A. TimeSeries Predictability Analysis
By far the most important result in the literature on estimating timevarying expected returns is the predictive power of dividend yields. For example, Campbell, Lo, and MacKinlay (1997) as well as Cochrane (2001) give center stage to empirical results involving the dividend price ratio. This evidence has been examined across asset classes, across industries, and across countries. While there is significant debate about the role of dividend yields as predictors, especially at long horizons, part of this extra scrutiny is due to the dividend yield being the most popular predictive variable.^{15}Goyal and Welch (2003), for example, provide a detailed and thorough analysis of various measures of dividend yields and argue that its predictive power has been overstated both in and outofsample. In particular, they document predictability prior to 1990 but show that this disappears when including the last decade. After considering various explanations, they argue that the most likely one is that the relation was spurious.
As the literature argues, the last 15 years have exhibited a dramatic shift in the breakdown between payout yields and dividend yields (see, e.g., Cochrane (2001, p. 391) and Allen and Michaely (2003)). Panel A of Table II presents the results of aggregate timeseries regressions of the market excess return on the dividend, payout, and net payout yield. From 1926 to 1984 (i.e., the “early sample”) the coefficient on the dividend yield is 0.296 with a tstatistic of 3.666 and R^{2} of 13%. However, when recent history is included, the coefficient and explained variation drop dramatically to 0.116 with a corresponding tstatistic of 2.24 and R^{2} of 5.5%. The temporary disappearance of statistical significance in the late1990s caused researchers to conclude that predictability based on dividend yields disappears. A series of high returns coupled with low dividend yields early in the new millennium brought statistical significance back (see our tstatistic of 2.240), but the breakdown between the explanatory power and the coefficient is still a resounding puzzle vis a vis early predictability stories. Moreover, this significance, when appropriately adjusted for the wellknown estimation bias inherent in this setting (see Stambaugh (1999)), disappears in the full sample but not in the early subsample.
Table II. Return Predictability The sample consists of all nonfinancial firms in the intersection of CRSP and Compustat with data for dividends paid to common shareholders, repurchases of common stock, and sales of common stock. The excess market return is the difference in the CRSP valueweighted total return (including dividends) and the return on a 3month Treasury bill. The dividend yield is computed as the difference in the cum and exdividend returns to the CRSP valueweighted index. The payout yield is the sum of dividend yield and repurchase yield, defined as the ratio of common share repurchases to yearend market capitalization. We measure repurchases in two ways, using the statement of cash flows (data from 1971) and the change in the Treasury stock (data from 1983). Because repurchases were negligible prior to and just after the passing of SEC Rule 10b18 in 1982, we assume that repurchases are zero before the availability of each measure. The net payout yield is the sum of the dividend yield and the repurchase yield (using the statement of cash flows measure) less the issuance yield, defined as the ratio of common share issuances to yearend market capitalization. Since repurchase and issuance data from the statement of cash flows goes back only to 1971, we use the monthly change in shares outstanding from the CRSP monthly stock file to capture net equity issuances prior to 1971. In particular, net equity issuance for month t is the product of the splitadjusted growth in shares and the average of the splitadjusted stock price at the beginning and end of the month. All variables are in logs (0.1 is added to net payout yield to avoid negative yields). Panel A presents results from univariate regressions of the log excess market return on the dividend yield, payout yield(s), and net payout yield. Panel B presents the results from multivariate regressions of the log excess market returns on the dividend yield and either the payout or net payout yield. All standard errors (SE) are heteroskedasticity consistent. The tstatistics are the ratio of the coefficient to the standard error. The adjusted coefficient (adj. coefficient) is computed using the method of Amihud and Hurvich (2004). Simulated pvalues (Sim pvalue) are computed via 10,000 simulations under the null of zero predictability, but accounting for the regressor's autocorrelation and the crosscorrelation of the errors. The R^{2} Sim pvalue is the corresponding R^{2} from simulations under the null. The rho is the crosscorrelation between the errors of the AR(1) and the errors of the predictive regression. Adjusted beta and confidence interval are calculated following Amihud and Hurvich (2004). The Bonnferoni Qtest confidence interval is calculated following Campbell and Yogo (2005). The GoyalWelch (2003, 2005) root mean squared error differential (dRMSE) uses 60 periods as the lookback window and an outofsample period of 1985 to 2003. The Sim pvalue of the dRMSE measure is calculated within the simulations under the null described above. SE is the standard error. Panel A: Univariate Predictive Regressions 

