Upside and Downside Components of Cash Flow Volatility: Implications for Corporate Policies* John Easterwood 1, Bradley Paye 1, and Yutong Xie 1 1 Pamplin College of Business, Virginia Tech, Blacksburg, VA 24061, USA January 16, 2017 JEL Classification: G1, C13 Keywords: Corporate policies, leverage, cash holdings, payout, dividends, repurchases. Abstract A large literature hypothesizes that corporate financial policies depend on the stochastic properties of future cash flows. We decompose traditional measures of total cash flow uncertainty into downside and upside components, and assess whether these components differentially impact three key corporate policies: 1) cash holdings; 2) leverage; and 3) corporate payout. We reject the hypothesis that total variation adequately captures future cash flow risk as it relates to these policies. In contrast, we find that cash holdings depend positively on the downside component of cash flow volatility, but negatively on the upside component. The reverse relations hold for leverage ratios. The marginal effect of downside uncertainty is economically and statistically stronger than that for upside uncertainty for both policies. While downside uncertainty is important for all payout decisions we investigate, upside uncertainty relates only to dividend increases and decreases. Our results indicate that cash flow characteristics beyond variance are important for a variety of corporate policy decisions. 1 Corresponding author. *We are grateful for comments and suggestions from Gregory Kadlec and Vijay Singal. Yutong Xie: yxie@vt.edu; John Easterwood: jceast@vt.edu; Bradley Paye: bpaye@uga.edu.
Upside and Downside Components of Cash Flow Volatility: Implications for Corporate Policies* JEL Classification: G1, C13 Keywords: Corporate policies, leverage, cash holdings, payout, dividends, repurchases. Abstract A large literature hypothesizes that corporate financial policies depend on the stochastic properties of future cash flows. We decompose traditional measures of total cash flow uncertainty into downside and upside components, and assess whether these components differentially impact three key corporate policies: 1) cash holdings; 2) leverage; and 3) corporate payout. We reject the hypothesis that total variation adequately captures future cash flow risk as it relates to these policies. In contrast, we find that cash holdings depend positively on the downside component of cash flow volatility, but negatively on the upside component. The reverse relations hold for leverage ratios. The marginal effect of downside uncertainty is economically and statistically stronger than that for upside uncertainty for both policies. While downside uncertainty is important for all payout decisions we investigate, upside uncertainty relates only to dividend increases and decreases. Our results indicate that cash flow characteristics beyond variance are important for a variety of corporate policy decisions. 1
1. Introduction This study investigates the impact of alternative measures of cash-flow uncertainty on corporate financial policies. In contrast to prior empirical work that focuses on total uncertainty, we decompose uncertainty into upside and downside components to determine if these aspects of uncertainty have different impacts on corporate actions. We study three corporate policies cash holdings, leverage, and payout. Our evidence suggests that corporate policies do respond differently to up and downside uncertainty and that downside uncertainty has a larger impact. Existing literature provides abundant evidence linking measures of the variance of future cash flows with various corporate policies. For example, previous studies find that firms with high total cash flow uncertainty hold more cash (Opler, et al 1999; Han and Qiu 2007), use lower leverage (Lemmon, Roberts, and Zender 2008; Baum et al, 2013), and are less likely to payout (Rozeff 1982; Hoberg and Prabhala 2009; Chay and Suh 2009; Walkup 2016). These studies implicitly interpret the variance of future cash flow as an adequate measure of cash flow risk. However, the appropriateness of variance as a risk measure has been questioned since at least Roy (1952) and Markowitz (1959). A large subsequent asset pricing literature emphasizes downside measures of risk (e.g., Ang, Chen, and Xing (2006)). In addition, financial theory recognizes that optimal investment decisions depend in general upon various aspects of return uncertainty, including, for example, potential skewness in returns. Although it has received little attention to this point, the distinction between downside and total variation in cash flow is likely to be important for corporate policies. Theoretical arguments emphasize the role of potential cash shortfalls in linking cash flow risk with various policies. In discussing the precautionary motive for cash holdings, for example, Bates, Kahle, and Stulz (2009) 1
note that Firms hold cash to better cope with adverse [cash flow] shocks when access to capital markets is costly (pp. 1988-1989). Potential low cash flow outcomes are similarly relevant for the servicing of debt and therefore in capital structure choice, as well in considerations regarding firm payout policies. Broadly, theory suggests that firms with higher future cash flow risk should adopt more conservative financial policies. Despite the importance of potential cash shortfalls, it is possible that variance fully captures relevant aspects of cash flow risk. This would be the case, for example, if future cash flows follow a normal distribution. In reality, the distribution of firm cash flows is likely to depart markedly from the normal. This suggests that downside cash flow variance may be a more appropriate risk measure than traditional variance, and that other stochastic features of cash flows (e.g., skewness) may be important for corporate policies. Empirically, we construct alternative measures of the variability of future firm cash flows. These include more traditional proxies based on the variance of past returns, and alternative proxies that decompose the total variance into downside and upside components. Our econometric tests include both downside and upside variation measures in standard panel data models for three key corporate policies: 1) cash holdings; 2) leverage choice; and 3) corporate payout decisions including dividend and share repurchases. We then test several key hypotheses of interest. The first is that total cash flow variation represents the correct risk measure. Under this hypothesis, slope coefficients for downside and upside cash flow variation measures should be identical. An alternative hypothesis asserts that downside cash flow variation is the correct risk measure. In this case, upside cash flow variation should play no role conditional on downside variation. In other words, the slope coefficient on this component should be equal to zero. The true distribution of future cash flows is unobserved. Thus, researchers typically construct measures based on the volatility of recent past performance, which implicitly assumes 2
