Economic Policy Uncertainty and Momentum

Economic Policy Uncertainty and Momentum Ming Gu School of Economics and WISE Xiamen University guming@xmu.edu.cn Minxing Sun Department of Finance University of Memphis msun@memphis.edu Yangru Wu Rutgers Business School Newark and New Brunswick Rutgers University yangruwu@business.rutgers.edu Weike Xu Department of Finance Clemson University weikex@clemson.edu June 9, 2017 Abstract Using a news-based measure of economic policy uncertainty (EPU), we demonstrate that EPU negatively forecasts momentum profits. An increase of one standard deviation in EPU is associated with a 1.13% decrease in returns for the winner-minus-loser portfolio. The effect of EPU on momentum is mainly driven by the short side portfolio. The predictive power of EPU on momentum payoffs is robust after controlling for market states, business cycle, market volatility, investor sentiment, market illiquidity, return dispersion and time-varying risk factors. Furthermore, a global EPU index forecasts momentum profitability for the international equity market and other asset classes. Our results suggest that economic policy uncertainty is an important determinant of time-series variations in momentum profits. Keywords: momentum; economic policy uncertainty; time-series variation of momentum

Economic Policy Uncertainty and Momentum June 9, 2017 Abstract Using a news-based measure of economic policy uncertainty (EPU), we demonstrate that EPU negatively forecasts momentum profits. An increase of one standard deviation in EPU is associated with a 1.13% decrease in returns for the winner-minus-loser portfolio. The effect of EPU on momentum is mainly driven by the short side portfolio. The predictive power of EPU on momentum payoffs is robust after controlling for market states, business cycle, market volatility, investor sentiment, market illiquidity, return dispersion and time-varying risk factors. Furthermore, a global EPU index forecasts momentum profitability for the international equity market and other asset classes. Our results suggest that economic policy uncertainty is an important determinant of time-series variations in momentum profits. Keywords: momentum; economic policy uncertainty; time-series variation of momentum

Introduction Momentum is the most robust and well-known anomaly in the finance literature. Jegadeesh and Titman (1993) document that stocks with high returns over past 3 to 12 month have abnormally high average returns over the next 3 to 12 month. This return pattern is robust within different size groups (Fama and French, 2008) and significant in major stock markets around the world (Rouwenhorst, 1998; Griffin, Ji and Martin, 2003; Chui, Titman and Wei, 2010). It has also been shown that momentum profits are dependent on several state of economy variables including business cycle (Chordia and Shivakumar, 2002), past market returns (Cooper, Gutierrez and Hameed, 2004), investor sentiment (Antoniou, Doukas and Subrahmanyam, 2013), market volatility (Wang and Xu, 2015), market illiquidity (Avramov, Cheng and Hameed, 2015) and return dispersion (Stivers and Sun, 2010). Recently, Baker, Bloom and Davis (2016) develop a new index of economic policy uncertainty (hereafter, EPU) based on newspaper coverage frequency. They find that EPU increases stock price volatility and decreases investment and employment in policy-sensitive sectors. Many studies show that EPU has the unique and significant time-series effects on financial markets and corporate operations. 1 For example, Pastor and Veronesi (2013) demonstrate that political uncertainty proxied by the EPU index, commands an equity risk premium, especially during the bad economy state. Brogaard and Detzel (2015) also employ the EPU index to measure the policy uncertainty, and argue that EPU is an economically important risk factor for equities. Thus, we raise a natural question whether 1 Other studies have linked policy uncertainty to stock prices (Pastor and Veronesi, 2012), daily jumps in stock and bond markets (Baker, Bloom and Davis, 2013), aggregate bank credit growth (Bordo, Duca and Koch, 2016), merger and acquisition activities (Bonaime, Gulen and Ion, 2016; Nguyen and Phan, 2016), corporate credit spreads (Kaviani, et. al, 2016), credit default swap spreads and liquidity (Wang, Xu and Zhong, 2016), mutual fund flow-performance sensitivity (Starks and Sun, 2016), the term structure of nominal interest rates (Leippold and Matthys, 2015) and bond option implied volatility (Ulrich, 2012). 1

momentum payoffs are related to EPU. We are interested in whether EPU have a unique power to predict time-series variations in momentum profits. At the same time, we are also interested in understanding how EPU will affect momentum returns. Lou (2012) proposes the fund flow-based mechanism to understand price momentum. He argues that past winning funds receive capital inflows and expand their existing holdings (mainly in winning stocks); at the same time, past losing funds lose capital and have to liquidate their holdings (mainly in losing stocks). As a result, performance-chasing mutual fund flows can lead past winning stocks to keep outperforming past losing stocks. Meanwhile, several recent studies show that investors infer mutual fund manager ability from signals of fund performance. Such learning in turn affects fund flow-performance sensitivity (e.g., Berk and Green, 2004; Pastor and Stambaugh, 2012; Starks and Sun, 2016). In particular, Starks and Sun (2016) employ EPU index as proxy for policy uncertainty, and show that investors learning about signals of fund performance weakens when uncertainty increases. Thus, investors have more difficulty moving their investments to the mutual fund manager with superior return generating ability during periods of higher uncertainty. In other word, investors may distribute their investments randomly to the mutual funds regardless of their past performance in the case of high policy uncertainty. Therefore, we expect that EPU may have the negative effect on momentum. Specifically, when EPU is low, the flow-based mechanism that winning funds keep investing in past winner stocks and losing funds liquidate their holdings in past loser stocks continues to work, and generate the momentum; when EPU is high, the fund flow-based mechanism may become disfunctional, and then momentum profits decline or even disappear. Our findings can be summarized as follows. First, EPU negatively forecasts momentum profits. An increase of one standard deviation in EPU is associated with a 1.13% decrease in 2