 Log(Dividend Yield)  Log(Payout [CF] Yield)  Log(Payout [TS] Yield)  Log(0.1 + Net Payout Yield 

Full Sample: 1926–2003 
Coefficient  0.116  0.209  0.172  0.759 
SE  0.052  0.062  0.060  0.143 
tstatistic  2.240  3.396  2.854  5.311 
pvalue  0.014  0.001  0.003  0.000 
Sim pvalue  0.170  0.045  0.080  0.000 
R^{2}  0.055  0.091  0.080  0.262 
R^{2} Sim pvalue  0.083  0.011  0.020  0.000 
ρ  −0.709  −0.671  −0.691  −0.301 
Adj. Coefficient  0.072  0.167  0.126  0.736 
SE  0.056  0.076  0.067  0.146 
tstatistic  1.281  2.192  1.872  5.058 
pvalue  0.102  0.016  0.032  0.000 
Bonferroni QLow  −0.007  0.035  0.014  0.313 
Bonferroni QHi  0.151  0.267  0.209  0.641 
dRMSE (GW)  −0.068  0.024  −0.017  0.048 
Sim pvalue dRMSE  0.932  0.082  0.703  0.022 

Early Sample: 1926–1984 
Coefficient  0.296  0.280  0.300  0.794 
SE  0.081  0.076  0.080  0.149 
tstatistic  3.666  3.688  3.741  5.342 
Bonferonni QLow  0.070  0.065  0.077  0.347 
Bonferonni QHi  0.389  0.387  0.390  0.735 
pvalue  0.000  0.000  0.000  0.000 
Sim pvalue  0.044  0.054  0.043  0.001 
R^{2}  0.130  0.121  0.135  0.300 
Panel B: Multivariate Predictive Regressions 

 Log(Dividend Yield)  Log(Payout [CF] Yield)  Log(Payout [TS] Yield)  Log(0.1 + Net Payout Yield  R^{2} 