that future volatility is similar to recent volatility. To measure the volatility of cash flows at the firm level requires accounting data for a sufficient number of periods. Since data is only available at quarterly or annual frequencies, use of accounting data necessitates a longer time-period for estimating risk. Longer time-periods are problematic if risk changes through time. To avoid using long windows to estimate risk, some researchers (e.g., Chay and Suh (2009) and Hoberg and Prabhala (2009)) use stock return volatility to measure cash flow volatility. We measure risk using the volatility of monthly stock returns over the prior 3 years. Additionally, we split the data into the upside and downside semi-deviations. Our procedure assumes that cash flow uncertainty is equally or more important than the volatility of discount rates as a cause of firm-level return volatility. Vuolteenaho (2002) provides evidence that cash flow news is the more important driver of firm-level stock returns. We summarize our empirical findings as follows. Cash holdings depend positively on the downside component of cash flow volatility and negatively on the upside component. The reverse relations hold for corporate leverage ratios: firms with high downside uncertainty use less debt financing and firms with high upside uncertainty use more debt financing. The marginal effect of downside uncertainty is economically and statistically stronger than for upside uncertainty for both policies. For both cash holdings and leverage choices, we resoundingly reject the null hypothesis that slope coefficients on downside and upside variation measures are equal. This constitutes strong evidence against the notion that total variance is the appropriate risk measure for these policies. Interestingly, we also find upside variation to be relevant for both policies, but in a direction that does not suggest interpretation of such variation as risk per se. A potential interpretation is that more nuanced aspects of uncertain cash flows, such as skewness, prove relevant for corporate policy decisions. 3
Turning to corporate payout, we find a strong negative relation between propensity to pay out and the downside component of cash flow volatility. In contrast, we do not find a significant relation between payout and upside cash flow volatility. We also investigate other dimensions of payout policy. We find that downside uncertainty negatively affects the propensity to pay cash flow to shareholders, to pay dividends to common shareholders, to initiate dividends and positively affects propensity to omit dividends. Upside uncertainty does not affect the decision to payout cash flow, to be a dividend payer, to initiate or omit a dividend. However, we find that upside uncertainty positively affects dividend increases and negatively affects dividend decreases and find the reverse relationship for downside uncertainty. One interpretation of these results is that firms are more conservative when starting to pay by only considering downside uncertainty. Firms seek to ensure that their earnings can support future dividends and they do not want to rely on the uncertain cash flow increases. When firms are having difficulty paying dividends, they do not want to rely on uncertain cash flow increases either. Once firms are paying, they consider both downside and upside uncertainties when contemplating alterations of the dividend level. Additionally, we investigate firms choice to payout via share repurchases only in a particular year as opposed to conducting other types of payout activity (a firm that pays a dividend only or a firm that both pays a dividend and repurchases shares). 2 Firms with higher downside and upside uncertainties are more likely to repurchase shares but not pay a dividend compared to other methods of payout. Repurchases are considered as more flexible than dividends and firms tend to payout temporary cash flows with repurchases (Jagannathan, Stephens, and Weisbach, 2000; Guay 2 We analyze firms that repurchase only because we want to distinguish firms that are repurchasing from firms that are both repurchasing and paying dividends or firms that only pay dividends. Many current dividend payers also repurchase. If we do not distinguish these two groups, the relation we find may be a mixed effect of dividend payers and repurchasers. 4
and Harford, 2000; Skinner, 2008; Floyd, Li, and Skinner, 2015; Kahle and Williams, 2016). Hence, when firms have higher downside uncertainty and they decide to payout, they are more likely to choose repurchases over dividends because of the flexibility advantage of repurchases. When firms have higher upside uncertainty and they decide to payout, they are more likely to choose to repurchase as well, because upside uncertainty does not represent a high stable level of cash flow but represent high chance of cash flow increases. Firms will only choose dividends to pay out cash flows when they are certain about future cash flow levels and have low level of uncertainty in both up and down directions. In summary, our findings strongly suggest that multiple aspects of future cash flow uncertainty are relevant for firms financial policies. We resoundingly reject the hypothesis that total cash flow variation adequately captures risk in the context of firm decisions regarding cash holdings, leverage, and corporate payout. Indeed, we find that downside and upside uncertainties affect different aspects of payout policy. The investments literature has recognized the limitations of variance as a risk measure, and acknowledged the distinction between upside and downside variation (Roy, 1952; Markowitz, 1959; Ang, Chen, and Xing, 2006). In contrast, few studies in the corporate finance literature distinguish between downside uncertainty, upside uncertainty, and total uncertainty. Almeida et al (2014) argue that cash holdings are more important when firms are more likely to evolve to bad states in a binomial outcome model but ignore upside uncertainty. Avramov, Li, and Wang (2016) use a text-based variable that captures managerial perception of downside possibility characterizing firm s fundamentals and find that this variable predicts various corporate policies but also ignore upside uncertainty. We differ from these papers by looking at the impacts from both upside and downside uncertainty and investigating the difference in magnitudes of the uncertainties. 5
