dollar in hiring a worker generates less workforce and low stock returns can be expected. After all, EPL is lower, the labor adjustment costs are higher. EPW can be accessed from Fraser Institute’s Economic Freedom of the World (“EFW”) database for total 26 countries from 2002 to 2013. 5Bv has the similar function like EPC and EPR that capture the ease of hiring, while 5Bi, 5Bii and 5Biv capture the ease of hiring. 5Biii captures the ability of labor unions that restricts both hiring and firing and 5Bvi measures the extent of intervention in supply-side. Since the regression of labor mobility and forecast accuracy is a cross-sectional regression, the average of this two measure will be used.

Table 2 reports the average of EPL and EFW indicator. EPL is negatively related to labor turnover while EPW is positively related to labor mobility. These two measures do not give me the same rank of labor mobility. For example, South Korea has more flexible labor market in EPL measure, but its labor market is the most inflexible in EPW measure. So I expect differences in the effect of labor market flexibility on CSV performance by using these two measures.

[ Insert Table 2 here]

For the data in robustness test, OCED.Stat provides the world indicator of skills for employment. Among all the indicators, Skill requirement indicators described the extent that a country depends on the skilled workers. Since I want to uncover whether the better CSV performance in countries with higher skilled workers, I adopt one sub-indicators of Skill requirement indicators. It is Educational attainment of employed persons. This indicator is classified as 6 components: no schooling, primary, lower secondary, upper secondary, tertiary first stage, tertiary second stage. A country relies on workers with higher education,

17

that country may face more labor adjustment costs. So I choose the indicator: Educational attainment of employed persons with tertiary second stage. This indicator is available for 17 countries in my sample from 2009 to 2013. Besides, this indicator is in percentage forms that represent how much percent a country required for high-skilled workers.

5. Empirical Results of Predictability Test

5.1. Pooled Panel Regression

The evidence of stock return predictability is very mixed and most studies only pay attention to one time-series at the market level. But, pooling all data can bring a more powerful result because the portion of the stock can be predicted is relatively small to the overall variance in the stock returns. So a part of the theoretical test of this paper is to estimate the pooled slope coefficient in equation (1). That is, by pooling monthly observations of excess returns and CSV from all countries, the estimation of joint coefficient β is produced. The results for pooled panel regression are reported in Table 3. Panel A shows the significant impact of CSV on future market returns in all time horizons without country-fixed effect. For example, the coefficient of CSV is -0.213 when k=12, which means one percent increase in CSV decreases next 1-year stock market excess return by 21.3 percent. After controlling for the country-fixed effect, the predictive relationship is still significant and we can see the coefficient and R-squared increased dramatically. For example, the coefficient for CSV jumped to -0.394 when k=36 and R-squared increased from 1.8% to 4.6%. The result of pooled panel regression shows that the link between labor adjustment costs and stock returns is material, thus I begin to run univariate time-series regressions for individual countries by using CSV and alternative predictors and expect the difference in forecast power of CSV.

[ Insert Table 3 here]

5.2. In-sample and Out-of-sample Return Predictability

Table 4 provides the results of in-sample (INS) and out-of-sample for 1, 3, 12, 24, and 36-month horizon. I perform the out-of-sample test for CSV at each horizon so that I can get a more robust result.In other words, I can effectively observe the horizon in which the stock returns are most predictable, then put more emphases on that horizon. Since the number of

