Journal of Empirical Finance

Transcription

1 Journal of Empirical Finance 16 (2009) Contents lists available at ScienceDirect Journal of Empirical Finance journal homepage: The cross section of cashflow volatility and expected stock returns Alan Guoming Huang School of Accounting and Finance, and Center for Advanced Studies in Finance, University of Waterloo, Waterloo, Canada ON N2L 3G1 article info abstract Article history: Received 20 May 2008 Received in revised 6 October 2008 Accepted 14 January 2009 Available online 23 January 2009 JEL Classifications: G12 G14 Keywords: Cashflow volatility Expected stock returns Idiosyncratic return volatility Cross section I show that historical cashflow volatility is negatively related to future returns cross-sectionally. The negative association is large; economically meaningful; long-lasting up to five years; robust to known return-informative effects of size, value, price and earnings momentums and illiquidity; and extends to both systematic and idiosyncratic cashflow volatilities. Using the standard deviations of cashflow to sales and of cashflow to book equity as proxies for cashflow volatility, the least volatile decile portfolio outperforms the most volatile decile portfolio by 13% a year relative to the Fama French four factors. The cashflow volatility effect is closely related to the idiosyncratic return volatility effect documented in Ang et al. [Ang, A., Hodrick, R.J., Xing, Y. and Zhang, X. The cross-section of volatility and expected returns. Journal of Finance, 51 (2006), ]. However, in portfolios simultaneously sorted on both cashflow and return volatilities, and in cross sectional regressions of returns at the firm level, these two effects neither drive out nor dominate each other. While the pricing of idiosyncratic cashflow volatility represents an anomaly against the traditional asset pricing theories, the pricing of historical cashflow uncertainty sheds light on potential fundamental risks embodied in the Fama French HML and SMB factors Elsevier B.V. All rights reserved. 1. Introduction There is a growing literature documenting a negative relationship between observed volatility and future stock returns. In a frequently cited paper, Ang et al. (2006) find that both systematic and idiosyncratic volatilities of stock returns are negatively associated with future returns cross-sectionally. Using a number of proxies for information uncertainty, such as firm age, size, analyst forecast dispersion and return volatility, both Jiang et al. (2005) and Zhang (2006) find that high information uncertainty, in the form of large analyst dispersion or return volatility, induces negative future returns. The negative association between analyst forecast dispersion and future returns is also documented in Diether et al. (2002). The literature also links idiosyncratic return volatility to earnings or cashflow volatility. Pastor and Veronesi (2003) argue that learning about firms' uncertainty in future profitability increases their idiosyncratic return volatility. Wei and Zhang (2006) find that the rise in idiosyncratic stock volatility documented in Campbell et al. (2001) and Morck et al. (2000) from the 1970s to the 1990s is largely attributable to a decreasing return-on-equity and an increasing volatility in return-on-equity. Moving a step further, Irvine and Pontiff (in press) argue that cashflow shocks and increased economy-wide market competition are primary drivers for the documented trend in return volatility. The Ang et al. (2006) results indicate a negative association between return volatility and future returns. If return volatility is positively associated with earnings or cashflow volatility, we should expect a An earlier version of this paper was titled The Cross Section of Earnings Volatility and Expected Stock Returns. This paper was developed from Chapter Three of my dissertation. I am grateful to my committee members Eric Hughson, Chris Leach, Martin Boileau, Michael Stutzer and Jamie Zender, as well as Geert Bekaert (the editor), Changling Chen, Alan Douglas, Ranjini Jha, Noah Stoffman, Ken Vetzal, Tony Wirjanto, two anonymous referees, and seminar participants at the University of Colorado at Boulder, the University of Waterloo and McMaster University for many helpful comments. I acknowledge financial support from the Social Sciences and Humanities Research Council of Canada (SSHRC). I am solely responsible for all remaining errors. Tel.: x address: aghuang@uwaterloo.ca /$ see front matter 2009 Elsevier B.V. All rights reserved. doi: /j.jempfin

2 410 A.G. Huang / Journal of Empirical Finance 16 (2009) similar negative association between earnings/cashflow volatility and future stock returns. This paper links the above findings and examines the impact of firms' cashflow volatility on returns. Iconfirm the negative association between return and cashflow volatility by examining the cross-sectional relationship between historical cashflow volatility and ex post returns. The negative association is strong and consistent. My findings consist of three pieces of evidence: (1) the economic magnitude, (2) the pricing of systematic cashflow volatility and idiosyncratic cashflow volatility, and (3) the relationship of cashflow volatility with return volatility in explaining returns. I briefly discuss these results in order. The magnitude of the negative association between historical cashflow volatility and future returns is large and long-lasting. In my benchmark cases I use two seasonality-adjusted measures of cashflow volatility: standard deviation of cashflow to sales and standard deviation of industry-adjusted cashflow to book equity of the past sixteen quarters. I form ten decile portfolios based on increasing values of cashflow volatility using all listed firms on the NYSE, NASDAQ and AMEX from 1980 to Between the value-weighted return on decile one and the value-weighted return on decile ten, the Jensen's alpha spreads relative to the Fama French three factors of market, SMB, and HML and Carhart's (1997) price momentum factor (hereafter Fama French four factors ) are respectively 1.19% and 1.06% a month for the two volatility measures, or equivalently 13% a year. The value-weighted raw returns have similar sizes of spread. Furthermore, the magnitude of the spread exists not only for one-month-ahead return, but also for six-month and one-year-ahead buy-and-hold returns. It is only at an investment horizon of two years that the spreads start to decline, although the effect largely lasts up to year five. The portfolio level relation between cashflow volatility and one-month to five-year returns survives at the firm level with multiple controls of return-informative variables that include the Fama French four factors of market, size, book to market and price momentum, earnings momentum (Chan et al., 1996), illiquidity (Amihud, 2002), and earnings yield (Haugen and Baker, 1996). In the Fama and MacBeth (1973) cross-sectional regression controlled for the Fama French four factors, a firm with average cashflow/sales (cashflow/book equity) volatility experiences a monthly return of 0.10 (0.20)% less due to its volatility. The pricing effect of cashflow volatility is due to both systematic volatility and idiosyncratic volatility. When total cashflow is decomposed into systematic and idiosyncratic cashflow relative to the industry mean, I find that both systematic volatility and idiosyncratic volatility are priced similarly with the total volatility. In fact, these three volatilities move together: Firms with large total volatility tend to display both large idiosyncratic volatility and large systematic volatility. The indifference of the results to the cashflow decomposition is perhaps because cashflow is not directly traded, and therefore firms are not motivated to peg their cashflows to the industry mean according to their exposures to industry cashflow. My final piece of evidence is that the cashflow volatility effect is different from the idiosyncratic return volatility effect at both the portfolio