The organization of this paper is as follows. The next section describes the data, presents hypotheses, and defines key variables. Section 3 describes our methodology and the relevant econometric issues. Section 4 conducts empirical analyses on cash and leverage. Section 5 presents empirical results on multiple dimensions of payout policy. Section 6 concludes the paper. 2. Hypotheses, variable definitions, and data 2.1 Data We begin with all firm-year observations in Compustat. To ensure that firms are publicly traded, the Compustat includes only firms with share codes of 10 or 11, and we use only the fiscal years a firm in the CRSP database at its fiscal year-end. We exclude small firms with book equity (BE) below $250,000 or assets (A) below $500,000 according to 2014 dollar. We also exclude utilities (SIC Codes 4900-4949) and financial firms (SIC Codes 6000-6999) from both samples. We further require all the variables to be non-missing in our regression models. We winsorize all Compustat variables at 1 st and 99 th percentiles. 2.2 Corporate Policies In this paper, we analyze three corporate policies cash holdings, leverage, and propensity to payout. Our proxy and our hypotheses for each policy are presented in this section. 2.2.1 Cash We use firms cash and short-term investment (CHE) standardized by lagged total asset (TA) as our cash measure. When firms have greater downside uncertainty, they are more likely to experience larger cash flow decreases or receive low cash flows. As a result, they are more likely 6
to have cash shortfalls. Hence, firms with high downside cash-flow uncertainty have more incentive to hold precautionary savings. On the other hand, firms with high upside uncertainty are more likely to have larger cash flow increases or receive higher future cash flows. Such firms have less incentives to of save. Hence, we hypothesize that downside uncertainty positively affects firms cash holding, and upside uncertainty negatively affects firms cash holding. 2.2.2 Leverage We use book value of debt divided by market value of total assets to measure leverage. Book value of debt is the sum of total long-term debt (DLTT) and debt in current liabilities (DLC). 3 Market value of total assets equals market value of equity plus book value of assets minus book value of equity (BE). Book value of equity (BE) is defined following Hoberg and Prabhala (2009). 4 Firms with high downside uncertainty tend to choose low leverage because firms commit to higher cash outflows by using more debt. On the other hand, firms with high upside uncertainty tend to use high leverage because firms can use the high potential cash inflows to support the required cash outflows from leverage. 2.2.3 Payout policy We identify a firm as a payer in a fiscal year if either its common dividends or share repurchases is greater than zero in that year. We identify a firm as a dividend payer in a fiscal year if its common dividends (DVC) is greater than zero in that year, and we identify a firm as a share 3 We also run our tests with total long-term debt (DLTT) as book value of debt or total liability (LT) as book value of debt. The results are qualitatively similar. 4 Book value of equity (BE) equals shareholder s equity (SEQ) minus preferred stock (PSTK) plus balance sheet deferred taxes and investment tax credit (TXDITC). If shareholder s equity is not available, it is replaced by either common equity (CEQ) plus preferred stock, or assets (AT) minus liabilities (LT). Preferred stock is preferred stock liquidating value (PSTKL), preferred stock redemption value (PSTKRV), or preferred stock par value (PSTK). 7
repurchaser in a fiscal year if it repurchases in that year. To identify stock repurchases, we use the purchase of common and preferred stock (PRSTKC) minus the purchase of preferred stock (PRSTKPC) 5. More generally, we identify a firm as a payer if it either pays dividends or repurchases in a year. We expect that downside uncertainty negatively affects propensity to payout, pay dividends, or repurchase. We further create 3 groups of payer firms based on the method of distributing cash to shareholders: firms that only pay a dividend (DO), firms that only repurchase shares (RO), and firms that both pay a dividend and repurchase shares (BO). These groups describe the mechanism for distributing cash. Banyi, Dyle, and Kahle (2008) find high error rates in commonly used repurchase estimators. Even the most accurate measure deviates from the actual number of shares repurchased by more than 30% in about 16% of the cases according to their sample. Hence, in un-tabulated robustness tests, we use all the measures discussed by Banyi, Dyle, and Kahle (2008) to identify repurchasers. The results are qualitatively the same. 2.3 Downside and Upside Uncertainty We consider a tradeoff between data availability and quality when choosing the proxies. Ideally, we should use the semi-deviations of cash flows. However, we will need a long time-series to have enough observations to obtain less noisy estimates of both semi-deviations, which will potentially introduce a survival bias. We do not use weekly or daily stock return data to mitigate microstructure problems (see for example, Scholes and Williams, 1976; Lo and MacKinlay, 1990; 5 Both the purchase of common and preferred stock (PRSTKC) and the purchase of preferred stock (PRSTKPC) are cash flow statement items. If the purchase of preferred stock (PRSTKPC) is missing, we replace the missing observations with the reductions in the value of any preferred stock outstanding (PSTK). We further identify a firm as a payer if it either pays dividend, repurchases, or does both in a year. 8