Coefficient  −0.088  0.318   0.098 
SE  0.111  0.129  
Coefficient  −0.394   0.641   0.112 
SE  0.216   0.251  
Coefficient  −0.042   0.830  0.267 
SE  0.064   0.108  
In contrast to the results for dividend yields, when we use the payout yield as a predictor for the entire sample period, the regression coefficients, tstatistics, and R^{2}s change only slightly and statistical significance is not lost. This is consistent with our explanation that measurement error and omitted variables drive the decline in dividend yield predictability. Interestingly, this is true irrespective of whether we use cash flow–based repurchases or Treasury stock–based repurchases to correct the dividend yield. For the cash flow (and Treasury stock)–based payoff yield measures the regression coefficient drops from 0.280 (0.300) to 0.209 (0.172), the tstatistic remains highly significant at any reasonable level, dropping from 3.688 (3.396) to 3.741 (2.854), and the R^{2} drops from 12.1% (13.5%) to 9.1% (8.0%).
Before we turn to the results for the net payout yield, it is useful to note that the regressor we use is constructed somewhat differently. In particular, the net payout yield is not necessarily positive anymore due to the netting out of equity issuances (see Section II.A above), hence, we can no longer use log(yield) as our regressor. Since we want to deviate from the literature as little as possible in this respect we simply bound the net yield away from negativity by defining the regressor as log(net yield + 0.1). The results are qualitatively robust to the precise size of the adjustment factor (0.1 in our case) but larger adjustments further remove the comparability of our estimates with those of the existing literature.^{16}
The results for net payout yield are quite striking. While the regression coefficient we obtain is not comparable to the ones computed above for the total payout and the dividend yield, the tstatistic and R^{2} are comparable. The tstatistic is 5.311, significant at any standard level. The R^{2} is 26.2%. This result is striking in light of the baseline level of predictability using dividends alone of 5.5%, or even in light of the payout measures' explanatory power of 8.0% or 9.1%.^{17}
There are two potential statistical objections to the evidence of predictability at short horizons, namely, the small sample bias of the predictive estimator (e.g., Stambaugh (1986)) and the breakdown of typical asymptotics in small samples due to the presence of highly persistent regressors (e.g., Elliott and Stock (1994)). Because these statistical issues arise as a result of the properties of the predictive variable, it is possible that the different results for the dividend versus payout measures may be due to these issues rather than fundamentals. It is important therefore to document the differential evidence with the appropriate corrections.
With respect to small sample bias, Stambaugh (1986, 1999) notes that the typically high persistence in regressors used in predictive regressions, coupled with the strong negative correlation between innovations to these regressors and asset returns themselves, create a bias in the predictive regression coefficient. When the bias is adjusted appropriately (e.g., downward), the regression coefficients are typically found to be insignificant. More recently, Lewellen (2004) points out that the autocorrelation of the regressor, when appropriately bounded below unity, would affect the Stambaugh bias by reducing the standard errors relative to those calculated while ignoring the constraint on the AR(1) coefficient of the regressor. In some cases (e.g., some periods and/or some regressors) this may have the effect of salvaging predictability.
Amihud and Hurvich (2004) suggest a simple method to implement the StambaughLewellen adjustment via ordinary least squares (OLS) regressions.^{18} Using their approach, we compute biasadjusted beta coefficients and standard errors, which appear in Panel A of Table II. Since at an annual frequency the persistence in the regressor is lower and the correlation between returns and the innovations to the regressor is lower, the effect of the adjustment is smaller. Interestingly, due to the relatively high persistence in the dividend yield series (namely, 0.94) and the relatively high crosscorrelation between the regressor's AR(1) errors and returns (namely, −0.709) relative to the other regressors discussed next, the bias adjustment has sufficient bite to diminish the statistical significance of the dividend yield predictability obtained in standard OLS regression. In particular, while the OLS beta is 0.116, the biasadjusted beta is 0.072; further, while, as noted above, the OLS tstatistic is 2.240, it is 1.281 for the biasadjusted beta. This lost significance is not the case, however, for all other regressors under consideration. In particular, the adjusted beta of the cash flow–based (Treasury stock–based) payout yield regressor is 0.167 (0.126) with a tstatistic of 2.192 (1.872). The statistical significance is not lost here because the adjustment necessary for the payout series is small. This is a result of the relatively low persistence of these regressors, coupled with the relatively weak correlation between the innovations to the regressors and returns.
With respect to possible size distortions of tstatistics due to nearunit root properties of the regressor, Elliott and Stock (1994) derive an alternative asymptotic theory in which they explicitly model the regressor as having a localtounit root. A number of recent finance papers apply this theory to the question of stock return predictability (see, e.g., Torous, Valkanov, and Yan (2001), Jansson and Moreira (2003), Campbell and Yogo (2005), and Polk, Thompson, and Vuolteenaho (2003)). Under this alternative methodology, the researcher can construct Bonferronibased tests that are robust to the persistence problem by directly incorporating DickeyFuller (1979) confidence intervals around the autoregressive parameter. In Panel A of Table II, we present the 5% confidence interval for the onesided Bonferroni Qtest using the test methodology and critical value in Campbell and Yogo (2005).^{19} For this relatively more conservative test we find that beta is bound away from zero for all three total payout variables, but not for the dividend yield. The Bonferroni Qtest cannot reject that beta is zero for the dividend price ratio, with a lower confidence interval below zero, namely, −0.007. The gross payout measures are bound away from zero, with the lower point of the confidence interval being 0.035 for the cash flow–based variable and 0.014 for the Treasury stock one. The 5% lower tail of the total net payout variable is bound well away from zero at 0.313, as one would expect given its low persistence and high R^{2}.
In conclusion, neither of the above statistical issues can explain the different predictability results using dividend yields versus the various payout measures. As an alternative comparison of the measures, we investigate their true predictive content from an economic perspective by turning to benchmarks recently set out in a series of papers by Goyal and Welch (2003, 2005). Goyal and Welch suggest an economically motivated, intuitive benchmark for predictability, namely, outofsample performance. They compute the root meansquarederror differential (dRMSE) between two competing models (i) a myopic model, where expected returns are just the historical mean risk premium, and (ii) a predictabilitybased MSE, where expected returns are based on a rolling regression of available past data at any point in time. Goyal and Welch argue that a reasonable economic benchmark for predictability to be interesting from an economic standpoint is dRMSE >0.