returns for the winner-minus-loser portfolio. Specifically, the average monthly return in the low EPU period is 1.67% (t-statistic=3.55), whereas the average momentum payoff in the high EPU period is -0.18% (t-statistic=-0.31). The difference in returns between the low- and high-epu periods is economically large and statistical significant. Second, the time-series predictive effect of EPU on momentum profits is mainly driven by the short side portfolio. For instance, the difference in monthly raw returns between the lowand high-epu periods is -2.44 % (t-statistic=-2.82) for the short side portfolio, compared with - 0.58 % (t-statistic=-0.92) for the long side portfolio. Third, we demonstrate that EPU has a unique power to predict momentum profits. The forecast power of EPU on momentum profits is robust after controlling other time-series variables, including business cycle, market states, investor sentiment, market volatility, market illiquidity, cross-sectional return dispersion, and time-varying factors. More importantly, the time-series regressions results show that EPU partially subsumes the explanatory power of selected state variables. Fourth, the EPU mimicking portfolio-ump alone performs well as compared to the Fama and French three factors in terms of a higher R-squared and a lower intercept, further supporting that EPU has a strong power to explain momentum. We then decompose the EPU index into three components, and find that the forecast power of momentum profits is mainly attributed to the news-based component and the tax-related component. The inflation and government spending component has little predictive power for momentum payoffs. Moreover, after controlling for the macroeconomic uncertainty factor from Jurado, Ludvigson and Ng (2015), we show that the effect of EPU on momentum remains unchanged, which implies that policy 3

uncertainty rather than economic uncertainty mainly drives the forecast power of EPU on momentum. Finally, using a global EPU index developed by Baker, Bloom and Davis (2016), we find evidence that EPU can predict momentum profits in the international equity market and many other asset classes. Additional robustness checks show that the effect of EPU on momentum is robust after controlling for firm size, institutional ownership and analyst coverage. Our paper makes several contributions to the literature. First, our study provides a new time-series explanation for momentum. We demonstrate that the predictive power of EPU on momentum profits is unaffected after controlling for previously documented state variables and time-varying risk factors and robust in the global equity market and other asset classes. Second, we shed light on the literature of how policy uncertainty impacts asset prices. Pastor and Veronesi (2013) show that policy uncertainty commands equity risk premium, especially during the bad economy state. Brogaard and Detzel (2015) document that policy uncertainty positively forecasts equity risk premium. Our paper differs from other studies by examining whether EPU can explain variations in profitability of stock price momentum. In addition, Addoum et al. (2015) argue that momentum payoffs are concentrated among politically sensitive stocks and industries. Our analysis differs from that of Addoum et al. (2015) in that we focus on a time-series determinant of momentum. The rest of this paper is organized as follows. In Section 2, we describe the data and summary statistics. We investigate the predictive effect of EPU on momentum profits in Section 3. Further analyses including EPU mimicking portfolios, decomposition of EPU and macroeconomic uncertainty are discussed in Section 4 and additional robustness checks are conducted in Section 5. We provide a plausible explanation and conclusion remarks in Section 6. 4

2. Data and Descriptive Statistics 2.1 Measuring economic policy uncertainty EPU index is a monthly news-based measure of economic policy uncertainty, developed by Baker, Bloom and Davis (2016) and calculated as a weighted average of three components. The first component is a normalized index of the volume of news articles discussing economic policy uncertainty in 10 large newspapers. An article is considered as a policy uncertainty article as long as it contains at least one of the terms uncertainty or uncertain, at least one of the terms economic or economy and at least one of the terms congress, legislation, white house, regulation, federal reserve, or deficit. Each month, the total number of policy uncertainty articles is normalized by the total number of articles in that newspaper. The second component of the index is based on the present value of future scheduled tax code expirations using data from the Congressional Budget Office. The third component of the index is based on disagreement among professional forecasters over future government purchases and consumer prices. It utilizes data in the Federal Reserve Bank of Philadelphia's Survey of Professional Forecasters to measure the forecast dispersion for Consumer Price Index, Federal Expenditures and State and Local Expenditures. The overall EPU index is obtained by applying the weights 1/2, 1/6 and 1/3 respectively to the above three components. The EPU index has been used widely in academic studies and carried by several commercial data providers. 2 2.2 The sample and momentum portfolio 2 Amengual and Xu (2014), Baker, Bloom and Davis (2013), Bloom (2014), Bordo, Duca and Koch (2016), Pastor and Veronesi (2012), Starks and Sun (2016), Stock and Watson (2012) and Ulrich (2012). We download the EPU index from the Economic Policy Uncertainty website. For more details regarding the EPU index, please refer to http://www.policyuncertainty.com/. In addition, commercial provider such as Bloomberg, FRED, Haver and Reuters, carries the EPU index. 5