Kadlec and Patterson, 1999; Han and Lesmond, 2011). Vuolteenaho (2002) decomposes firm level stock returns into two parts: changes in cash-flow expectations and changes in discount rates, and finds that firm-level stock returns are mainly driven by cash-flow news. Hence, we use stock returns semi-deviations as our proxies. The key variables in this study are downside and upside uncertainties of cash flows. We use semi-deviations of monthly stock returns over the most recent 36 months excluding the current fiscal year to measure downside and upside uncertainties. We use the following definitions of semi-deviations: T DN 1 T (min(r t B DN, 0)) 2 t=1 (1) where R t indicates an arbitrary firm s excess stock return and B DN is a specified benchmark return around which the downside semi-deviations is computed. We define an upside measure of cash flow uncertainty in similar fashion: T UP 1 T (max(r t B UP, 0)) 2 t=1 (2) We consider several possibilities for specifying the benchmark returns B DN and B UP. Our first set of measures set B DN = B UP = R, where R equals the mean excess return, 1 T T t=1 R t. A second set of measures excludes returns close to the mean, and identifies downside variation with negative returns. Specifically, this approach sets B DN = 0 and B UP = 2 R. We refer to this set of measures as modified semi-deviations. 9
2.4 Control Variables In addition to the aforementioned dependent variables and proxies for downside and upside uncertainty, we use a number of control variables in our analyses, including profitability, operating profitability, size, investment, market-to-book ratio, earned/contributed capital mix, etc. Detailed definitions for each control variable for each corporate policy are listed in the Appendix. Due to simultaneity concerns, all control variables are lagged one period in the regressions presented below. 2.5 Descriptive Statistics Table 1 presents the descriptive statistics of our dependent variables and risk measures. One observation is a firm-year. Panel A in Table 1 contain summary statistics for dependent variables. In our sample, 23% of the observations are dividend payers, 28.1% of the observations are repurchasers, and 39.6% of the observations are payers. Among the dividend payers, half of the observations are dividend only, and the other half both pay dividends and repurchase. There is wide variation in the ratio of cash and marketable securities to assets. The median firm holds cash equal to approximately 9.5% of total assets. This ratio is 2.6% at the 25 th percentile and 28.6% at 75 th percentile. The mean market leverage ratio is 11% in our sample, with a standard deviation of 17.4%. The variation in leverage is also high: 1.2% at the 25 th percentile and 26.3% at the 75 th percentile. Panel B in Table 1 reports our measures for upside and downside uncertainties. For both groups of measures, the mean of upside uncertainty is higher than the mean of downside uncertainty, and the standard deviation of upside uncertainty is lower. Upside uncertainty calculated using modified semi-deviation has a significantly lower number of observations, because we use 2 R as our upside threshold. 10
3. Econometric Methods Our empirical analysis follows a wide range of prior studies that link various corporate policies with firm characteristics. As with most of these studies, econometric inference occurs in a panel data environment. We focus on the relationship between alternative measures of future cash flow uncertainty and multiple corporate policies (cash holdings, leverage, and corporate payout). A large empirical literature exists with respect to each of these policies, and in each literature a variety of econometric concerns have been raised in prior studies. Some of the econometric concerns overlap, e.g., potential concerns regarding omitted variables bias and other forms of endogeneity. However, there remains considerable heterogeneity both within and across these literatures in terms of model specifications and inference techniques. We seek to adopt a relatively consistent approach to inference for the policies we examine, while remaining attentive to particular econometric concerns associated with each policy. The following provides a stylized representation of the panel data models we consider in order to frame the key econometric issues: Y i,t = λ t + ϕy i,t 1 + β 1 DN i,t + β 2 UP i,t + ψ X i,t + η i + ϵ i,t (3) where Y i,t represents a corporate policy measure of interest (e.g., a measure of leverage) for firm i in year t, λ t is a time-varying intercept, DN i,t and UP i,t are measures of downside and upside cash flow uncertainty that are of focal interest, X i,t denotes a set of additional control variables (potentially including lagged variables), η i denotes a (time-invariant) firm effect, and ϵ i,t are idiosyncratic shocks. The main coefficients of interest are β 1 and β 2. For comparative purposes, we sometimes consider a version of Eq. (3) in which a total cash flow 11
uncertainty measure (TU i,t ) appears in place of the variables DN i,t and UP i,t. The set of control variables X i,t included in each model depends on the specific corporate policy under analysis, and generally follows prior literature. The specification in Eq. (3) is obviously linear. This is convenient for discussing various econometric approaches to inference. Below we discuss extensions to various types of nonlinear specifications. A wide variety of approaches are available to estimate parameters in the panel model of Eq. (3). A key dimension along which the approaches differ involves the relative strength of the assumptions required in order for the estimator to be consistent. This is not the only relevant dimension for comparison; however, as more general estimators can be subject to other types of concerns (e.g., increased sensitivity to errors-in-variables). To explore and document robustness of key results, we apply several alternative approaches. We summarize these below, providing for each a brief synopsis of relative strengths and weaknesses: 1. Pooled OLS with Industry Fixed Effects: This approach estimates model parameters using pooled OLS. The estimator is consistent under the standard conditions for the consistency of OLS, i.e., when errors u i,t = η i + ϵ i,t are uncorrelated with the regressors. The model includes industry fixed effects (dummies) to provide additional control for potentially relevant differences across firms. We note two significant drawbacks to this approach. First, the assumption that η i is invariant within industry groups is strong. A violation of this assumption will generally lead to inconsistent estimates. A second objection is that pooled OLS requires the errors to be strictly exogenous. 