In Table II we compute this measure with a rolling lookback window of 60 years and a forecast period starting in 1985. Consistent with Goyal and Welch, the dividend yield series does not provide sufficient predictive information to overcome statistical or modeling errors; hence, not only is its statistical validity questionable, but its economic relevance is doubtful. Interestingly, two out of the three alternative series that we examine do manage to beat the GoyalWelch benchmark. The dRMSE is positive for all but the Treasury stock–based total payout variable. The cash flow–based payout forecasts exhibit a dRMSE of 2.4%, while that of the net payout is a remarkable 4.8% on a per annum basis.
The statistical significance of the dRMSE measure is examined in Goyal and Welch (2003) using asymptotic statistical theory developed by Diebold and Mariano. In some recent work, for example, Clark and West (2005), authors note that statistical noise in the repeated estimation of predictive coefficients in the rolling regression framework can introduce a bias into the dRMSE measure. This is because in finite samples the RMSE under the null of no predictability is not expected to be zero, but instead, negative. Motivated by smallsample concerns, Goyal and Welch (2005) provide a bootstrapping analysis of the DieboldMariano dRMSE statistic and provide corresponding cutoff values. Similarly, we conduct a Monte Carlo simulation under the null of no predictability and obtain a simulationbased pvalue for the dRMSEs given the relevant parameters.^{20} Not only are the dRMSEs of the cash flow–based total payout series and those of the net payout series positive, but they also have impressive pvalues of 8.2% and 2.2%, respectively. That is, in simulations under the null of no predictability, only 8.2% (2.2%) of the time is the dRMSE greater than 2.4% (4.8%).
Panel B of Table II provides a “horse race” between the dividend yield and the various payout yield measures. Consistent with our main thesis, all three payout measures are highly significant, while the dividend yield is insignificant in each case. For example, for the dividend yield and cash flow–based total payout bivariate regression the pvalue (not shown) on dividends is 0.790 while that of total payout is 0.007. Comparing the R^{2}s across univariate and bivariate regressions, we conclude that dividend yield's contribution to the regression is negligible in the presence of any of the other payout regressors. For example, the univariate R^{2} for the cash flow–based payout yield regressor is 9.2% while in the bivariate case it is 9.8%—not a remarkable difference. This result, that dividends disappear when pitted against payout series, carries through to the other two regressions.
B. CrossSectional Analysis
The idea that dividends can be a useful measure for expected stock returns has its roots in early finance research (e.g., Dow (1920)). More recent research on the crosssectional relation between dividend yields and returns is motivated not only by Dow's findings or the implication of the Gordon growth model, but by the presence of market imperfections. For example, Litzenberger and Ramaswamy (1979), among others, use tax motives to find a positive relation between expected returns and dividend yields in the context of a taxbased CAPM. Others studies (e.g., John and Williams (1985), Allen et al. (2000), Grullon et al. (2002)) turn to agency problems and information asymmetries in motivating a crosssectional relation between equity returns and dividend yields. In this section, we explore (i) whether yields are useful measures for describing crosssectional variation in expected returns, and (ii) whether the different yield measures (i.e., dividend vs. (net) payout) lead to different conclusions. We close the section by comparing the performance of simple trading strategies based on our three yield measures.
Our first set of analyses examines the characteristics of stocks as a function of our different yield measures. Each year at the end of June, we form 10 portfolios based on the ranked values of the dividend yield, payout yield, and net payout yield from December of the previous year. Breakpoints for the decile portfolios are determined using only NYSE stocks with a nonzero yield. Stocks with zero yields comprise their own portfolio.
Table III presents the average monthly return, postranking beta, log firm size, log booktomarket, yield, and number of firms for each of 10 positive yield portfolios, as well as for a portfolio of zeroyield stocks, for the July 1984 to December 2003 period. With the exception of the first decile portfolio, there seems to be little crosssectional return variation based on these portfolios, for example, the lowest three deciles' mean monthly return is 1.15% monthly, the middle four deciles' monthly return is 1.28%, and the highest three is 1.33%. This contrasts with the July 1963 to June 1984 period (not reported in the tables), in which these same portfolios increase sharply from the low deciles (1.23%) to the high deciles (1.63%). In both periods, the average beta decreases with the dividend yield while the average booktomarket ratio increases with dividend yields.
Table III. Yield Portfolio Returns and Characteristics At the end of June of each year t, 10 portfolios are formed on the basis of ranked values of the dividend yield, payout yield, and net payout yield. The dividend (payout) [net payout] yield is the ratio of common dividends (dividends plus common share repurchases) [dividends plus repurchases minus common share issuances] in year t to yearend market capitalization. There are two measures of payout yield, one based on the statement of cash flows, the other based on the change in Treasury stock. For the net payout yield, we use the cash flow–based measure of repurchases. To mitigate the effect of outliers, we trim the upper and lower 0.5% of the log(booktomarket) distribution, the upper 5% of the dividend and payout yield distributions, and the upper and lower 2.5% of the net payout yield distribution. All stocks containing nonmissing data for the ratio of booktomarket equity, common share dividends, common share repurchases, and common equity sales, and at least 2 years worth of historical return data are then allocated to the yield portfolios using NYSE breakpoints based on positive yields (nonzero yields in the case of the net payout). Each portfolio's monthly equalweighted return for July of year t to June of year t+ 1 is calculated, after which the portfolios are reformed. ME is the market capitalization, in millions of dollars, as of June in year t. BE/ME is the ratio of book equity in December of year t− 1 to market capitalization in December of year t− 1. β is the postranking beta for one of 100 size × booktomarket portfolios and is computed as the sum of the coefficients from a timeseries regression of portfolio returns on contemporaneous and lagged excess market return. Firms is the average number of stocks in the portfolio in each month. A portfolio consisting of zeroyield stocks is presented as well. Panel A: Portfolios Formed on the Dividend Yield during July 1984 to December 2003 