18

stock markets that can be predicted in terms of CSV is the largest when k=12, the out-of-sample results are only available for alternative predictors when k=12. Turn to table 4, for each predictor, the first column is the in-sample predictive coefficient with Newey-West (1987) standard errors, the second column is in-sample R^2 and if any, the third column is OOS R^2. First focusing on in-sample results, the predictive power varies from different countries and time periods as hypothesis 2. The results suggest CSV has enduring and significant performance in US stock market from k=3 to k=36, while results in Eiling et al. (2016) show CSV has predictive power for all five time horizons. The reason can be that industry returns used to construct CSV in this paper are different from returns used in Eiling et al. (2016). They use 49 industry portfolio returns from 1952 to 2013, while I use level 3 industry returns from 1973 to 2017. But the out-of-sample OOS R^2 is positive for CSV in US stock market when k=3. Valuation predictors do not have in-sample predictive power under Newey-West standard errors in U.S. over the five horizons even vast studies give significant evidence for those valuation variables. Actually, Paye and Timmermann (2006) proposed that from 1990s DP ratio has less ability to forecast stock return because of model uncertainty. Kellard et al. (2010) compared the performance of DP ratio in US and UK, pointing out the weaker performance in US is due to larger evaporation of dividends. My results also suggest dividend-related ratio has both in-sample and out-of-sample forecast ability in UK. Besides, for other alternative predictors in US, the out-of-sample test cannot reject the null-hypothesis expect that price pressure and short-term interest rate have positive OOS R^2, 0.38% and 2.21% respectively. CVM can predict future stock returns with 5% significance level over longer time horizons (e.g. 24 and 36 months ahead). Then, I start analysis this empirical results from Australia stock market. CSV significantly and negatively predictive 1-month and 3-month ahead stock returns both in INS and OOS forecast test, in the meanwhile, only DP ratio and DY ratio have forecast power among other predictors but they do not outperform the historical average. When k=24, even there do not have INS

evidence, CSV in Australia outperforms the benchmark and the 𝑂𝑂𝑆 𝑅2 = 0.09%. INS and

OOS tests give similar evidence for CSV in Austria for every horizon: INS coefficients are all significant at least at 5% significance level with INS 𝑅2 and positive OOS 𝑅2 increased gradually. Despite upward bias in the 𝑅2 appears in overlapping horizons, the predictive power of CSV is material in Austria. Belgium, New Zealand, Norway and Singapore are markets where no INS evidence appears and OOS forecast accuracy is only positive at one

19

horizon. ( k=3 for Belgium, k=36 for New Zealand and Norway, and k= 12 for Singapore). Besides, valuation predictors have apparent forecast ability in Singapore that can be

observed from positive OOS 𝑅2 when k=12. Canada, Finland and Germany are three

countries with INS and OOS forecast power for CSV at the short horizons, while Finland has higher OOS forecast accuracy in longer horizon. Besides, among all alternative predictors, the risk-free rate is the most significant variable in these countries. CSV exhibit stronger predictive ability in Denmark and France when k=12 and 24. For example, the out-of-sample

𝑅2 _{equals to 13.19% in France when k=12, which indicates that cumulative squared error of}

the model is 13.19% lower than that of historical average prediction. INS 𝑅2 of CSV range

from 0.4% (k=1) to 27.8% (k=12) of CSV for HK, while OOS 𝑅2 of CSV is negative when k=12

and k=24. Interestingly, no out-of-sample significance of CSV can be observed in Spain, Sweden, and Switzerland, while INS test alludes that CSV can predict returns in longer time horizon. Other than valuation variables, CSV, technical variables and risk-free rate have no predictive power in Japan. The same trend can be seen in Italy where CSV can only predict future returns when k=36. For rest developed countries - Netherlands and UK, CSV shows stronger forecast ability, especially at one-year horizon. One thing needs to mention is that CSV is negative related to future return in all predictable countries other than Netherlands. Turn to developing countries, CSV has both INS and OOS evidence in Thailand, while CSV in other developing countries has the less predictive ability.

After all, there is no same trend within developing countries and developed countries I have researched now, nor same trend in European countries. CSV always negatively predict future stock returns for each country and significant INS coefficients range from -1.315 in US to -0.22 in Canada. The results give implications for investors to allocate their assets. On the other hand, except US and Austria, CSV does not have comparable strength with alternative variables in other stock markets. The most reasonable and general driver for the difference is different labor conditions caused by labor policies. Thus, the next step is to econometrically relate labor condition to CSV performance.