6 This precludes simultaneity and feedback from shocks to included variables. It is important to construct appropriate standard errors under pooled panel approaches. To account for potential (unmodeled) firm effects, we cluster standard errors by firm. (Year effects are included in the model.) 6 Let Z i,t collect all variables in the model. The assumption of strict exogeneity can be stated as E[ε i,t Z i,1,, Z i,t ] = 0, t = 1,, T 12
2. Traditional Fixed Effects Estimation: This approach applies the traditional (static) fixed effects or within estimator. This approach accommodates time-invariant unobserved heterogeneity that is potentially correlated with observed covariates. There are at least two potential weaknesses of the fixed effects approach. First, consistency of the estimator requires the relatively strong assumption that firm variables are strictly exogenous conditional on thee unobserved firm effect. 7 This precludes lagged dependent variables and the realistic possibility of feedback between the dependent variable and explanatory variables. Second, Griliches and Hausman (1986) note that biases associated with errors-in-variables can be exacerbated under fixed effect estimators. This concern is relevant in our application, because the stochastic properties of future cash flows are unobserved and the proxies we rely upon are inherently subject to measurement error. In some cases, we also include the lag of the dependent variable, despite the fact that doing so results in the so-called Nickell bias (Nickell (1981)). 8 3. Dynamic Panel GMM Estimation: This approach amounts to a form of instrumental variables estimation using lagged dependent variables and lagged endogenous variables as instruments. In typical cases, the associated orthogonality conditions overidentify the parameters of interest, and estimation proceeds via a GMM approach (Holtz-Eakin, Newey, and Rosen (1988) and Arellano and Bond (1991)). Appealingly, the dynamic panel GMM approach offers the possibility of consistent estimation without requiring strict exogeneity as in static fixed effects approach. This permits the inclusion of lagged dependent variables as well as predetermined or endogenous variables. For example, when the errors ε i,t are uncorrelated, sufficiently lagged dependent endogenous variables serve as valid instruments in the first-differenced 7 This assumption of strict exogeneity conditional on η i can be stated as E[ε i,t Z i,1,, Z i,t, η i ] = 0, t = 1,, T 8 The bias documented by Nickell (1981) does not disappear asymptotically as N increases with T fixed. However, it does disappear as T becomes large. Although N > T in our application, T is sufficiently large that it is at least plausible to hope that the associated bias is not too severe. 13
model. 9 Unfortunately, the validity of using lagged dependent and endogenous variables as instruments is questionable in our setting. For example, Roberts and Whited (2013) critique the validity of lagged endogenous variables as instruments in the context of leverage models similar to those entertained in our paper. Even if the underlying moment conditions are valid, dynamic panel GMM estimators can be subject to a weak instrument problem, potentially leading to imprecise estimates. 4. Fama-MacBeth Approach: Many prior studies estimate coefficients in specifications similar to Eq. (3) using the Fama-MacBeth approach (Fama and MacBeth (1973)). There are at least two concerns with the Fama-MacBeth approach in our settings. First, Fama-MacBeth relies on a sequence of cross-sectional regressions. In the presence of unobserved heterogeneity that is correlated with variables of interest, the corresponding cross-sectional OLS estimates are generally inconsistent. Since this applies to each cross-section, the associated time-series averages are also inconsistent. To help address this concern, we implement the Fama-MacBeth approach using specifications that include a set of industry dummies. Consequently, the conditions under which this approach produces consistent estimates are similar to those discussed above for the pooled OLS estimator with industry fixed effects. Second, Petersen (2009) shows that Fama-MacBeth standard errors are downward biased in the presence a firm fixed effect. In light of these points, we focus on alternative inference approaches in the main paper. Nevertheless, the Internet Appendix shows that our main empirical results remain robust to using standard implementations of the Fama- MacBeth method. Linear panel data models of the form of Eq. (3) are very common in the empirical literature concerning the determinants of firm leverage and cash holdings. However, it is worth noting that the linear form is potentially, if not likely, misspecified in these contexts because the leverage and cash holdings variables of interest generally take the form of bounded ratios (Ramalho and da 9 Wintoki, Linck, and Netter (2012) discuss dynamic panel GMM approaches relative to the more traditional static fixed effects estimator with applications to corporate governance research. 14