Dividend Yield Deciles 

 Zero Yield  1  2  3  4  5  6  7  8  9  10  All 

Average return  1.43  1.04  1.27  1.13  1.26  1.26  1.29  1.32  1.36  1.37  1.25  1.26 
β  1.48  1.29  1.25  1.21  1.20  1.18  1.17  1.14  1.13  1.03  1.08  1.17 
ln(ME)  3.75  6.17  5.92  6.04  6.08  6.10  6.06  6.04  5.77  5.82  5.53  5.95 
ln(BE/ME)  −0.56  −0.92  −0.70  −0.61  −0.53  −0.46  −0.38  −0.29  −0.18  −0.05  −0.56  −0.36 
Dividend yield  0.00  0.00  0.01  0.01  0.02  0.02  0.03  0.03  0.04  0.06  0.18  0.04 
Firms  2,357  153  148  132  128  122  122  121  132  123  118  130 

Panel B: Portfolios Formed on the Cash Flow–Based Payout Yield during July 1984 to December 2003 
Cash Flow–Based Payout Yield Deciles 
 Zero Yield  1  2  3  4  5  6  7  8  9  10  All 

Average return  1.37  1.22  1.33  1.28  1.33  1.38  1.41  1.46  1.52  1.41  1.74  1.41 
β  1.49  1.42  1.33  1.29  1.26  1.23  1.21  1.19  1.13  1.11  1.21  1.24 
ln(ME)  3.61  4.79  5.21  5.39  5.54  5.67  5.73  5.60  5.70  5.64  5.15  5.44 
ln(BE/ME)  −0.57  −0.73  −0.60  −0.49  −0.43  −0.40  −0.37  −0.31  −0.25  −0.18  0.04  −0.37 
Total payout yield  0.00  0.00  0.01  0.02  0.02  0.03  0.04  0.05  0.06  0.08  0.15  0.05 
Firms  1,611  292  213  180  166  158  153  157  148  148  167  178 

Panel C: Portfolios Formed on the Treasury Stock–Based Payout Yield during July 1984 to December 2003 
Treasury Stock–Based Payout Yield Deciles 
 Zero Yield  1  2  3  4  5  6  7  8  9  10  All 