and firm levels and that neither effect drives out the other, although they are highly correlated. At the portfolio level, high cashflow volatility portfolios display high idiosyncratic return volatility. However, in double sorting of portfolios on cashflow volatility controlled for idiosyncratic return volatility, the cashflow volatility effect clusters in medium to high return volatility firms. Similarly, the return volatility effect clusters in medium to high cashflow volatility firms. The cashflow volatility effect is different from the return volatility effect also in that in firms of the top 30% of cashflow volatility, only 30% of them rank in the top 30% of idiosyncratic return volatility. Furthermore, controlling for return volatility, the overall Fama French four-factor alpha spread between the least volatile cashflow quintile and the most volatile cashflow quintile is 0.70% a month at 1% significance level. This spread size is similar to the unconditional spread when no control is imposed. The cashflow volatility effect is also different from the return volatility effect at the firm level. In Fama MacBeth cross sectional regressions of stock returns, both effects co-exist. I also decompose cashflow volatility to a component related to and a component orthogonal to contemporaneous and lagged return volatilities. The cashflow volatility effect remains for the orthogonal component at the firm level. In sum, the evidence at the portfolio and firm levels points to a separate cashflow volatility effect that is no weaker than the idiosyncratic return volatility effect. My results survive a wide range of robustness checks. Bali and Cakici (2008) dispute Ang et al.'s (2006) findings on idiosyncratic volatility, emphasizing the sensitiveness of Ang et al.'s results to different estimation windows of volatility, weighting schemes for portfolio returns, breakpoints for portfolio sorting, and size and liquidity controls. In light of Bali and Cakici's criticism, I check the robustness of my results against the following alternatives: (1) eight measures of cashflow volatility that also include measures based on accounting earnings, (2) two other estimation windows for cashflow volatility using either twelve or twenty quarters of past cashflows, (3) controlling for size, book-to-market equity, price momentum, earnings momentum, and liquidity, (4) two schemes of portfolio breakpoints based on either all CRSP stocks or only NYSE stocks, (5) sub-period breakdown by year, and (6) value-weighting versus simple average of returns for portfolios. In all of these cases, the results stand up well. This paper contributes to the growing literature that historical volatility is negatively correlated with future stock returns. The pricing of cashflow volatility itself has not received much attention in the empirical asset pricing literature. Earlier literature focuses on the cross-section pricing implication of the level of earnings or the change in earnings. For example, Basu (1977) and Haugen and Baker (1996) report that the earnings to price ratio (earnings yield) is positively related to future stock returns. 1 Chan et al. (1996) confirm the long-standing post-earnings-announcement-drift phenomenon that firms who release positive earnings surprises tend to experience subsequent positive abnormal returns, and relabel it earnings momentum. To the best of my knowledge, Haugen and Baker (1996) is the only paper that includes volatilities in earnings and cashflow yields in a cross-sectional regression of returns. In more than fifty firm attributes, these authors find that the variability in cashflow yield is negatively related to future stock returns, but do not find significant relationship between returns and volatilities in earnings and dividend yields. 1 However, Fama and French (1992) find that the usefulness of earnings yield is subsumed by size and book to market ratio.

3 A.G. Huang / Journal of Empirical Finance 16 (2009) Haugen and Baker (1996) provide no further discussion on this finding. This paper is different from Haugen and Baker (1996) in that it shows that both earnings volatility and cashflow volatility (not just cashflow yield volatility) matter, and that the economic significance of the pricing of cashflow volatility is non-negligible. In addition, Haugen and Baker (1996) use market size as the scalar to calculate cashflow volatility, making it hard to separate their results from the familiar size effect. That cashflow volatility, return volatility or analyst forecast dispersion is negatively correlated with future stock returns contradicts the traditional notion that volatility connotes risk, and therefore should be compensated with higher returns. In particular, the pricing of idiosyncratic cashflow volatility presents yet another anomaly against the traditional asset pricing theories. Some potential explanations can be offered for the results found in this and other similar papers. First, one may argue that historical volatility is different from expected volatility in the sense that volatility is itself a mean-reverting process. Second, in a follow-up paper to Diether et al. (2002), Sadka and Scherbina (2007) argue for limits to arbitrage such as transaction costs (that arbitrageurs cannot or would not short overpriced stocks that are high in forecast dispersion due to arbitrage limits). Jiang et al. (2005) resort to investor overconfidence (that investors are even more over confident on high-volatility firms and hence overprice these firms). Finally, cashflow volatility may represent the cashflow uncertainty risk component in several interpretations of the Fama French's HML and SMB factors. Fama and French (1992, 1993, 1995, 1996) and Chen and Zhang (1998), among others, interpret HML as a measure of distress risk. Cashflow uncertainty captures distress risk at least partly firms with volatile cashflows suffer from higher default probability and hence larger distress risk. Chan and Chen (1991) argue that small firms tend to have high financial leverage and cashflow problems and are less likely to survive economic downturns. Thus it is also likely that cashflow uncertainty is a component in SMB. Nonetheless, this paper focuses on presenting a robust empirical result and leaves for future research a credible theoretical explanation for the phenomenon. The rest of the paper is organized as follows. Section 2 describes the data and defines cashflow volatility variables. Section 3 details returns on and properties of portfolios sorted on cashflow volatility. Section 4 reports the firm-level results. Section 5 provides robustness checks of longer-term returns and the decomposition of cashflow volatility into systematic and idiosyncratic volatilities. Section 6 explores the relationship between idiosyncratic return volatility and cashflow volatility in explaining returns. Section 7 concludes. 