Our sample consists of all common stocks (with share code of 10 or 11) listed on NYSE, AMEX and NASDAQ obtained from the Center for Research in Security Prices (CRSP). The sample period is from February 1985 to December 2014. We exclude stocks with price lower than $5 at the beginning of portfolio formation period. Stock returns are adjusted for delisting by using delisting return from CRSP. 3 To form momentum portfolios, we sort stocks based on their cumulative return from month t-7 to month t-2 into ten portfolios, following previous literature (Jegadeesh and Titman, 1993). We skip one month between the end of the ranking period and the start of the holding period to avoid the short-term reversals effect (Jegadeesh, 1990; and Lehmann, 1990). The decile portfolio breakpoints are determined by sorting momentum using NYSE firms. To calculate the momentum strategy returns, we long the winner portfolio, short the loser portfolio and hold the portfolios for one month. Then we calculate the value-weighted average excess returns and the risk-adjusted returns under the Fama-French three factor model for each portfolio. < Table 1 > Panel A of Table 1 presents the monthly average returns of the momentum strategy in the full sample. For the monthly average excess returns, the long portfolio (W) earns 0.52 percent and the short portfolio (L) yields -0.23 percent; the Long-Short (WML) portfolio earns 0.74 percent with the t-statistic of 1.94. For the Fama-French alpha, the winner portfolio yields -0.13 percent, the loser portfolio yields -1.26 percent; the Long-Short (WML) portfolio earns 1.12 percent and the t-statistic of 3.07. As can be seen in Panel A, the momentum strategy earns a significantly positive abnormal return from 1985 to 2014. 3 Following Shumway (1997) and Shumway and Warther (1999), delisting return is -55% if trading on Nasdaq or - 30% if on NYSE/Amex when delisting is for performance-related reasons. 6

In Panel B of Table 1, we present the correlations between WML portfolio returns at month t and the EPU index at month t-1. EPU is significantly negatively correlated with WML portfolio returns, with a correlation of -0.16, implying that momentum payoffs are low when EPU is high. In addition, we report the correlations between the WML portfolio returns at month t and other state variables at month t-1. These variables have been proposed to predict timeseries variations in momentum payoffs in the previous literature, including UP market dummy (Cooper, Gutierrez and Hameed, 2004), market volatility (Wang and Xu, 2015), business cycle (Chordia and Shivakumar, 2002), investor sentiment (Antoniou, Doukas and Subrahmanyam, 2013), market illiquidity (Avramov, Cheng and Hameed, 2015) and return dispersion (Stivers and Sun, 2010). We define the UP market dummy variable as one if the prior two-year cumulative market return is non-negative. Consistent with Cooper, Gutierrez and Hameed (2004), the correlation between WML and UP market dummy is significantly positive (0.14), suggesting that momentum strategy is stronger following UP market. The market volatility is defined as the lagged 12-month daily return standard deviation. Wang and Xu (2015) demonstrate that the momentum strategy returns are lower following high market volatility period. As can be seen, WML is significantly negatively correlated with market volatility (-0.13), confirming findings in Wang and Xu (2015). The NBER recession dummy variable is equal to one if periods represent recession. Chordia and Shivakumar (2002) show that momentum profits tend to be lower following recessions. The correlation between WML and NBER recession is -0.09, consistent with Chordia and Shivakumar (2002). We also find evidence that WML is positively related to Baker and Wrugler (2006) s investor sentiment index, consistent with Antoniou, Doukas and Subrahmanyam (2013) that momentum profits are higher when investors are optimistic. Finally, 7

consistent with Stivers and Sun (2010), we demonstrate that momentum payoffs are negatively correlated with lagged 3-month moving average of return dispersion 4. In addition, these state of economy variables are correlated, with correlations ranging from -0.48 to 0.40. In sum, we document that the univariate correlation between WML and EPU is significantly negative, implying that EPU may negatively predict the momentum payoffs. However, it is critical to evaluate whether EPU has a distinctive power to forecast momentum profits. We will show that the effect of EPU on momentum is robust after controlling for other state variables in the later section. 3. Main Results In this section, we provide a detailed analysis of how EPU forecasts momentum profitability. First, we employ the portfolio-sort analysis and investigate momentum portfolio returns and the Fama and French alphas among high and low EPU periods. Second, we conduct time-series regression and Fama-Macbeth regression analyses. Third, we challenge our results by controlling for other time series determinants of momentum profits and time-varying risk factors. 3.1 Portfolio sort analysis We classify the whole sample periods into the high and low EPU months. High EPU months are those in which the values of the EPU index in the previous month are above the median value for the sample period and low EPU months are those with below the median values. Then we compute the average excess returns and risk-adjusted returns for the low and high EPU 4 The correlation between WML and market illiquidity is insignificantly positive, which is inconsistent with Avramov, Cheng and Hameed (2015). The reason is that we use a different sample period. When we use the same sample period from 1925 to 2014 in Avramov, Cheng and Hameed (2015), we confirm a significantly negative correlation between WML and market illiquidity. 8