Average return  1.42  1.16  1.35  1.29  1.35  1.30  1.34  1.36  1.46  1.47  1.60  1.37 
β  1.49  1.40  1.31  1.27  1.25  1.21  1.19  1.17  1.12  1.08  1.17  1.22 
ln(ME)  3.71  5.03  5.45  5.57  5.65  5.81  5.85  5.76  5.75  5.70  5.26  5.58 
ln(BE/ME)  −0.59  −0.76  −0.63  −0.54  −0.43  −0.41  −0.36  −0.31  −0.23  −0.12  0.18  −0.36 
Total payout yield  0.00  0.00  0.01  0.02  0.02  0.03  0.03  0.04  0.05  0.07  0.15  0.04 
Firms  1,986  251  191  168  158  149  144  153  145  142  151  165 

Panel D: Portfolios Formed on the Net Payout Yield during July 1984 to December 2003 
Net Payout Yield Deciles 
 Zero Yield  1  2  3  4  5  6  7  8  9  10  All 

Average return  1.77  0.95  1.41  1.36  1.42  1.28  1.36  1.37  1.48  1.56  1.68  1.38 
β  1.45  1.49  1.48  1.42  1.32  1.26  1.23  1.20  1.16  1.13  1.22  1.29 
ln(ME)  2.81  4.22  4.42  4.54  5.16  5.45  5.61  5.70  5.65  5.62  5.12  5.15 
ln(BE/ME)  −0.09  −0.82  −0.78  −0.68  −0.57  −0.46  −0.41  −0.35  −0.28  −0.20  0.03  −0.45 
Total net payout yield  0.00  −0.10  −0.01  0.00  0.01  0.02  0.03  0.04  0.05  0.06  0.14  0.02 
Firms  505  551  481  383  246  198  185  173  173  174  195  276 
As far as the average return is concerned, the portfolios formed on (net) payout yield measures (Panels B through D) tell a different story. Most important, there is measurable crosssectional variation in expected returns, the result being an almost monotonic relation between returns and the payout yield. For the cash flow (Treasury stock)–based measure of payout yield, the lowest three deciles' mean is 1.28% (1.27%), the middle four is 1.40% (1.34%), and the highest three is 1.56% (1.51%). Note that finding higher payout yield portfolios that have higher realized returns is consistent with the timeseries results documented in Section III.A, where we document higher returns during periods of high payout yields for the aggregate market. Like the dividend yield portfolios, the payout yield portfolios are negatively correlated with beta and positively correlated with booktomarket: High payout yield portfolios have lower betas and higher booktomarket ratios than low payout yield portfolios. Similarly, these inferences carry over to net payout yields, whose portfolio (low, medium, high) returns are 1.24%, 1.36%, 1.57%, and are negatively correlated with beta and positively correlated with the booktomarket ratio.
As is now standard in the literature, Table IV performs FamaMacBeth (1973) monthly return regressions on postranking betas, booktomarket, size, and either the dividend, payout, or net payout yield, over the July 1984 to December 2003 period. Again, we focus on this later period corresponding to the period in which share repurchase activity is largely protected from legal action. Specifically, we run crosssectional regressions for each month in order to generate a time series of parameter estimates. As mentioned above, for each year we trim the smallest and largest 0.5% of the observations for booktomarket, the largest 5% of the observations for the dividend yield and payout yield, and the largest and smallest 2.5% of the observations for the net payout yield. This trimming procedure avoids giving extreme observations excessive weight in the regressions, although we also address this issue further by using a robust regression technique discussed below. Table IV presents the average value of each estimated parameter's time series, along with a corresponding standard deviation and tstatistic. Panel A presents the results for the entire sample and Panel B presents the results when only the nonzero yield firms are included.
Table IV. Fama–MacBeth Monthly Return Regressions The sample consists of all nonfinancial firms in the intersection of CRSP and Compustat with data for dividends paid to common shareholders, repurchases of common stock, and sales of common stock. The dividend (payout) [net payout] yield is the ratio of common dividends (dividends plus common share repurchases) [dividends plus repurchases minus common share issuances] in year t to yearend market capitalization. There are two measures of payout yield, one based on the statement of cash flows, the other based on the change in Treasury stock. For the net payout yield, we use the cash flow–based measure of repurchases. To mitigate the effect of outliers, we trim the upper and lower 0.5% of the log (booktomarket) distribution, the upper 5% of the dividend and payout yield distributions, and the upper and lower 2.5% of the net payout yield distribution. We also require that firms have at least 2 years worth of historical return data available on CRSP. Crosssectional regressions are estimated each month. Mean is the timeseries mean of the estimated coefficients, Std. is its timeseries standard deviation, and t(Mn) is Mean divided by its timeseries standard error. Market capitalization is denoted by ME, book equity is denoted by BE, common dividends is denoted by D, cash flow–based (Treasury stock–based) common share repurchases is denoted by RCF (RTS), and stock issuances is denoted by S. β is the postranking beta for one of 100 size × booktomarket portfolios and is computed as the sum of the coefficients from a timeseries regression of portfolio returns on contemporaneous and lagged excess market return. The table provides estimates based on ordinary least squares (OLS) and least absolute deviation (LAD) regressions. Panel A presents results for all nonmissing yield values, including zero yields. Zero dividend and repurchase yields are adjusted by adding 0.01 before converting to percentages and taking logs. The net payout yield is measured in percentages but not converted into logarithmic scale because of negative values. Panel B presents results for all positive yield values for dividend and payout, and all nonzero yields for net payout. Panel A: Entire Sample 