2. Data and variable definitions 2.1. Cashflow volatility measures My initial sample consists of all NYSE/NASDAQ/AMEX-listed firms for which I find data on the merged CRSP/Compustat monthly stock returns and quarterly accounting items from 1973 to I select the survival-bias-free combined Compustat data, which include research files of extinct and acquired companies. Large-scale quarterly data exist on Compustat from 1973, and quarterly data items which are necessary for computing cashflow exist in large numbers from In a recent study that links the return volatility trend to the earnings volatility trend, Wei and Zhang (2006) use the same quarterly data from 1976 to To ensure that accounting information is known prior to trading, I match stock returns to accounting numbers of the prior fiscal quarter. I eliminate financial services companies (SIC code between 6000 and 6999) and observations with negative shares outstanding, negative assets and negative book equity. I investigate the relationship between ex post returns and historical cashflow volatility. The nature of this study requires the estimation of cashflow volatility, for which the sample should ideally contain as many time-series observations as possible. Therefore, unlike many previous studies that use annual data, I use quarterly data to increase the number of observations. Matching quarterly Compustat data with monthly returns also implies that accounting information is impounded into stock prices more promptly than is the case when annual accounting data are matched with monthly returns. I compute cashflow volatility as the rolling standard deviation of the standardized cashflow over the past sixteen quarters (four years). I require at least eight nonmissing observations of cashflow within this estimation window. Although the choice of the estimation window of four years is somewhat arbitrary, I can report that virtually all of my results are robust to estimation windows of three years (twelve quarters) and five years (twenty quarters). Adjusted for the first four years needed for cashflow volatility calculation, my final sample period covers the period I choose cashflow to proxy for firms' economic earnings. Direct use of accounting earnings may disguise firms' operational profit due to the pervasive earnings management documented in the accounting literature and may subsequently underestimate the volatility of operational profit (e.g., Healy, 1985; Dechow et al., 1995). I define cashflow from operations, CF, as the sum of earnings before extraordinary items, depreciation and amortization, and change in working capital, where following Fama and French (1992), Idefine accounting earnings as income before extraordinary items minus preferred dividends. 2 For the purpose of cross-sectional aggregation, cashflow needs to be standardized by firm size. The question is, which variable should be used as the scalar? Some choices of the scalar in previous studies are book equity (e.g., Shroff, 1999; Wei and Zhang, 2006), shares outstanding (e.g., Allayannis et al., 2005; Waymire, 1985), and the absolute value of the variable's own mean (e.g., Barnes, 2001; Minton and Schrand, 1999; Minton et al., 2002). My focus is on operating variables. I therefore exclude shares outstanding as the scalar. Similarly, I do not use market equity as the scalar because market equity contains information about 2 In their analysis using annual data, Fama and French (1992) define accounting earnings as earnings before extraordinary items plus preferred dividends minus deferred taxes. However, in the Compustat quarterly data file a substantial amount of deferred taxes observations are missing. Including deferred taxes would then result in a severe loss of observations. That said, I can report that including deferred taxes in the earnings definition does not change my results.

4 412 A.G. Huang / Journal of Empirical Finance 16 (2009) return and may be different from operating variables (see, e.g., Berk, 1995). Following Wei and Zhang (2006), I use book equity as the scalar. In addition, I also use sales as the scalar. I choose sales for two reasons. First, sales is also used in prior studies as a measure of firm size (e.g., Berk, 1997). Second, using sales as the scalar has the advantage of addressing seasonality in cashflow. There is substantial evidence in the accounting literature suggesting that operating variables exhibit significant seasonality (e.g., Brown, 1993). The data confirms its existence. The full-sample lag-four autocorrelation between the current quarter cashflow and the cashflow of a year ago is high at Sales positively co-move with earnings, which is confirmed in my sample: the contemporaneous correlation between cashflow and sales is Scaling cashflow by sales greatly reduces the autocorrelation in cashflow. The lag-four autocorrelation of cashflow to sales reduces to only. I also deseasonalize cashflow to book equity. I adjust it by its industry mean; that is, I use the excess cashflow to book equity over and above the mean cashflow to book equity of the industry defined by the first two-digit SIC code. 3 This method accounts for the seasonality common to the industry that the firm is in. The lag-four autocorrelation of the adjusted cashflow to book equity is 0.20, again substantially lower than the autocorrelation of CF. Based on the analysis above, I construct these two cashflow volatility measures: standard deviation of cashflow to sales, labeled as CFSALES, and standard deviation of seasonality-adjusted cashflow to book equity, labeled as CFBE. To mitigate the effect of extreme observations, I follow the literature (e.g., Minton et al., 2002) and winsorize the volatility measures at the 1st and 99th percentiles over the full sample period. Such winsorization greatly reduces the excess kurtosis and the upper bound of the volatility measures. For example, the maximum of CFSALES before the winsorization is , and after the winsorization reduces to only I winsorize monthly stock returns similarly. 4 It is noteworthy to point out that since most of my results depend on the ranking of cashflow volatility of a firm, winsorization of volatility would not affect those results Return informative variables I control for the following return-informative variables when studying the effect of cashflow volatility on returns: size, book to market equity, price momentum, idiosyncratic return volatility, earnings momentum, illiquidity, and earnings yield. I select these variables based on three considerations. First, size, book to market, price momentum and illiquidity are now well acknowledged factors affecting stock returns. 5 Second, earnings momentum and earnings yield are profitability-related variables that may subsume cashflow volatility in predicting stock returns. 6 Finally, I control for idiosyncratic return volatility to differentiate this paper from Ang et al. (2006). Idefine size (ME) as the beginning of the period market equity (lagged one-month market equity), book to market equity (BEME) as book equity to market equity, and earnings yield (EY) as earnings to market equity. Following Ang et al. (2006), Idefine price momentum (PMOM) as the past twelve-month return qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi and idiosyncratic return volatility (IRV) relative to the Fama French threefactor model. Specifically, monthly IRV is defined as var i t in the following regression of daily excess stock returns for firm i: R i t R f t = α i + β i MKTMKT t + β i SMBSMB t + β i HMLHML t + i t; ð1þ where R t i is the stock return at day t during the month, R f is the riskfree rate, MKT, SMB and HML are the daily Fama French three factors of market, size and value, respectively, and ε is the residual. Eq. (1) is estimated with the previous one-month daily returns and is updated every month for each stock. I follow Chan et al. (1996) to define earnings momentum and Amihud (2002) to define illiquidity. Earnings momentum at month t is defined as the standardized unexpected earnings (SUE) in the following expression: SUE i t = ei q e i q 4 ; ð2þ σ i t i i where e q is the most recent quarterly earnings, e q 4 is earnings four quarters ago, and σ i t is the standard deviation of unexpected earnings, e i i q e q 4, over the preceding sixteen quarters. Illiquidity (ILLIQ) for month-t is defined as: ILLIQ i t = previous within month mean of jdaily returnj daily volume i cross sectional average of the numerator : i The ILLIQ ratio gives the absolute percentage price change per dollar of daily trading volume, or the daily price impact of the order flow, standardized by the cross sectional mean. 3 I thank an anonymous referee for suggesting this measure. 4 The subsequent return informative variables are also winsorized analogously. The results remain virtually the same if these variables are winsorized month by month. 5 See, e.g., Fama and French (1992) for the size and value (book to market) effects, Jegadeesh and Titman (1993) for the price momentum effect, and Amihud (2002) for the illiquidity effect. 6 See, e.g., Haugen and Baker (1996) for the relationship between earnings yield and returns, and Chan et al. (1996) for the relationship between earnings momentum and returns.