months. Following Stambaugh, Yu and Yuan (2012), the Fama-French alphas in the low and high EPU periods are estimated using the following regression: R t = β H α H,t + β L α L,t + β 1 MKT t + β 2 SMB t + β 3 HML t + ε t (1) where α H,t and α L,t are dummy variables that identify the high and low EPU periods. Specifically, α H,t (α L,t ) is equal to 1 if the value of the EPU index in the previous month is above (below) the median value. R t is the excess return at month t on either the long portfolio, the short portfolio or the long-short portfolio. MKT t is the value-weighted market excess return at month t, SMB t is return spread between low and high stocks at month t. HML t is the return spread between high and low value stocks at month t. The results for the average excess returns and risk-adjusted returns are presented in Panel A and Panel B of Table 2, respectively. < Table 2 > Results in Table 2 reveal that momentum profits are only significant following low EPU period. In Panel A, among the low EPU period, the average hedge portfolio return is 1.67 percent per month with a t-statistic of 3.55, which is statistically significant at the 1% level. Among the high EPU period, the mean return for WML portfolio is -0.18 percent per month with a t-statistic of -0.31. The return difference between the low and high EPU months is 1.85 percent per month with a t-statistic of 2.45, which is statistically significant at the 5% level. We find a similar pattern for the Fama-French alpha. In Panel B, among the low EPU months, the risk-adjusted return of WML portfolio is 1.97 percent per month with a t-statistic of 3.99. Among the high EPU months, the Fama-French alpha for WML portfolio is 0.26% and insignificant. The difference in the risk-adjusted returns between the low and high EPU months is 1.71 percent per month with a t-statistic of 2.35. 9

Table 2 shows that the predictive power of EPU on momentum profits is mainly driven by the short side portfolio. In Panel A, the mean return difference between the low and high EPU periods is insignificant for the long side, whereas the difference between two groups being -2.49 percent per month (t-statistic=-2.82) for the short side. In Panel B, the difference in risk-adjusted returns between the low- and high-epu months is 0.53% (t-statistic=1.64) for the long side portfolio, while the difference between two groups being -1.17% per month (t-statistic=-2.34) for the short side portfolio. The short side accounts for about 68% (1.17/1.71) of return difference between long and high EPU periods. The evidence in Table 2 suggests that EPU plays a crucial role to predict momentum profits. We demonstrate that momentum profits are significant only when EPU is low. In addition, the forecast power of EPU on momentum payoffs is mainly driven by the short portfolio. 3.2 Time-series regression analysis The results reported above are computed by averaging within low EPU and high EPU months, where this classification is simply a binary measure. Here we conduct an alternative analysis, using time-series regressions to investigate whether the level of EPU forecasts momentum payoffs. We run the following regressions: R t = α + β 1 EPU t 1 + β 2 MKT t + β 3 SMB t + β 4 HML t + ε t (2) where R t is the value weighted excess return at month t on either the long, the short, or the longshort portfolio for momentum strategy; EPU t 1 is the standardized value of the EPU index at month t-1. The EPU index is scaled to have zero mean and a standard deviation of one. To examine whether EPU can predict the momentum strategy, we regress the excess returns on the 10

lagged EPU index as well as the Fama and French three factors. The regression results are reported in Table 3. < Table 3> For the long side, the slope of EPU is insignificant in column (1), while the coefficient estimate of EPU is significantly negative in column (2) after controlling for the Fama and French three factors. Consistent with the portfolio sorting results in Table 2, we find that the short portfolio has substantially lower returns following low levels of EPU. In column (3), the slope of EPU is 1.27 (t-statistic= 2.22), implying that an increase of one standard deviation of EPU is associated with a 1.27% increase in payoffs of the loser portfolio. After controlling for the Fama and French three factors in column (4), the slope of EPU is 0.79 (t-statistic= 2.92). More importantly, we document a significantly negative association between momentum profits and EPU for the WML (long-short) portfolio 5. As can be seen in in columns (5) and (6), the slope coefficients on EPU are negative and significant. In column (5), the coefficient of EPU is -1.13 with a t-statistic of -2.62, statistically significant at the 1% level. We find a similar result after controlling for the Fama and French three factors in column (6). The coefficient of EPU is - 1.11 with a t-statistic of -2.89, implying that one standard deviation decrease in EPU is related to a 1.11% monthly increase in abnormal return for the momentum strategy. The evidence in Table 3 demonstrates that EPU negatively predicts momentum profits and the forecast power is mainly driven by firms in the short side portfolio 6. 3.3 Fama-Macbeth regression analysis 5 We obtain qualitatively similar results using the UMD factor in the Kenneth French s website. The UMD factor constructed based on six value-weight portfolios formed on size and prior (2-12) returns. The winner-minus-loser portfolio return is the average return on the two high prior return portfolios minus the average return on the two low prior return portfolios 6 We find qualitatively similar results using equal-weighted portfolio returns. 11

In addition to time-series regression analysis, we conduct cross-sectional regression analysis. Specifically, we first split the whole sample into the low- and high- EPU months. The low (high) EPU months are those in which the value of EPU at the previous month is above (below) median value. Then we run the Fama and MacBeth (1973) regressions for the full sample period, the low- and high- EPU months as follows: R t+1 = α + β 1 Mom t 2,t 7 + β 2 controls t 1 + ε t (3) where R t+1 is the value-weighted excess return at month t+1; Mom t 2,t 7 is calculated as the firm specific cumulative returns from month t-7 to month t-2. We include other firm characteristics that have roles in explaining the cross-section of average stock returns in control variables. Specifically, these variables are natural logarithm of book-to-market ratio (Fama and French, 1992), natural logarithm of firm size (Banz, 1981), turnover ratio (Lee and Swaminathan, 2000), Idiosyncratic volatility (Ang et. al., 2006), short-term reversal (Jegadeesh, 1990), growth profitability premium (Norvy-Marx, 2010), asset growth ratio (Cooper, Gulen and Schill, 2008), net stock issuance (Loughran and Ritter, 1995), net operating assets (Hirshleifer et. al, 2004) and investment-to-capital ratio (Xing, 2008). Definitions of variables are described in Appendix. < Table 4 > Table 4 reports the Fama-Macbeth regression results. Column 1 evaluates the forecast power of momentum on future stock returns for the whole sample period. This result is consistent with the portfolio-sorts analysis in Panel A of Table 1, momentum plays a strong role in explaining the cross-section of average stock returns after controlling for other anomaly variables. The coefficient of interest, the slope of momentum, is 0.77 (t-statistic=3.65) and statistically significant at the 1 % level. Furthermore, we find that momentum is significant only following low EPU period, confirming our findings in the portfolio sorts and time-series 12