20

6. Economic Channel for Different CSV Performance

6.1. INS CSV Performance and Unemployment Condition

At the first place, I provide a more general picture of the relationship between the unemployment rate ( unemployment growth) and the country’s stock market predictability. The unemployment rate is available for 21 countries in my sample and it cannot be acquired in India, Malaysia, Philippines, Singapore and Thailand. Figure 1 shows the relationship between the one-year horizon predictive coefficient of CSV and unemployment conditions for 21 available countries, while graph A focuses on the average of the unemployment rate and graph B is about the average of unemployment growth. The point I want to make here is to see where the unemployment rate or unemployment growth is higher, the predictive power of CSV is stronger. The empirical results of Eiling et al. (2016) suggest that CSV negatively predicts future returns of low to high performing industries in US, while CSV has more forecast power in top performing industries that facing higher adjustment costs. On the other hand, higher adjustment costs may lead to higher future unemployment because workers were trained and did not work. As the same logical, my global investigation is that aggregate stock returns are more predictable by CSV in countries with higher unemployment rate or unemployment growth. If above analysis is true, I expect a negative relationship between unemployment condition and the negative size of the predictive coefficients. This is indeed what both graph A and graph B reveal: unemployment rate and unemployment growth negatively related to forecast coefficients, while the correlation coefficients are -0.09 and -0.33 respectively. In other words, the higher the unemployment rate (unemployment growth), the more negative the predictive magnitude. However, these two correlation coefficients are not statistically significant. Especially the coefficient of the unemployment rate tends to be 0. So it is necessary to turn to more empirical tests.

[ Insert Figure 1 here ]

Motivated by Rangvid et al. (2011) who is the first test the link between characteristics and in-sample predictability for the dividend, I adopt the same methodology to test labor condition by using unemployment rate and unemployment growth, that is the fixed-effects, unbalanced predictive panel regressions for excess returns. I show the results in Table 5.

21

Panel A shows the results in which I use the unemployment rate as the measure of labor condition in the country. Panel B shows the results from using unemployment growth within a country. For each independent variables, the upper value is coefficient and the lower value is the standard deviation. It is interesting to see the empirical results: Panel A indicates higher unemployment rate with a stronger predictive power of CSV when k=1, 24 and 36 (the interaction terms are negative). Besides, statistically significant interaction terms can be observed when k=24 and 36. For example, the interactive coefficient of CSV and unemployment rate is -0.470 when k=24. It means one percent increase in unemployment rate lead to additional 47 percent decrease in one-year horizon stock market returns that predicted by CSV. In contrast to the clear effect of the unemployment rate on CSV performance, Panel B confirms my hypothesis only in short horizons without significant coefficients. From the degree of fit of the model, 𝑅2 of the model with the

unemployment rate is almost three times the 𝑅2 of model with unemployment growth. In

other words, the model with unemployment rate performs econometrically better than the model with the growth of unemployment rate. In addition to concerned about how the unemployment conditions affect the CSV performance, my empirical tests also show unemployment conditions are positively related to future stock market returns. It is in the agreement of the result of Arnott, Li, and Liu (2016) that stock returns are weak when labor conditions are strong.

[ Insert Table 5 here ]

After all, these tests to investigate whether differences in unemployment conditions can explain the differences in stock return predictability across countries do not give us a strong evidence. Only the unemployment rate associated with stronger CSV performance in longer horizons, while the growth of unemployment rate shows no influence on the predictive power of CSV. Combing the results of two measurements, I am more inclined to doubt that the impact of the unemployment rate on CSV performance is the result of time series rather than the differences between countries. That is to say, the significant coefficients of the interactive term are from the time when the unemployment rate is high rather than from where the unemployment rate is high.