5 A.G. Huang / Journal of Empirical Finance 16 (2009) Table 1 Time-series averages of cross-sectional summary statistics. Number of observations Mean Std dev Minimum Median Maximum CF 3, , ,398.8 ME 4, , ,874.8 BEME 4, PMOM (%) 4, SUE 3, ILLIQ 4, EY 4, IRV (%) 4, CFSALES 2, CFBE 2, RET (%) 4, This table presents the time-series averages of cross-sectional summary statistics of the variables. The data covers all NYSE, NASDAQ, and AMEX-listed firms from the quarterly files of COMPUSTAT and monthly return files of CRSP from January 1980 to December CF = cashflow from operation (in million $); ME = market value of equity (in million $); BEME = book to market equity, measured as book equity of last quarter to market equity at the beginning of the month; PMOM = price momentum, measured as the past 12-month return; SUE = standardized unexpected earnings; ILLIQ = the illiquidity measure in Amihud (2002); EY = earnings yield, measured as earnings of last quarter divided by beginning-of-the-month market equity; IRV = idiosyncratic volatility of the previous month's daily returns relative to the Fama French three factors; CFSALES = standard deviation of cashflow to sales over the past 16 quarters; CFBE = standard deviation of the industryadjusted cashflow to book equity over the past 16 quarters; and RET = monthly stock return. PMOM, SUE, ILLIQ, IRV, CFSALES, CFBE and RET are winsorized at the 1st and 99th percentiles over the full sample Summary statistics and correlation Table 1 reports the time series averages of the cross-sectional summary statistics of the above variables. Two observations are in order. First, the sample size is large. On average, the sample has about 4000 firm observations in returns and 3000 observations in cashflow volatility per month. Second, the sample consists of more small firms than large firms. The mean firm size is significantly higher than its median. This skewness of firm size distribution is consistent with prior studies. Table 2 presents the time series averages of the cross-sectional correlations of the variables in Table 1. A quick examination reveals several unsurprising observations. First, returns are positively correlated with book to market, price momentum, SUE, ILLIQ and earnings yield. Second, corroborating the findings of Ang et al. (2006) and Wei and Zhang (2006), the correlation between IRV and return is negative, and the correlation between cashflow volatility and IRV is highly significant. Consistent with these two correlations, cashflow volatility is significantly negatively correlated with return. The cross-sectional correlation between return and size (lagged market equity) is insignificant, suggesting a weak size effect during the sample period. 3. Portfolios sorted on cashflow volatility 3.1. Returns In this section, I examine the raw and risk-adjusted returns on portfolios sorted on cashflow volatility. I form ten decile portfolios every month using all stocks in the sample based on increasing cashflow volatility breakpoints. Following Ang et al. Table 2 Time-series averages of cross-sectional correlations of variables. CF ME BEME PMOM SUE ILLIQ EY IRV CFSALES CFBE RET CF 1 ME BEME 1 PMOM SUE ILLIQ EY IRV CFSALES CFBE RET This table presents the time-series averages of cross-sectional correlations of the variables. The data covers all NYSE, NASDAQ, and AMEX-listed firms from the quarterly files of COMPUSTAT and monthly return files of CRSP from January 1980 to December CF = cashflow from operation; ME = market value of equity; BEME = book to market equity, measured as book equity of last quarter to market equity at the beginning of the month; PMOM = price momentum, measured as the past 12-month return; SUE = standardized unexpected earnings; ILLIQ = the illiquidity measure in Amihud (2002); EY = earnings yield, measured as earnings of last quarter divided by beginning-of-the-month market equity; IRV = idiosyncratic volatility of the previous month s daily returns relative to the Fama French three factors; CFSALES = standard deviation of cashflow to sales over the past 16 quarters; CFBE = standard deviation of the industry-adjusted cash-flow to book equity over the past 16 quarters; and RET = monthly stock return. Underscored numbers are significant at the 5% level. Bold faced numbers are insignificant. All other numbers are significant at the 1% level.