regression analyses. Specifically, the coefficient of momentum is 1.43 (t-statistic=6.42) for the low EPU period in column (3), whereas the slope of momentum is 0.25 (t-statistic=0.76) for the high EPU period in column (2). Our Fama and Macbeth regression analysis demonstrates that momentum anomaly is significant only in the low EPU period. 3.4 Controlling for other time-series determinants of momentum In this subsection, we investigate the role of EPU in forecasting momentum profits by controlling for other time-series determinants. We consider six state variables, including market state (Cooper, Gutierrez and Hameed, 2004), business cycle (Chordia and Shivakumar, 2002), market volatility (Wang and Xu, 2015), investor sentiment (Antoniou, Doukas and Subrahmanyam, 2013), market illiquidity (Avramov, Cheng and Hameed, 2015) and crosssectional return dispersion (Stivers and Sun, 2010). To confirm the distinctive effect of EPU on time-series variations in momentum profits, we regress monthly momentum returns on the lagged EPU index and one of the lagged state variables. The regressions are as follow: R wml,t = α + β 1 EPU t 1 + β 2 MKT t + β 3 SMB t + β 4 HML t + β 5 X t 1 + ε t (4) where R wml,t is the long-short hedge portfolio return for momentum strategy at month t; EPU t 1 is the standardized value of EPU using the EPU index at month t-1. The EPU index is scaled to have zero mean and a standard deviation of one. X t 1 is one of the following variables at month t-1: UP market dummy variable, NBER recession dummy variable, market volatility, investor sentiment and market illiquidity. Specifically, UP market dummy variable (UP market) is defined as one if the prior two-year cumulative market returns is non-negative. Business cycle (BC) is a NBER recession dummy variable that equals one if periods represent recession. Market 13

volatility is defined as the lagged 12-month daily return standard deviation. Investor sentiment employs Baker and Wrugler (2006) s investor sentiment index. Aggregate market illiquidity is defined as the value-weighted average of each stock s monthly Amihud illiquidity. The return dispersion is the lagged three-month moving average of the cross-sectional standard deviation of the monthly returns for the Fama and French 100 portfolios formed on size and book-to-market ratio. The results are presented in Table 5. < Table 5 > As can be seen in Table 5, the predictive power of EPU on momentum profits is robust after controlling for other time-series determinants of momentum. For example, the coefficients of EPU are -0.90 (t-statistic = -2.21) and -0.84 (t-statistic = -2.35) for columns (1) and (2), respectively. The coefficient of market state is significantly positive in column (2), consistent with Cooper, Gutierrez and Hameed (2004) that momentum profits are higher in the bullish market. However, the slope of market state is not significant in column (1), suggesting EPU partially absorb the forecast power of the market state. The evidence in columns (1) and (2) demonstrate that the predictive effect of EPU on momentum profits remain significantly negative after controlling for the market state. Furthermore, from column (3) to column (12), we find that EPU subsumes the predictive effect of business cycle, market volatility, investor sentiment, market illiquidity and return dispersion. The slope coefficients of NBER recession, market volatility, investor sentiment, market illiquidity and return dispersion become insignificant, although the signs of these variables are as expected. Additionally, we control for these variables together in columns (13) and (14). The coefficient on EPU is still negative and significant at the 10% level. In sum, the evidence in Table 5 indicates that EPU is a unique predictor of momentum payoffs. 14

3.5 Controlling for time-varying risk factor The literature has documented that momentum profits have time-varying exposures to the risk factors (Grundy and Martin, 2001; Korajczyk and Sadka, 2004). This subsection addresses the concern that the effect of EPU could be explained by variations in the loadings on the Fama and French three factors. Grundy and Martin (2001) find that the conditional factor risk for the momentum portfolio is a linear function of the ranking-period return. Following Korajczyk and Sadka (2004), we run the following regression to control the time-varying factor risk: R wml,t = α + β 1 EPU t 1 + β 2 MKT t + β 3 MKT t MMKT t 2,t 7 + β 4 MKT t MSMB t 2,t 7 + β 5 MKT t MHML t 2,t 7 +β 6 SMB t + β 7 SMB t MSMB t 2,t 7 + β 8 SMB t MMKT t 2,t 7 + β 9 SMB t MHML t 2,t 7 + β 10 HML t + β 11 HML t MHML t 2,t 7 + β 12 HML t MMKT t 2,t 7 + β 11 HML t MSMB t 2,t 7 + ε t (5) where R wml,t is the long-short hedge portfolio return for momentum strategy at month t; MMKT t 2,t 7, MSMB t 2,t 7 and MHML t 2,t 7 are the average cumulative returns of the Fama and French three factors from the month t-7 to the month t-2. We present the results in Table 6. < Table 6> Specification (1) presents the regression results excluding the level of EPU. Consistent with Grundy and Martin (2001) and Korajczyk and Sadka (2004), the momentum portfolios indeed have time-varying exposures to the risk factors. In addition, the loadings of the risk factors are generally higher following positive ranking-period factor returns. The evidence in the specification (2) shows that the forecasting power of EPU on momentum profits is robust after controlling for time-vary factor risks. The loading of the lagged level of EPU is -0.53 (t-statistic 15