22

6.2. OOS CSV Performance and Labor Market Flexibility

After uncovering the role of unemployment condition in cross-country differences of INS performance, I conclude that CSV has stronger power in higher unemployment countries or at times when unemployment is higher. This section examines how labor market flexibility affects the relationship between CSV OOS performance and future stock markets. This test is a cross-sectional test. As I described in section 3, I employ two measures of labor market flexibility. Figure 2 provides a general picture of how EPL (EFW) relates to out-of-sample forecast accuracy when k=12. The correlation coefficients tell us that EPL which negatively represent labor market flexibility is associate with lower out-of-sample performance. Besides, higher EFW (higher labor mobility) implies higher out-of-sample performance. If this is a true story, my hypothesis that lower labor flexibility is related to higher CSV performance is wrong. However, the correlation coefficients are not significant and they are both close to 0. As I mentioned before, positive OOS R-squared means that CSV beats the historical mean to predict future stock market returns. After carefully observing figure 2, I found the fact that positive 𝑂𝑂𝑆 𝑅2s are more likely to concentrate in the place with middle value of EPL and EFW. As a result, I can expect the effect of labor market flexibility on CSV is non-linear. In an extremely inflexible market, workers are still able to get minimum wage in the face of an economic recession or a worse performance of their company. Hiring also faces a series of restrictions. For instance, the job type and duration of fixed-term and temporary employment contracts have strict rules. So the need of labor reallocation can be relatively small in an inflexible labor market, even no sectoral reallocation shocks to produce the labor adjustment cost. In such circumstances, firms will not easily dismiss employees and recruit staff. So the company’s recruitment decision and behavior may have a very small influence on company’s productivity, shareholders will naturally not consider the impact of human resources costs on future stock returns. In contrast, extremely flexible market means the higher capability of changing according to needs and staying operational in reaction to changes in economic circumstances. If economic shocks happen, workers are eager to move into top performing industries without restriction. Even workers should cost time and other resources to get new skill to move into higher performing industries, firms in very flexible labor markets do not have many restrictions on hiring. For example, without restriction on working contract, firms can only sign short-term temporary contracts with employees or pay

23

the lower salary and allowances. So that firms in such labor market invest one capital in hiring will get the same or more workforce than in less flexible markets. That is to say, labor adjustment cost will not become a factor that leads to the decrease in stock rewards.

[ Insert Figure 2 here ]

Then, I use OLS regressions to get the empirical results of the relationship between labor market flexibility and OOS CSV performance. I still use EPL and EFW in a country to explore whether differences in labor market flexibility can explain the difference in return predictability. I also control for three other country characteristics: stock market development (MV/GDP), log equity market capitalization (Log_MV) and turnover (VOL/MV). This paper is the first to link the OOS performance of predictor to labor market conditions, and also the first to control for other country characteristics. Prior studies have implied that in-sample predictability is more likely to exist in large financial markets or developed markets. So the reason why I control for other country characteristics is that they may have influences on return predictability. The results are reported in Table 5. In panel A, EPL shows no statistically significant relationship with OOS performance of CSV in all horizons except when k=36. The largest and statistically significant coefficient of EPL is -0.13 when k=36. R-squared is also the largest when k =36, that is 29.9%. Since higher EPL captures lower labor market flexibility, Panel A shows labor mobility only plays a slight significant but positive role in CSV performance in 3-year aggregate excess returns. However, panel B does not give a consistent evidence with panel A. EFW which positively represent labor market flexibility is not statistically significant at any significance level, but it has the positive effect on the forecast ability of CSV when k=1, 3, 12 and 24, while it has the negative effect when k=36. There are two possible explanations of the difference between panel A and panel B: one is as I mentioned before, the components that consist of these two measures are different. EPL pays attention to regulations of hiring and firing, while EFW, in addition, focuses on regulatory intervention to the supply-side. The other explanation is sample countries are all OCED countries in panel A but panel B includes both OCED and non-OCED countries. Presumably, it is because of the diversity of the labor and the stock market in OCED and non-OCED countries that lead to inconsistent evidence of the role of labor market flexibility in return predictability. So panel C shows the result in which I use EFW in OCED countries to re-examine the effect of labor market flexibility. The negative sign of EFW when k=36 in

24

panel B becomes positive in panel C and R-squared at all horizons are doubled. As a consequence, the differences in panel A and panel B are mainly because of the different sample.