6 414 A.G. Huang / Journal of Empirical Finance 16 (2009) (2006), the breakpoints are determined by the ranked values of all stocks and are updated month by month. I then calculate the monthly value-weighted return of each portfolio weighted by each stock's market equity at the beginning of the month, as well as the monthly simple average return. The first two rows of Panels A and B of Table 3 report the value-weighted and simple average returns of portfolios sorted on CFSALES and CFBE, respectively. The column labeled D1 D10 shows the spread between the smallest decile portfolio (D1) and the largest decile portfolio (D10), and the column labeled D1:5 D6:10 shows the spread between the simple average of deciles 1 to 5 and the simple average of deciles 6 to 10. One common feature across both panels is a positive, large D1 D10 spread in both value-weighted and simple average returns. In Panel A, the D1 D10 spread in value-weighted return is 1.35% per month (with a t-statistics of 3.78), and in Panel B the spread is 0.64% per month (with a t-statistics of 2.63). This translates into an annual return difference of about 9 15%. The return spreads for the simple average returns are about the same magnitude: the D1 D10 spread for CFSALES (CFBE)-sorted portfolios is 1.25% (0.92%) a month. Furthermore, as shown in the column D1:5 D6:10, the first half of portfolios, which have lower cashflow volatility, displays higher returns than the second half of the portfolios. The t-statistics for all D1:5 D6:10 return spreads are significant. For the rest of the paper I choose to present value-weighted returns. Using simple average returns instead would not change the results. The next two rows of Panels A and B, labeled CAPM alpha and FF-4 alpha, report respectively the alphas of the valueweighted portfolios relative to the CAPM and the Fama French four-factor model. The return spreads are even stronger after I control for these risk factors. In Panel A, the D1 D10 spread increases to 1.64% a month after controlled for the market return and slightly decreases to 1.19% after controlled for the FF-4 factors. In Panel B, the D1 D10 spread increases to 0.87% and 1.06% respectively. All these D1 D10 alpha spreads are highly significant. The D1:5 D6:10 spreads increase similarly in both panels and are highly significant as well. A further examination of individual portfolios reveals that although statistically significant alphas cluster in portfolios with large cashflow volatility, the value of alpha is generally decreasing in cashflow volatility decile. These results suggest that returns on cashflow volatility sorted portfolios are not attributable to the Fama French risk factors. In the next row I report the Sharpe ratio of each portfolio. The Sharpe ratio is defined as the mean to the standard deviation of the excess return. This ratio can be interpreted as a measure of total risk-adjusted return. Confirming the results of previous rows, the Sharpe ratio decreases with cashflow volatility in both panels. The D1 D10 spread is 0.24 in Panel A and 0.13 in Panel B. These Table 3 Returns on the ten portfolios sorted on cashflow volatility. Decile 1 (S) (L) D1 D10 D1:5 D6:10 Panel A: Portfolios sorted on CFSALES Value-weighted ret. (%) [3.78] 0.46 [3.12] Simple average ret. (%) [3.70] 0.59 [3.71] CAPM 0.14 [1.40] [] [ 0.45] 0.14 [1.34] 0.01 [0.12] 0.12 [ 1.1] 0.14 [ 1.11] 0.44 [ 3.31] 0.33 [ 1.58] 1.51 [ 4.98] 1.64 [4.84] 0.56 [3.92] FF-4 [1.07] 0.00 [ ] 0.09 [0.67] 0.15 [1.31] [0.67] 0.14 [ 1.2] 0.12 [ 0.91] 0.31 [ 2.46] 0.16 [ 0.82] 1.08 [ 4.46] 1.19 [4.27] 0.44 [3.58] Sharpe ratio #offirms Market weight (%) Panel B: Portfolios sorted on CFBE Value-weighted ret. (%) [2.63] 0.19 [1.81] Simple average ret. (%) [3.58] 0.40 [3.23] CAPM [ 0.18] [ 1.36] [1.24] [ 0.3] [0.32] [ 0.73] 0.16 [1.31] [ 0.39] 0.57 [ 3.38] 0.89 [ 4.65] 0.87 [3.76] 0.28 [2.79] FF [1.24] [ 0.65] 0.09 [1.029] 0.39 [ 0.34] [ 0.33] [ 0.72] 0.19 [1.543] [ 0.54] 0.58 [ 3.88] 0.91 [ 5.38] 1.06 [4.91] 0.25 [3.65] Sharpe ratio #offirms Market weight (%) This table reports returns on 10 monthly portfolios formed with all stocks based on increasing cashflow volatility breakpoints. The breakpoints are determined by the ranked values of cashflow volatility of all stocks in the sample and are updated month by month. CFSALES (CFBE) is the standard deviation of cash-flow to sales (industry-adjusted cashflow to book equity) over the past 16 quarters. CAPM alpha and FF-4 alpha are Jensen's alpha relative to, respectively, CAPM and Fama French four-factors of market, SMB, HML and momentum. Sharpe ratio is the mean portfolio excess return divided by its standard deviation. D1 D10 is the spread between decile 1 and decile 10, and D1:5 D6:10 is the spread between the mean of deciles 1 5 and the mean of deciles Numbers in square brackets are t- statistics.,, and indicate significance at the 1%, 5% and 10% levels, respectively.

7 A.G. Huang / Journal of Empirical Finance 16 (2009) spreads are not trivial: As a benchmark for comparison, the Sharpe ratio of the S&P 500 index return for the same period is If we assume independence in portfolio return distribution, the D1 D10 spread in the Sharpe ratio indicates that a zero-investment strategy that goes long in decile 1 stocks and short in decile 10 stocks produces a total-risk adjusted return comparable to the S&P 500 index! The findings identified in Table 3 largely hold for individual years. Fig. 1 depicts the mean D1 D10 return spread averaged across each year from 1980 to Panel (a) shows the D1 D10 spread of portfolios sorted on CFSALES, and Panel (b) shows the spread Fig. 1. Monthly D1 D10 return spread averaged across year.