Coefficient  OLS Estimates  LAD Estimates 

July 1984 to December 2003 (234 Months)  July 1984 to December 2003 (234 Months) 

Mean  Std.  t(Mn)  Mean  Std.  t(Mn) 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln (D_{it}/ME_{it}) +e_{it} 
Intercept  2.33  7.27  4.90  −0.81  6.47  −1.91 
β  −0.03  6.06  −0.07  −0.46  5.52  −1.26 
ln(ME)  −0.16  1.08  −2.34  0.30  0.81  5.72 
ln(BE/ME)  0.26  1.28  3.10  0.39  0.90  6.57 
ln(D/ME)  0.03  1.17  0.38  0.23  0.98  3.52 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RCF)_{it}/ME_{it}) +e_{it} 
Intercept  2.11  7.07  4.56  −1.01  6.23  −2.49 
β  0.05  5.92  0.12  −0.35  5.43  −0.98 
ln(ME)  −0.16  1.04  −2.37  0.29  0.78  5.77 
ln(BE/ME)  0.26  1.22  3.25  0.37  0.86  6.58 
ln((D+RCF)/ME)  0.15  0.99  2.24  0.32  0.85  5.73 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RTS)_{it}/ME_{it}) +e_{it} 
Intercept  2.24  7.18  4.77  −0.91  6.37  −2.20 
β  0.02  5.99  0.06  −0.41  5.50  −1.14 
ln(ME)  −0.17  1.07  −2.43  0.30  0.80  5.74 
ln(BE/ME)  0.24  1.28  2.92  0.38  0.92  6.36 
ln((D+RTS)/ME)  0.11  1.11  1.50  0.27  0.94  4.43 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RCF−S)_{it}/ME_{it}) +e_{it} 
Intercept  2.11  7.10  4.55  −0.73  6.45  −1.74 
β  0.10  6.12  0.26  −0.40  5.59  −1.10 
ln(ME)  −0.15  1.03  −2.22  0.31  0.76  6.17 
ln(BE/ME)  0.27  1.25  3.29  0.39  0.92  6.52 
(D+RCF − S)/ME  0.03  0.10  4.14  0.04  0.08  8.09 
Panel B: Positive Yield Subsample 

Coefficient  OLS Estimates  LAD Estimates 

July 1984 to December 2003 (234 Months)  July 1984 to December 2003 (234 Months) 

Mean  Std.  t(Mn)  Mean  Std.  t(Mn) 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln (D_{it}/ME_{it}) +e_{it} 
Intercept  1.47  6.60  3.41  0.25  6.41  0.59 
β  −0.08  5.29  −0.23  −0.30  5.02  −0.92 
ln(ME)  −0.02  0.72  −0.33  0.13  0.69  2.88 
ln(BE/ME)  0.08  0.91  1.31  −0.00  0.86  −0.01 
ln(D/ME)  0.02  0.63  0.43  0.07  0.63  1.58 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RCF)_{it}/ME_{it}) +e_{it} 
Intercept  1.73  6.39  4.15  −0.26  6.26  −0.63 
β  0.08  5.81  0.21  −0.26  5.35  −0.76 
ln(ME)  −0.08  0.85  −1.40  0.20  0.71  4.22 
ln(BE/ME)  0.18  1.10  2.49  0.14  0.83  2.63 
ln((D+RCF)/ME)  0.06  0.48  1.94  0.12  0.46  3.99 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RTS)_{it}/ME_{it}) +e_{it} 
Intercept  1.78  6.23  4.37  −0.11  6.07  −0.29 
β  0.04  5.51  0.11  −0.29  5.09  −0.88 
ln(ME)  −0.08  0.84  −1.38  0.18  0.71  3.89 
ln(BE/ME)  0.15  1.03  2.18  0.10  0.82  1.86 
ln((D+RTS)/ME)  0.05  0.51  1.45  0.11  0.45  3.56 