Silva (2013)). It is also common to model corporate payout decisions (the decision to pay a dividend, repurchase shares, or both) using an indicator variable. In such cases, although the linear probability model might be applied, limited dependent variable models are prevalent in the empirical literature. Consequently, we estimate panel logit models in these cases, relying primarily on pooled estimation approaches with industry (and year) fixed effects included. 4. Cash and leverage In this section, we investigate how downside and upside uncertainties affect cash holdings and leverage policies. In Table 2, we report the results of baseline specifications with total stock return volatility as our measure of uncertainty and do not include downside or upside uncertainty measures. Since we have different control variables for each model, we do not report the estimated coefficients for the individual control variables to conserve space. Results for our cash model, leverage model, and payout model are reported in columns 1, 2, and 3 respectively. The control variables we use in columns 1, 2, and 3 are the same as those in Table 3, Table 4, and Table 5, respectively. We use past 36-month stock return standard deviation to measure total uncertainty. In the cash model, the coefficient on total uncertainty is insignificant, which is inconsistent with the results generally found in the literature. 10 In the leverage and payout models, the coefficients on total uncertainty are negative and significant, which confirms existing results. 10 Gao and Grinstein (2015) find a positive and significant relationship between realized stock return volatility and firms cash holdings. Gao and Grinstein (2015) use weekly stock returns during a certain year to calculate volatility, while we use monthly returns in the past 3 years. Second, Gao and Grinstein (2015) do not employ fixed effects in their regression models, while we include industry and year fixed effects. Third, Gao and Grinstein (2015) restrict their samples to large firms, while we impose weaker size restrictions. Finally, we estimate the model over a different (longer) sample period. In unreported results, we confirm that the coefficient on total return volatility is positive and significant for our sample when we remove industry and year fixed effects and cluster standard errors at the firm-year level. Hence, the coefficient is sensitive to the model specification. In addition, we obtain a positive coefficient (although insignificant) for our specification using a subsample ending in 2005 (to exclude the effect of financial crisis). 15
Our results for models linking firms cash holdings with downside and upside uncertainties and additional controls are reported in Table 3. We follow Bates, Kahle and Stulz (2009) and control for market-to-book ratio, size, cash flow to assets, net working capital to assets, capital expenditure to assets, leverage, dividend payout dummy, R&D to assets, and lagged dependent variable in our cash model. Bates, Kahle and Stulz (2009) also control for industry cash flow volatility. Since our interest is to test the effect of other measures of uncertainties, we do not simultaneously include industry-level risk measures. Our results for cash level with downside and upside uncertainties replacing total uncertainty are reported in Table 3. Columns 1 to 3 use semi-deviations to proxy for downside and upside uncertainties, and columns 4 to 6 use modified semi-deviations to proxy for downside and upside uncertainties. In columns 1 and 4, we estimate panel regression models with industry and year fixed effects with either set of downside and upside uncertainty proxies. The standard errors are clustered at the firm level. In columns 2 and 5, we estimate panel regression models with firm and year fixed effects. In columns 3 and 6, we obtain the Arellano-Bond dynamic panel estimators including a single lag of the dependent variable. We treat leverage, dividend dummy, capital expenditure, upside uncertainty, and downside uncertainty as endogenous in constructing moment conditions associated with the Arellano-Bond estimation. In all the specifications, the coefficients on downside uncertainty are positive and significant, indicating that firms with more downside uncertainty hold more cash, and the coefficients on upside uncertainty are negative and significant, indicating that firms with more upside uncertainty hold less cash. The coefficients are economically significant as well. According to column 1 of Table 3, a 10% shift in downside uncertainty will lead to a 2.63% increase in corporate cash holding; a 10% shift in upside uncertainty will lead to a 1.17% decrease in corporate cash holding. As compared with the results 16
in our baseline regression shown in Table 2, our results indicate that splitting total uncertainty into downside and upside parts is important in predicting firm cash policy. We further test whether the effect of downside uncertainty is stronger than the effect of upside uncertainty. Because the two coefficients have different signs, the null hypothesis we test is whether the sum is zero (H 0 : β UP + β DN = 0). If the sum is positive and significant, the positive effect from downside uncertainty is stronger than the effect from upside uncertainty. Test statistics are reported in the last row of each column, and the results generally confirm our hypothesis. The sum is positive and significant in all columns but column 6. The inefficient dynamic panel estimators may cause the insignificant difference. The results generally indicate that the effect of downside uncertainty is stronger than the effect of upside uncertainty. We run the same analyses for leverage and the results are reported in Table 4. In our Arellano-Bond estimation, we treat cash, dividend dummy, tangibility, upside uncertainty, and downside uncertainty as endogenous variables since our dependent variable is leverage now. The results confirm our hypothesis. The coefficients on downside uncertainty are negative and significant in all specifications. When firms have high chance of cash flow decreases, they tend to use lower leverage to avoid financial distress or accessing external financial market in case of cash shortfalls. The coefficients on upside uncertainty are positive and significant in all specifications, which indicates that firms with more upside potential are willing to use higher leverage and commit more cash out flows. The results are also economically significant. According to column 1 of Table 4, a 10% shift in downside uncertainty will lead to a 1.33% decrease in market leverage; a 10% shift in upside uncertainty will lead to a 0.28% increase in leverage. As compared with the results in our baseline regression shown in Table 2, our results indicate that splitting total uncertainty into downside and upside parts is important in predicting firms leverage policy. Furthermore, the 17