8 416 A.G. Huang / Journal of Empirical Finance 16 (2009) on portfolios sorted on CFBE. Although the spread is not always positive, we observe that the frequency of years with positive spread roughly triples the frequency of years with negative spread, and that those negative spreads are generally small in magnitude. These patterns suggest that the findings in Table 3 are not driven by one or multiple outlier years of disparately large, positive spreads Characteristics of cashflow volatility sorted portfolios I now examine the properties of the decile portfolios sorted on cashflow volatility. I first look at the market weight in each decile portfolio. The last two rows of Panels A and B of Table 3 provide the average number of firms in each decile and its market weight. By definition, the firm number is constant at around 300 for each decile. The portfolio market weight decreases with decile, with decile 10 portfolio accounting for 1.63% of the market weight for CFSALES and 1.86% for CFBE. By comparison, Ang et al. (2006) form five portfolios sorted on idiosyncratic return volatility. Their smallest idiosyncratic return volatility quintile accounts for 1.9% of the market share. Furthermore, the second half of the portfolios, which has significantly lower return than the first half of the portfolios, accounts for about 25% of the market share. Based on the weight of the second half of the cashflow volatility portfolios, the cashflow volatility effect uncovered in Table 3 does not seem to belong to only marginal firms. I next show the standard return-informative variables of these deciles. Fig. 2 depicts the value-weighted means of size, book to market, price momentum, illiquidity, and earnings momentum for each decile portfolio. In both CFSALES and CFBE sorted portfolios, high cashflow volatility is associated with low market value and SUE, and high price momentum and illiquidity. The association between book to market and cashflow volatility in these decile portfolios is sensitive to the proxy for cashflow volatility, as shown in Panel (b) that the portfolio book to market is somewhat increasing in CFSALES but decreasing in CFBE. Fig. 2 just reveals that the inverse relation between cashflow volatility and returns documented in Table 3 is largely unrelated to some existing empirical regularities. From the figure, firms whose cashflow are highly volatile are small, illiquid and with a strong price momentum. This set of firms is supposed to have higher returns. However, Table 3 shows that returns on these decile portfolios decrease with cashflow volatility. Hence, Fig. 2 suggests that the cashflow volatility effect is different from the size, price momentum and illiquidity effects. The portfolio properties that may subsume the cashflow volatility effect is earnings momentum. This is because decreasing return is associated with both decreasing earnings momentum and increasing cashflow volatility implying that earnings momentum and cashflow volatility is inversely related, which is confirmed in Fig. 2. Book to market equity of the CFBE sorted portfolios may also be treated similarly, as high CFBE portfolios display low book to market ratios and thus low returns. Nonetheless, in the next section I formally show that the cashflow volatility effect is not driven out by any of these control variables Robustness to double sorts To show that the results in Table 3 are not subsumed by a control variable, I construct 5 5 portfolios sorted first on the control variable and then on cashflow volatility. At the first step, I group stocks into 5 quintile portfolios based on the control variable. At the second step I further sort each such quintile portfolio into 5 quintile portfolios based on cashflow volatility. The sorting is updated every month. If the cashflow volatility effect is capturing the effect due to the control variable, then the risk-adjusted return spread among various cashflow volatility quintile portfolios inside each control quintile should not be different from zero. Panel A of Table 4 reports the FF-4 alphas on the 5 5 portfolios first sorted by size then by cashflow volatility. The column V1 V5 shows the alpha spread between the portfolio with the least volatile cashflow and the portfolio with the most volatile cashflow inside each size quintile. In each size quintile, the highest cashflow volatility quintile has substantially lower FF-4 alpha than the lowest size quintile. The V1 V5 spread is statistically significant at below 1% level for all but the largest stocks. Hence, it is not small stocks that are driving the results. The overall effect of controlling for size is shown in the row labeled Controlling for Size, where I average across the five size quintiles to produce cashflow volatility quintile portfolios. Each of these cashflow volatility quintile covers all sizes of firms. After controlling for size, the overall V1 V5 alpha spread is still high at 0.89% (0.94%) a month for CFSALES (CFBE) with a significance level of 1%. To put the overall significance of the size-controlled alpha spread into perspective, the last row of Panel A, labeled Unconditional CFV alpha, shows the unconditional FF-4 alpha spread of cashflow volatility quintile portfolios when no control is imposed. In comparison, the unconditional FF-4 alpha spread is 0.55% (0.74%) for CFSALES (CFBE). Thus, Panel A shows that controlling for size does not weaken the overall significance of the cashflow volatility effect. The rest of Table 4 repeats the same exercise for book-to-market, SUE, PMOM and ILLIQ. In all of these controls, the cashflow volatility effect holds up well, with the overall V1 V5 alpha spreads in the same order of magnitude as the size-controlled spread. An examination of individual control quintile reveals that the cashflow volatility effect spreads to firms differing widely in book to market, price momentum and illiquidity, as for these controls, the cashflow volatility effect is present within almost all of the control quintiles. One noteworthy finding in Table 4 is that while the overall effect of cashflow volatility holds after controlling for SUE, the effect is concentrated in medium to low (often negative) SUE stocks, as shown in Panel C that the alpha spread is significantly positive only in the smallest quintile of SUE for CFSALES and in the smallest two quintiles of SUE for CFBE. In other words, low earnings surprise firms must have some property that is both related to cashflow volatility and able to induce lower future returns that high earnings surprise firms lack. One possible explanation is persistence of earnings surprise. In untabulated results, I find that among the lowest SUE firms, the persistence of earnings surprise increases with cashflow volatility. That is, among firms with negative