[ Insert Table 6 here ]

Summing up, the cross-sectional tests do not give us a significant linear relationship between labor market flexibility and out-of-sample R-squared of CSV. It is also inconsistent with my hypothesis that CSV can show higher predictive ability in the less flexible labor market. I have two explanations for this result. The first one is I described before: labor adjustment costs in particularly flexible or inflexible labor markets can be relatively small, thus the costs do not have an important role in firm’s stock returns. The second one is my measures for labor market flexibility are not accurate captures how labor mobility affects the labor adjustment costs. For instance, OOS R-squared is calculated by using a longer time series data but the two proxy for labor market flexibility are only available for very short time period. Synchronization on the time interval may lead to deviations in the prediction results.

7. Robustness Tests

In my main empirical analyses, I find a large variation in the forecast power of CSV across countries. Since labor adjustment cost will affect the company’s hiring productivity, the market value can be underestimated by their shareholders. The expected stock rewards of that company will decrease. Similarly, if one country has a higher labor adjustment cost, then CSV will play an important role in future stock market returns. Reallocation cost is due to the fact that workers spend time and resources to get more skill. For instance, people cost long time to get a higher educational degree. So I can logically expect labor adjustment costs can strongly predict future stock market returns in countries that depend heavily on higher-skilled workers. More specifically, I use the indicator that represents how much percentage the countries require for advanced educational employees to proxy for the need of high-skilled workers. Table 7 reports the result of robustness test. The slope coefficients of Educational attainment of employed persons with tertiary second stage are all positively related to OOS performance of CSV. That means if labor adjustment costs in one country

25

have a strong predictive relationship with future stock market returns, then this country may have a labor market that relies heavily on highly educated workers. However, even the magnitude of the coefficient and the R-squared are relatively large, the coefficients are not significant at any significance level. From the perspective of econometrics, my robust test and even the above test connecting labor market flexibility and OOS R-squared of CSV all suffer small sample problems. Also, we should keep in mind that the variable I choose to proxy for the demand for high-skilled workers is too narrow and cannot describe the overall situation. Unfortunately, there are no other available indicators that can accurately measure the demand for high-skilled workers across countries. Besides, there are no efficient data with such a long history as the stock returns.

[ Insert Table 7 here ]

8. Conclusion

This paper is dedicated to explore the predictive power of CSV for stock returns of international markets. CSV is a proxy for labor adjustment costs that produced by sectoral shift, so my main empirical test is to see how labor adjustment costs affect future stock market returns. The reason why CSV may have predictive ability can be simplified as different substitution effect across countries: Fewer shareholders expect to earn profit from investments if their firm needs to pay more for hiring. I use the most recent data for total 26 countries around the world ( 20 developing markets and 6 emerging markets). Besides, I also adopt 7 well-known alternative variables. They can be classified as fundamental ratios, macro variable and technical variables. By using the traditional predictive regression model, I found a consistent evidence with prior studies that CSV has the strongest forecast power compared with alternative predictors in US market. However, in some countries, CSV does not show a strong predictive power, even no predictive power. Pick the predictive regression on one-year-ahead stock returns as an example, the forecasting power of CSV is significant at 1% significance level under Newey-west standard errors for Spain, French, Hongkong, Austria, Switzerland, UK, and the US, while there is a slight significance for Finland The Netherland, and Thailand. For other countries in my sample, no significant slope coefficients of CSV appear. Alternative variables that confirmed to have forecasting power in