coefficients on the downside uncertainty are statistically significantly bigger than the coefficients on the upside uncertainty, as indicated by the tests in the last row of Table 4. To summarize, these results suggest that downside and upside uncertainties differentially affect cash and leverage policies in a firm. The level of cash is positively affected by downside uncertainty but negatively affected by upside uncertainty. The reverse relations hold for corporate leverage ratios. The effect from downside uncertainty is stronger than the effect from upside uncertainty. As compared with our baseline results using total uncertainty, our results also indicate that split uncertainties are more important in predicting cash and leverage policies. 5. Payout policy In this section, we study how downside and upside uncertainties affect firms payout policies. There are multiple dimensions of payout policy. Firms need to decide whether to pay or not, the amount of payment, and the method of payment. Hence, in this section, we document how downside and upside uncertainties affect the propensity to pay, propensity to make dividend changes, and propensity to choose repurchase over other methods of payments. 5.1 The propensity to pay We start with the propensity to payout and the propensity to pay dividends. The first two columns of Table 5 report the results of estimating a logit model where the dependent variable is one if a firm pays out in a year (either by dividends or repurchases) and zero otherwise. We follow Fama and French (2001), Hoberg and Prabhala (2009), and Chay and Suh (2009) and control for ROA, market-to-book ratio, size, cash, leverage, earned/contributed capital mix, investment, and the lagged dependent variable in all specifications. All specifications include industry and year 18
fixed effects and standard errors are clustered by firm. When we use semi-deviations as our downside and upside uncertainty proxies (column 1), the coefficient on downside uncertainty is negative and significant at the 1% level, and the coefficient on upside uncertainty is positive but marginally significant. When we use modified semi-deviations as our proxies (column 2), the coefficient on downside uncertainty is negative and significant at the 1% level, and the coefficient on upside uncertainty is insignificant. The results indicate that firms with high downside cash-flow uncertainty are less likely to payout, but upside uncertainty only has a marginal positive impact. The results on downside uncertainty coincide with the effect of total uncertainty in our baseline specification, but the results on upside uncertainty indicate that the upside part of total uncertainty does not affect this aspect of payout policy. Columns 3 through 6 of Table 5 present the results of estimating similar logit models in which the dependent variable are indicators for dividend payout and share repurchases, respectively. More specifically, in columns 3 and 4, the dependent variable equals one if a firm is identified as a dividend payer in a year and zero otherwise; in columns 5 and 6, the dependent variable equals one if a firm is identified as a repurchaser in a year and zero otherwise. Both panels show similar results. The coefficients on downside uncertainty are negative and significant, but the coefficients on upside uncertainty are insignificant. The results indicate that firms with higher downside uncertainty are less likely to pay dividends, and upside uncertainty does not affect firms propensity to pay dividends. 5.2 Repurchase only When distributing cash flows to shareholders, firms can choose to pay dividends, repurchase, or use both. In this section, we investigate how downside and upside uncertainties 19
affect firms choices between repurchase only and other methods of payout. We estimate logit regressions where the dependent variable equals one if a firm is identified as repurchase only in a fiscal year. The results are reported in Table 6. The first two columns are based on the subsample of firms that are RO and BO. The coefficients on downside and upside uncertainties are both positive and significant. The results indicate that firms with high downside uncertainty are more likely to only repurchase shares instead of bother repurchasing and paying a dividend because of the flexibility that repurchasing allows. Firms with high upside uncertainty are also more likely to only repurchase because when firms are more likely to have upward cash flow changes they should be more likely to use repurchase to payout these cash flows (Jagannathan, Stephens, and Weisbach, 2000; Guay and Harford 2000). The results in columns3 and 4 are based on the subsample of firms that are RO and DO. The results indicate that firms with higher downside and upside uncertainties are more likely to choose repurchasing shares over paying dividends. Another way of generalizing the results in Table 6 is to ask what they indicate about the choice to pay a dividend. The positive coefficients on both up and downside uncertainty suggest that any uncertainty makes firms less likely to pay a dividend and more likely to repurchase shares if it has cash to distribute. However, the substantially larger coefficient on downside uncertainty indicates that it has a more dampening effect on dividend payment than a similar amount of upside uncertainty. 5.3 Dividend changes In this subsection, we test whether downside and upside uncertainties are related to dividend initiations, omissions, increases, and decreases. Table 7 reports the results. Columns 1 and 2 of Table 7 investigates the decision to initiate dividends. We identify a firm as an initiator 20