9 A.G. Huang / Journal of Empirical Finance 16 (2009) Fig. 2. Value-weighted properties of the 10 decile portfolios sorted on cashflow volatility. past earnings surprises, those with past high cashflow volatility is more likely to continue to have negative earnings surprises in the future than those with low cashflow volatility, leading to lower future returns of the high cashflow volatility firms. In contrast, firms in other SUE quintiles have stable or decreasing persistence of earnings surprise along cashflow volatility quintiles.

10 418 A.G. Huang / Journal of Empirical Finance 16 (2009) Table 4 FF-4 alphas of 5 5 portfolios sorted first on a control variable and then on cashflow volatility. Panel A: Controlling for size Size quintile CFSALES quintile CFBE quintile V1 (S) V2 V3 V4 V5 (L) V1 V5 V1 (S) V2 V3 V4 V5 (L) V1 V5 1 (S) 0.77 [3.86] 0.47 [2.18] [0.19] 0.22 [ 0.76] 0.28 [ 0.83] 1.05 [3.80] 0.71 [3.52] 0.73 [3.32] [0.70] 0.15 [ 0.55] 0.60 [ 1.9] 1.31 [5.75] [3.24] 0.21 [1.29] [ 0.41] 0.29 [ 1.42] 1.01 [ 3.6] 1.49 [5.40] 0.49 [3.12] 0.22 [1.35] [0.12] 0.47 [ 2.39] 1.02 [ 4.08] 1.51 [7.96] [1.81] 0.26 [2.22] 0.16 [1.18] [ 0.81] 0.79 [ 3.57] 1.03 [3.93] 0.31 [2.97] 0.20 [1.70] 0.00 [ ] [ 1.26] 0.58 [ 3.25] 0.88 [5.27] [1.66] 0.18 [1.5] 0.08 [0.66] [ 0.87] 0.51 [ 2.75] 0.72 [2.76] 0.21 [1.96] [1.71] 0.14 [1.39] [ 0.44] 0.68 [ 5.98] 0.89 [6.17] 5 (L) [0.32] [0.26] [1.27] 0.15 [ 1.47] 0.14 [ 1.07] [0.99] 0.13 [1.06] [ 0.73] [0.27] [ 1.44] [0.33] 0.10 [0.63] Controlling for size 0.34 [3.67] 0.23 [2.49] [0.67] [ 1.53] 0.55 [ 3.06] 0.89 [4.46] 0.37 [4.37] 0.25 [2.89] [0.77] 0.19 [ 1.8] 0.57 [ 4.1] 0.94 [8.78] Unconditional CFV alpha [0.46] 0.10 [1.12] [ 0.58] 0.19 [ 1.82] 0.51 [ 2.32] 0.55 [2.54] [0.52] [0.57] [ 0.79] 0.09 [0.91] 0.70 [ 5.56] 0.74 [4.70] Panel B: Controlling for book-to-market (BM) BM quintile CFSALES quintile CFBE quintile V1 (S) V2 V3 V4 V5 (L) V1 V5 V1 (S) V2 V3 V4 V5 (L) V1 V5 1 (S) [0.26] 0.15 [1.20] 0.27 [ 1.51] 0.42 [ 1.82] 1.29 [ 4.28] 1.32 [3.91] [0.53] 0.23 [1.64] [ 0.33] 0.66 [ 3.57] 1.12 [ 5.19] 1.19 [4.78] 2 [0.14] 0.15 [ 0.95] 0.25 [ 1.57] 0.56 [ 3.36] 0.60 [ 2.2] 0.62 [1.99] [ 0.91] [0.13] 0.31 [ 1.91] [] 1.16 [ 5.4] 1.03 [4.40] [] 0.28 [ 1.64] 0.24 [ 1.44] [ 0.76] 0.80 [ 3.71] 0.80 [3.07] 0.19 [ 1.24] 0.15 [ 0.9] 0.37 [ 2.69] 0.08 [ 0.45] 0.16 [ 0.77] [ 0.16] [1.60] 0.10 [0.67] 0.16 [ 1.09] [ 0.28] [ ] 0.29 [0.99] [1.19] [ 0.37] [0.71] 0.00 [] [] 0.13 [0.50] 5 (L) 0.47 [2.57] 0.81 [4.43] 0.72 [3.26] 0.43 [1.85] 0.26 [1.04] 0.22 [0.72] 0.73 [3.84] 0.76 [3.85] 0.64 [3.32] 0.48 [2.10] 0.10 [ 0.38] 0.83 [2.44] Controlling for BM 0.16 [1.72] 0.13 [1.54] [ 0.51] 0.15 [ 1.41] 0.49 [ 3.18] 0.65 [3.33] 0.13 [1.55] 0.16 [2.24] 0.00 [] [ 0.56] 0.50 [ 4.08] 0.63 [4.43] Panel C: Controlling for SUE SUE quintile CFSALES quintile CFBE quintile V1 (S) V2 V3 V4 V5 (L) V1 V5 V1 (S) V2 V3 V4 V5 (L) V1 V5 1 (S) 0.59 [ 3.17] 0.38 [ 2.01] 0.76 [ 3.97] 1.65 [ 7.79] 2.08 [ 7.37] 1.49 [4.49] 0.28 [ 1.67] 0.86 [ 4.69] 0.93 [ 4.63] 1.81 [ 7.7] 2.80 [ 11.8] 2.52 [9.02] [ 2.59] 0.40 [ 2.64] 0.52 [ 2.99] 0.38 [ 2.03] 0.64 [ 2.3] 0.21 [0.62] 0.47 [ 3.11] 0.20 [ 1.3] 0.26 [ 1.4] 0.93 [ 4.92] 1.54 [ 6.89] 1.07 [4.02] [ 0.49] 0.12 [0.69] [ 0.14] 0.35 [ 2.02] [0.12] 0.12 [ 0.38] [] [ 0.23] 0.14 [ 0.69] 0.25 [1.34] 0.33 [ 1.39] 0.31 [1.08] 4 [0.70] [1.00] [0.49] 0.49 [2.73] 0.29 [ 1.08] 0.40 [1.20] [0.68] 0.12 [0.73] 0.20 [1.28] 0.41 [2.02] [0.82] [0.23] 5 (L) 0.50 [3.51] 0.56 [3.38] 0.51 [2.98] 0.50 [3.08] 0.20 [0.84] 0.30 [1.04] 0.56 [3.58] 0.50 [3.89] 0.48 [2.82] 0.75 [4.41] 0.53 [2.66] [0.12] Controlling for SUE 0.10 [ 1.2] 0.01 [0.18] 0.14 [ 2.08] 0.28 [ 2.86] 0.56 [ 3.19] 0.46 [2.20] [ 0.28] 0.10 [ 1.58] [ 1.61] 0.27 [ 3.1] 0.80 [ 6.73] 0.77 [5.36] Panel D: Controlling for price momentum (PMOM) PMOM CFSALES quintile CFBE quintile quintile V1 (S) V2 V3 V4 V5 (L) V1 V5 V1 (S) V2 V3 V4 V5 (L) V1 V5 1 (S) 0.97 [4.43] 0.46 [1.69] [0.08] 0.37 [ 1.11] 1.07 [ 3.31] 2.04 [5.22] 1.14 [4.59] 0.57 [2.24] [ 0.46] 0.65 [ 2.25] 1.35 [ 3.69] 2.49 [6.07] [2.84] 0.44 [2.32] 0.18 [0.88] 0.10 [0.47] 0.91 [ 3.71] 1.38 [4.33] 0.53 [3.12] 0.41 [2.23] 0.14 [0.721] 0.20 [ 0.99] 0.75 [ 3.16] 1.28 [4.20] [0.56] [] [0.47] 0.24 [ 1.53] 0.70 [ 3.53] 0.77 [3.26] [0.30] [0.13] 0.39 [ 2.54] [0.15] 0.49 [ 2.56] 0.53 [2.30] [0.52] [ 0.14] 0.08 [ 0.61] 0.25 [ 1.59] 0.45 [ 2.31] 0.53 [2.15] [0.16] 0.19 [ 1.24] [ 0.81] [ 0.15] 0.30 [ 1.65] 0.32 [1.37] 5 (L) [ 0.18] [ 0.39] [ 0.16] 0.48 [ 2.33] 0.49 [ 1.88] 0.46 [1.38] 0.14 [ 0.74] 0.23 [ 1.4] 0.24 [ 1.61] 0.12 [0.66] 0.61 [ 2.71] 0.47 [1.68] Controlling for PMOM 0.31 [3.44] 0.16 [1.69] [0.36] 0.25 [ 2.22] 0.72 [ 4.8] 1.04 [5.47] 0.32 [3.37] [1.31] 0.15 [ 1.73] 0.15 [ 1.52] 0.70 [ 5.9] 1.02 [7.13]