=-2.03), which is negative and significant at the 5% level. The economic impact of implies that one standard deviation increase in EPU reduces the momentum profits by 0.53% per month. In sum, the findings in Tables 5 and 6 further confirm that EPU predicts the time-series variations in momentum profits and the predictive power of EPU is robust after controlling for other states variables and time-varying risk factors. 4. Further analyses of the momentum-epu relation In this subsection, we provide further analyses of the relationship between EPU and momentum profits. First, we investigate whether an asset pricing model with the EPU factor can explain the momentum anomaly in Section 4.1. Second, we examine which component of the EPU index drives the impact of EPU on momentum profits in Section 4.2. Third, we test whether the relationship between EPU and momentum profits is driven by policy uncertainty in Section 4.3. Finally, the predictive power of EPU on momentum in global equity markets and other asset classes are discussed in Section 4.4. 4.1 The EPU factor-mimicking portfolio 4.1.1 Construction of the EPU mimicking factor-ump We project EPU onto the traded return space to form an EPU factor-mimicking portfolio, UMP. Then we aim to test whether a contemporaneous UMP factor can explain momentum. To construct the EPU factor-mimicking portfolio, following Adrian, Etula and Muir (2014), we then run the following regression: log (EPU t ) = a + b [MKT, SMB, HML, RMW, CMA] t + ε t (6) 16

where [MKT, SMB, HML, RMW, CMA] are the Fama and French five factors (Fama and French, 2015) 7. We choose these return factors because they represent a large spread of the return space (from -34.6% to 22.3%). By construction, Cov(log (EPU t ), R t ) = cov(ump t, R t ) + cov(ε t, R t ) = cov(ump t, R t ), for all R t [MKT, SMB, HML, RMW, CMA] t. Ideally, the error term, ε t, is orthogonal to the space of returns so that the covariance of any asset with EPU is identical to its covariance with the UMP. Thus, the EPU mimicking factor, UMP t, is the fitted value of the regression. In addition, we normalize the weights, b, to sum to one for the convenience of units. The monthly return for the UMP factor is estimated by UMP t = a + b [MKT, SMB, HML, RMW, CMA] t where b = b = [ 0.31, 0.38, 1.95, 1.17,1.72]. b 4.1.2 The pricing results using the UMP We examine whether the contemporaneous EPU mimicking factor-ump can reduce momentum alpha. In addition, we compare the performance of explaining momentum using several factor models. Specifically, we regress the long-short portfolio returns of momentum on the several factors and compute the alpha and R-squared for each model. The regressions are as follows: R wml,t = α + β 1 UMP t + ε t (7) where R wml,t is the long-short hedge portfolio return of momentum strategy at month t; UMP t is the EPU mimicking factor estimated in the subsection 4.1.1. We report the results in Table 7. < Table 7> 7 We download the five factors from Kenneth French s website. 17

The evidence in Table 7 delivers a clear message that the contemporaneous EPU mimicking factor-ump explains momentum well. For example, the monthly alpha in column (1) is 0.59% (t-statistic=1.61). The single UMP factor model has higher explaining power for momentum profits than the Fama and French three-factor model. Specifically, the R-squared in column (1) is 0.10, compared with 0.08 in column (2). Furthermore, we add the UMP factor to the Fama-French three-factor model in column (3). Alpha is 0.74% (t-statistic=1.77) in column (3), compared to 1.13% (t-statistic=3.08) in column (2). The R-squared increases from 0.08 in column (2) to 0.11 in column (3). Overall, the UMP performs well compared with the Fama and French three factors in terms of a relatively high R-squared and a low alpha. Table 7 shows that returns of momentum substantially decrease after adjusting for the policy uncertainty risk, suggesting that the EPU mimicking factor-ump has the power to explain momentum anomaly. 4.2 Decomposing EPU into three components The EPU measured by the EPU index has three components. The first component, EPU news-based component, is a normalized index of the volume of news articles discussing economic policy uncertainty in 10 large newspapers. The second component, EPU related to taxcode, is based on the present value of future scheduled tax code expirations using data from the Congressional Budget Office. The third component, EPU related to CPI and government purchase, is based on disagreement among professional forecasters over future government purchases and consumer prices. As the EPU index is a weighted average of three components, it is important to determine which component contributes the predictive effect of momentum profits. We repeat the time-series regression analysis in Table 3 for each component of the EPU index. For each component of the EPU index, we run the following regressions: 18