in a year if the firm does not pay a dividend in year t 1 and pays dividends in year t, and we do not consider the payment of special dividends in year t as a dividend initiation. We follow DeAngelo, DeAngelo, and Skinner (2000) and classify a cash distribution as a special if it has distribution code 1262 or 1272. Our tests are based on the subsample of firms that do not pay a dividend in year t 1. The results show that dividend initiations are less likely when firms have higher downside uncertainty and they are unaffected by upside uncertainty. Columns 3 and 4 of Table 7 examines dividend omissions. Here, we base our analyses on the subsample of firms that do pay dividends in year t 1. The dependent variable is one if the given firm ceases dividend payments in year t, and we do not consider special dividends either. The results show that dividend omissions are positively affected by downside uncertainty but not significantly affected by upside uncertainty. If a firm faces higher chance of cash flow decreases, it might omit dividend and preserve cash. Columns 5 through 8 of Table 7 examine the decision to increase or decrease dividends. In both panels, we limit the subsample to firms that pay dividends in both year t 1 and year t, and we do not consider special dividends either. In columns 5 and 6, the dependent variable is one if the firm increased its dividend in year t and zero otherwise. In columns 7 and 8, the dependent variable is one if the firm decreased its dividend in year t, and is zero otherwise. Downside uncertainty positively predicts dividend decreases and negatively predicts dividend increases. In contrast to initiation and omission decisions, dividend increases are positively affected by upside uncertainty and dividend decreases are negatively affected by upside uncertainty. One interpretation is that firms are more conservative with dividend initiations and are reluctant to omit dividends, but they to adjust dividend levels based on cash flow variability and do so differentially for up and downside cash flow uncertainty. 21
In summary, upside uncertainty does not affect the decision to payout cash flow, to be a dividend payer, to initiate or omit a dividend. On the other hand, we find that upside uncertainty positively affects the likelihood of dividend increases and negatively affects the likelihood of dividend decreases and find the reverse relationship for downside uncertainty. Furthermore, conditional on paying cash to shareholders, firms are more likely to choose repurchase only when they have both higher upside and downside uncertainty. 6. Conclusion In this paper, we decompose measures of future cash flow variation into downside and upside components. We document that these components differentially affect firms cash, leverage, and payout decisions. In particular, we consistently find that corporate policies are more sensitive to downside uncertainty relative to upside. Theoretical treatments of corporate policies such as leverage and cash holdings emphasize the importance of potential cash shortfalls. Although this motivates our decomposition, it is important to emphasize that the results we obtain are not obvious, because (total) variance can still be a sufficient measure of risk in important benchmark settings (e.g., normally distributed future cash flows). Moreover, we find that upside uncertainty differentially, but nontrivially, impacts key corporate policies including cash holdings and leverage. Our overall pattern of results indicates that relatively nuanced stochastic properties associated with future cash flows (e.g.. skewness and kurtosis) are important drivers of corporate policy decisions. 22
A.1 Definition of control variables A.1.1 Cash A.1.2 Leverage A.1.3 Payout Policy Market-to-book ratio for assets: book value of assets (AT) minus the book value of equity (BE) plus the market value of equity as the numerator of the ratio and the book value of assets as the denominator. Size: logarithm of total book assets (Ln(AT)) to proxy for firms size. Cash flow to assets: cash flow is defined as earnings after interest, dividends, and taxes but before depreciation divided by lagged book assets. Net working capital to assets: current assets (ACT) excluding cash (CHE) minus current liabilities (LCT) divided lagged total assets. Capital expenditure to assets: capital expenditure (CAPX) divided by lagged total assets as our measure of capital expenditure. Leverage: The numerator of our leverage measure is the book value of long-term debt (DLT) plus the book value of debt in current liabilities divided (DLC). The denominator of our leverage measure is market value of equity plus book value of long-term debt plus the book value of debt in current liabilities. Acquisition: the value of acquisitions (AQC) divided by lagged total assets. Dividend payout dummy: a dummy variable equal to one in years in which a firm pays a common dividend. Otherwise, the dummy equals zero. R&D to assets: R&D (XRD) divided by lagged assets, and set to zero when R&D is missing. Operating cash flow: EBIT divided by lagged assets to measure operating cash flow. Cash holding: cash and short term investment (CHE) divided lagged assets to measure cash holding. Size: logarithm of total book assets (Ln(AT)) to proxy for firms size. Research and development: research and development expense (XRD) divided by lagged total assets. Depreciation: depreciation (DP) divided by lagged assets to measure depreciation Tangibility: net plant, property, and equipment (PPENT) divided by lagged total assets to measure tangibility. Market-to-book ratio for assets: the book value of assets (AT) minus the book value of equity (BE) plus the market value of equity as the numerator of the ratio and the book value of assets as the denominator. Dividend payout dummy: a dummy variable equal to one in years in which a firm pays a common dividend. Otherwise, the dummy equals zero. 23
ROA: earnings before extraordinary items (IB) plus long-term and short-term interest expense (XINTD + XINST) plus income statement deferred taxes (TXDI) divided by assets. Market-to-book ratio for assets: the book value of assets (AT) minus the book value of equity (BE) plus the market value of equity as the numerator of the ratio and the book value of assets as the denominator. Size: logarithm of total book assets (Ln(AT)) to proxy for firms size. Cash holding: cash and short term investment (CHE) divided lagged assets to measure cash holding. Leverage: The numerator of our leverage measure is the book value of long-term debt (DLT) plus the book value of debt in current liabilities divided (DLC). The denominator of our leverage measure is market value of equity plus book value of long-term debt plus the book value of debt in current liabilities. Earned/contributed capital mix: retained earnings (RE) divided by book equity as our measure. Investment: growth in fixed assets as our investment measure. We use total assets minus current assets as fixed assets, and then we calculate the growth rate. Namely, our investment measure is fixed assets t fixed assets t 1 fixed assets t 1. 24