11 A.G. Huang / Journal of Empirical Finance 16 (2009) Table 4 (continued) Panel E: Controlling for illiquidity (ILLIQ) ILLIQ quintile CFSALES quintile CFBE quintile V1 (S) V2 V3 V4 V5 (L) V1 V5 V1 (S) V2 V3 V4 V5 (L) V1 V5 1 (S) [0.25] 0.05 [0.43] 0.10 [1.10] 0.15 [ 1.39] 0.15 [ 1.07] 0.18 [0.98] [1.41] 0.12 [ 1.24] [0.71] [ 0.44] [ 1.08] 0.28 [1.68] [1.47] [0.59] 0.05 [0.46] 0.33 [ 3.02] 0.42 [ 2.15] 0.60 [2.26] 0.09 [0.82] [ 0.24] [0.31] [ 0.51] 0.59 [ 5.01] 0.68 [4.47] 3 [ 0.25] 0.18 [1.40] 0.16 [1.20] 0.33 [ 2.35] 0.83 [ 3.69] 0.80 [2.95] [0.88] 0.05 [0.40] 0.01 [] 0.44 [ 3.69] 0.73 [ 4.27] 0.84 [4.41] [0.99] [ 0.32] 0.26 [ 1.63] 0.40 [ 2.14] 1.27 [ 4.74] 1.42 [4.66] 0.00 [ ] 0.10 [ 0.63] 0.16 [ 1.06] 0.54 [ 3.24] 0.93 [ 4.00] 0.93 [4.08] 5 (L) 0.20 [1.19] 0.16 [ 0.84] 0.41 [ 1.99] 0.73 [ 3.03] 1.33 [ 4.34] 1.53 [5.07] [ ] [ 0.24] 0.32 [ 1.67] 0.71 [ 3.02] 1.36 [ 4.88] 1.33 [5.29] Controlling for ILLIQ [1.25] [0.22] [ 0.85] 0.39 [ 3.88] 0.80 [ 4.72] 0.90 [4.45] [0.90] [ 0.61] 0.08 [ 0.97] 0.36 [ 4.22] 0.75 [ 6.03] 0.81 [6.89] This table reports the FF-4 alphas of 5 by 5 portfolios formed with all stocks in the sample sorted first by a control variable and then by cashflow volatility. Based on increasing value of the control variable, firms are first broken into 5 control quintiles. Each control quintile is then further broken into 5 quintiles based on increasing value of CFSALES (the standard deviation of cashflow to sales over the past 16 quarters) or CFBE (the standard deviation of the industry- adjusted cashflow to book equity over the past 16 quarters). The breakpoints are determined by all stocks and are updated every month. In each panel, the detailed 5 5 alphas are first reported. At the end of each panel, the five cashflow volatility quintiles are averaged across each control quintile so that each cashflow volatility quintile contains all values of the control variable. The control variables are size, book to market equity, standardized unexpected earnings (SUE), past twelvemonth return (PMOM), and illiquidity (ILLIQ). In Panel A, the row Unconditional CFV alpha reports the unconditional FF-4 alpha on the cashflow volatility-sorted quintile port- folios without any control. t-statistics are in square brackets.,, and indicate significance at the 1%, 5% and 10% levels, respectively. In summary, the evidence in this section shows a strong, negative association between cashflow volatility and return at the portfolio level. This relationship cannot be accounted for by a number of known factors, including size, value, price momentum, earnings momentum and illiquidity. By following a mechanical, zero-cost investible strategy of going long in stocks in cashflow volatility decile 1 and short in stocks in cashflow volatility decile 10, one could have generated value-weighted returns of 9 15% a year or alpha returns of 13% a year during The cashflow volatility effect at the firm level In this section, I explore the cashflow volatility effect at the firm level using regressions that allow for multiple controls. I follow the standard approach in Fama and MacBeth (1973) and Fama and French (1992) to run the two-step cross sectional regressions. I keep each stock's β, size and book-to-market equity as the original Fama and French (1992) variables. I augment this set of variables with previously mentioned return-informative variables and cashflow volatility. I therefore run the following crosssectional regression for every month t: R i t = αi t + γ 1;t βi t + γ 2;t ln ð ME Þi t + γ 3;tln ð BEME Þi t + γ 4;t PMOMi t + γ 5;t SUEi t + γ 6;t ILLIQ i t + γ 7;t EYi t + γ 8;t Cashflow Volatilityi t + i t ; ð3þ where sup-i indexes stock i, sub-t indexes time, R is the raw return, ME is lagged market equity, BEME is book-to-market equity, PMOM is price momentum, SUE is standardized unexpected earnings, ILLIQ is illiquidity, and EY is earnings yield. The time-series of regression coefficients are then averaged to compute the mean estimates and the associated t-statistics. To map Fama and French (1992) accurately, I use their definitions of β, ME and BEME. Specifically, ln(me) is measured as the logarithm of market equity of June of the latest fiscal year, ln(beme) is measured as the logarithm of book equity of the latest fiscal quarter divided by market equity of December of the same year, and β is estimated with the Fama MacBeth two-pass procedure. Note that with the exception of β, the value of every other right hand side variable in Eq. (3) is known prior to the stock transaction. I also replace the forward-looking β in Eq. (3) with historical beta measured from the previous four years of excess return, and find that the results are similar. Table 5 presents the results for different sets of regressors in Eq. (3). Model 0 provides the benchmark Fama and French (1992) results, with price momentum as an additional regressor. Models 11 and 12 add CFSALES and CFBE respectively. Models 21 and 22 further add earnings momentum. Finally, models 31 and 32 use the full specification in Eq. (3). Most tellingly, Table 5 shows consistent evidence that cashflow volatility negatively predicts expected stock returns at the firm level, regardless of model specification. In every regression that I estimate, cashflow volatility loads negatively with significant t- statistics. In models 11 and 12, the coefficient of cashflow volatility is highly significant when it is combined with the traditional four factors of beta, market, book-to-market and price momentum. The cashflow volatility effect is somewhat weakened with the addition of SUE, ILLIQ and EY but still holds. The loading of cashflow volatility is economically meaningful. I emphasize the results in models 11 and 12. In model 11, the loading of CFSALES is. Over the full sample, the mean of CFSALES is 1.74 and the standard deviation is A loading would give rise to an average return of 1.74 or 0.10% of stock return every month. And if CFSALES increases by one standard deviation, return decreases by 7.82 or about 0.47% a month. The results with CFBE are similar: the loading of CFBE is The full sample mean of CFBE is. This means on average, CFBE explains 0.20% of stock return a month.