R wml,t = α + β 1 EPU Component t 1 + β 2 MKT t + β 3 SMB t + β 4 HML t + ε t (8) where R wml,t is the value weighted hedge portfolio at month t for momentum strategy. EPU Component t 1 is the standardized value for each component of the EPU index at month t- 1. Each component of the EPU index is scaled to have zero mean and a standard deviation of one. We report the results in Table 8. < Table 8> The evidence shows that the major explaining power of EPU on momentum payoffs comes from the news-based component and tax-based component. In column (2), the slope of EPU news-based component is -1.14 with a t-statistic of -2.85, suggesting that a one standard deviation increase in the EPU news-based component is associated with a 1.14% decrease in hedge portfolio return for momentum. Columns (3) and (4) suggest that uncertainty related to tax-code also contributes to the predictive effect of momentum payoffs. For instance, in column (3), the coefficient estimate of EPU tax-code based component is -0.71 with a t-statistic of -2.08, significant at the 5% level. After controlling for the Fama and French three factors in column (4), the slope of uncertainty related to tax reduces to -0.64 (t-statistic=-1.94) and significant at the 10% level. Furthermore, we find that uncertainty related to inflation and government purchase does not have explaining power for momentum payoffs. The slopes of inflation and government purchase components are negative but insignificant in columns (5)-(8). 4.3 Controlling for macroeconomic uncertainty Which uncertainty contributes to the significant predictive effect of EPU on momentum, economic uncertainty or policy uncertainty? We address this question in this subsection, by introducing the Jurado, Ludvigson and Ng (2015) index to measure economic uncertainty. Jurado, 19

Ludvigson and Ng (2015) s measure is constructed using the aggregation of individual conditional volatilities, which are estimated based on unpredictable component of the future value of 132 macroeconomic series. We download the one-month, three-month and 12-monthahead economic uncertainty indices (EU1, EU2 and EU3) from Sydney Ludvigson s website. Then we run the following regressions to control for economic uncertainty: R wml,t = α + β 1 X t 1 + β 2 EPU t 1 + β 3 MKT t + β 4 SMB t + β 5 HML t + ε t (9) where X t 1 is one of the three economic uncertainty indices. Each index is scaled to have zero mean and a standard deviation of one. We present the results in Table 9. < Table 9> We first examine whether economic uncertainty can forecast the momentum payoffs, in which we find mixed results. We find no evidence that economic uncertainty predicts momentum profits using monthly long-short hedge portfolio returns regress on economic uncertainty alone. For instance, in columns (1), (5) and (9), the slopes of three economic uncertainty measures are all negative but insignificant. However, we find that economic uncertainty negatively forecasts momentum profits after adding the Fama and French three factors in regressions. For example, among columns (2), (6) and (10), the coefficients of economic uncertainty are all significantly negative. Then we test whether the negative predictability of EPU on momentum profits is robust after controlling for economic uncertainty. Our regression results find that the effect of EPU on momentum payoffs remains unaffected. Furthermore, EPU subsumes the effect of economic uncertainty on momentum in models using the Fama and French three factors and EPU as regressors. For instance, in columns (4) and (8), the coefficients of economic uncertainty are significant only at the 10% level. In column (12), the slope of economic uncertainty is 20

insignificant. Table 9 demonstrates that the predicting effect of EPU on momentum is robust after controlling for economic uncertainty. 4.4 Global equity and other asset classes A recent study of Asness, Moskowitz and Pedersen (2013) documents strong momentum effects exist not only among global equity markets but also in other asset classes including global equity index futures, currencies, global government bonds and commodity futures. In this subsection, we investigate whether the predictive power of EPU on momentum is robust in global equity markets and other asset classes. We use the monthly Global Economic Policy Uncertainty index (GEPU) to capture the overall policy uncertainty in the global economy from January 1997 to December 2016. The GEPU is a GDP-weighted average of national EPU indices for 18 countries: Australia, Brazil, Canada, Chile, China, France, Germany, India, Ireland, Italy, Japan, the Netherlands, Russia, South Korea, Spain, Sweden, the United Kingdom and the United States 8. We download the monthly global momentum factor from Kenneth French s website. The momentum returns for other assets are from Asness, Moskowitz and Pedersen (2013), which contain equity country index futures across 18 developed equity markets, 10 currencies across developed markets, 10 country government bonds and 27 different commodity futures 9. The market index for each asset is as follows: it is the global market excess return from Kenneth French s website for equity, the MSCI World Index for country index futures, an equal-weighted average of the securities for 8 See Baker, Bloom and Davis (2016) for a detailed discussion of how to construct national EPU indices. The GEPU index is downloaded from the Economic Policy Uncertainty website. 9 We download the updated monthly momentum returns for different assets from AQR capital management: https://www.aqr.com/library/data-sets/value-and-momentum-everywhere-portfolios-monthly. This data set is an updated and extended version of Asness, Moskowitz and Pedersen (2013). Details on data description and sources can be found in Asness, Moskowitz and Pedersen (2013). 21

currency, the Bloomberg Barclays global treasury total return index for government bonds and the S&P Goldman Sachs Commodity Index (S&P GSCI) for commodity. Additionally, we add nonstock assets and all assets in our analysis. The momentum returns for nonstock assets are calculated by taking average of momentum premiums across all nonstock assets. Similarly, the momentum profits for all assets are obtained by taking average of momentum returns for each asset. To investigate the predictive effect of GEPU on momentum for each asset, we regress the monthly momentum returns on the GEPU alone. Then we add the market excess returns for each asset in our regression analysis to control for the market risk. Table 10 reports the regression results. < Table 10> We find evidence that the GEPU significantly negatively forecasts momentum profits among five asset classes including global stocks, country index futures, commodity, nonstock assets and all assets. For instance, the coefficient estimator of GEPU is -0.73 (t-statistic=-2.74) for global stocks in column (1), implying that a one-standard-deviation increase in GEPU is associated with a 0.73% decrease in global momentum returns. Additionally, the coefficients on GEPU for country index futures, commodity, nonstock assets and all assets are significantly negative. The results are unaffected after controlling for the market excess returns. However, the relations between GEPU and momentum returns are insignificant for government bonds and currency. From Table 10, we show that the effects of EPU on momentum profits are robust in the global equity market and among many other asset classes. 5. Other Robustness Checks 22

We conduct several additional robustness checks in this subsection. The relationships between EPU and momentum at a variety of holding periods are discussed in Section 5.1. We also examine the forecasting power of EPU on momentum profits in three subsamples based on size, institutional ownership and analyst coverage. 5.1 EPU and momentum at various holding periods We examine the predictability of EPU on momentum payoffs using different holding period returns. The holding periods for momentum are ranging from one- to twelve-months. For each holding period, we compute the average value-weighted hedge portfolio returns and the risk-adjusted returns for the low- and high-epu months. Figure 1 presents the results for average hedge portfolio return and risk-adjusted returns up to twelve-months holding period. It is clear to observe that differences in average hedge portfolio returns between the low- and high-epu periods decrease with holding periods. We present the time-series predict regressions using different holding period returns in Table 11 using the same setting as Table 2. For instance, in Panel A of Table 11, the regression coefficients of EPU are -1.13 (t-statistic = -2.62) for one month, -0.49 (t-statistic =-1.65) for six months and -0.24 (t-statistic =-0.89) for twelve months, respectively. Panel A suggests that EPU could negatively predict momentum profits for up to 6 months. After controlling for the Fama and French three factors in Panel B, the negative predictive of EPU on momentum also persists around 6 months. 10 Overall, we show that the predictive power of EPU on momentum profits decreases with portfolio holding periods. The negative relation between EPU and momentum profits are significant up to next 6 months. 10 In an un-reported table, the raw return difference in monthly momentum profits between the low- and high-epu months is 1.86% for the one-month holding period, compared to 0.22% for the 12-months holding period. The riskadjusted return difference in momentum profits between the low- and high-epu months is 1.71% for the one-month holding period, compared to 0.10% for the 12-months holding period. 23

< Figure 1> < Table 11> 5.2 Subsamples by firm size, institutional ownership and analyst coverage We investigate the predictability of EPU on time-series variations in momentum profits in subsamples based on three characteristics: size, institutional ownership and analyst coverage. Small (large) stocks are those whose sizes are lower (higher) than NYSE 50 percentile. Low (high) institutional ownership stocks are those whose institutional ownership is below (above) the median value for each quarter. Similarly, low (high) analyst coverage stocks are those whose analyst coverage is below (above) the median value for each month. For each subsample, we regress the value-weight long-short portfolio returns on the standardized value of EPU as well as the Fama and French three factors. Table 12 presents the results for size, institutional ownership and analyst coverage in Panels A, B and C, respectively. We find that EPU negatively predicts momentum profits for each subsample. For instance, in Panel A, for large stocks, the slope of EPU is -1.15 (t-statistic=-2.82) in column (4), suggesting that an increase of one standard deviation in EPU is related to a 1.15% decreases in momentum profits. Also, the coefficient estimate of EPU is -1.07 (t-statistic=-2.57) in column (2) for small stocks. We find consistent results for low and high institutional ownership stocks as well as low and high analyst coverage firms. In Panel B, the coefficient of EPU is -1.10 (t-statistic=-2.68) in column (4), compared with -1.16 (t-statistic=-2.37) in column (2). In Panel C, the slope of EPU is -1.07 (t-statistic=-2.60) in column (4), as opposed to -0.83 (t-statistic=-2.09) in column (2). < Table 12> 24

Overall, we further show that EPU negatively forecasts momentum profits for small and large stocks; for low and high institutional ownership firms; and for low and high analyst coverage firms. 6. Conclusion Economic policy uncertainty plays a significant role in predicting momentum profits. Using a news-based measure of EPU, we demonstrate that EPU negatively forecasts momentum profits. Specifically, an increase of one standard deviation in EPU is associated with a 1.13% decrease in momentum returns for the winner-minus-loser portfolio. We show that momentum is significant only following low levels of EPU. The difference in returns between the low- and high-epu periods is significant. Furthermore, we find that the effect of EPU on momentum is mainly driven by the short-side portfolio. The difference in returns between the low-and high- EPU periods is -2.49 % (t-statistic=-2.82) for the short portfolio, compared with -0.58 % (tstatistic=-0.92) for the long portfolio. To understand our findings, we borrow the fund flow-based mechanism of momentum from Lou (2012) that winning funds keep investing in past winner stocks and losing funds liquidate their holdings in past loser stocks. Combining with the implication of Starks and Sun (2016) that investors learning about signals of fund performance weakens when uncertainty increases, the flow-based mechanism may work only in the state of Low EPU, leading to the significant payoffs of momentum. We further demonstrate that the predictive power of EPU on momentum remains significantly negative after controlling for the state of economy variables and time-varying risk factors. EPU subsumes the predictive effect of business cycle, market volatility, investor 25

sentiment, market illiquidity and return dispersion. We also construct an EPU mimicking portfolio and show that the single mimicking factor model explains momentum profits well. Further analysis shows that the explanatory power of EPU on momentum payoffs mainly comes from the news-based component and tax-based component; the effect of EPU on momentum is mainly attributed to the policy uncertainty rather than economic uncertainty. Finally, we show that a global EPU index helps explain momentum profits in the global equity market and other asset classes. Overall, our findings suggest that EPU is an important determinant of time-series variations in momentum profits. 26