R_{it}=a+b_{1t}β_{it}+b_{2t} ln (ME_{it}) +b_{3t} ln (BE_{it}/ME_{it}) +b_{4t} ln ((D+RCF−S)_{it}/ME_{it}) +e_{it} 
Intercept  1.77  7.20  3.76  −0.77  6.68  −1.76 
β  0.21  6.27  0.50  −0.27  5.71  −0.73 
ln(ME)  −0.11  1.00  −1.62  0.29  0.76  5.85 
ln(BE/ME)  0.28  1.32  3.21  0.39  0.99  6.02 
(D+RCF−S)/ME  0.03  0.10  4.03  0.04  0.08  8.64 
For the OLS regressions on the entire sample, the standard results appear in the significantly negative coefficient on size and significantly positive coefficient on booktomarket. The market beta coefficient is insignificant across all four specifications. More to the point of this paper, however, are the differences across our yield measures. Both the (cash flow–based) payout yield and net payout yield coefficients are positive and highly significant, whereas the dividend yield coefficient is insignificant. The coefficient on the log payout yield is 0.15 with a tstatistic of 2.24, and the coefficient on net payout yield is 0.03 with a tstatistic of 4.14. The difference in coefficient magnitude is due to the fact that we cannot use the log transformation on the net payout yield because of negative values, which prevents a direct comparison of the coefficients.^{21} The Treasury stock–based payout yield is also positive (0.08) and more than twice the magnitude of the dividend yield, but statistically insignificant. However, the important point is that relative to the dividend yield (coefficient of 0.03 with a tstatistic of 0.38), the payout yield and net payout yield show significantly stronger associations with stock returns.
One of the commonly cited problems in measuring the cross section of returns is the extent to which the results are robust. In particular, Knez and Ready (1997) argue that robust estimation should be applied due to outliers and find that, indeed, the size effect reverses (becomes positive) when such a technique is applied. As such, we apply a similar method here by reestimating the model using a least absolute deviation (LAD) regression. Similar to Knez and Ready (1997), the standard size effect reverses sign. The booktomarket effect remains positive and significant. Interestingly, the coefficient on the log dividend yield is now significantly positive, though relatively less so than the (net) payout yield.
Finally, in Panel B we report OLS and LAD regression results for the case in which the sample is restricted to only positiveyield stocks. Interestingly, in the dividend yield specification of the OLS regression, the standard result no longer applies. That is, size and booktomarket are not significantly related to returns. In the payout and net payout yield specifications, the size coefficient is still insignificant but the booktomarket coefficient is now significant. This change is due in large part to a larger sample of firms that pay dividends or repurchase shares compared to firms that only pay dividends. More to the point of this paper though, where the dividend yield is insignificant, the payout and net payout yields are significantly related to returns, but for the OLS estimate of the Treasury stock–based payout yield. The LAD estimates for the positive yield subsample show a similar pattern: The payout coefficients are positive and highly significant while the dividend yield appears relatively less important. This is consistent with the declining rank correlation between payout yields and dividend yields during the 1984 to 2003 period described above and in Figure 2.
Thus far, our results suggest that payout and net payout yields have explanatory power for crosssectional variation of returns over and above the standard firm characteristics, and that the payouts' coefficients are robust to sample specification, as well as outliers. On the other hand, for most cases during the 1984 to 2003 period, the dividend yield coefficient is not able to explain crosssectional variation in returns and is more sensitive to sample specifications.^{22}
Given the evidence of crosssectional covariation between stock returns and the payout yield, we develop measures of dividend, payout, and net payout yields as potential factors. We begin by sorting firms into three dividend yield groups and three payout yield groups each year, based on their deciles discussed earlier. These low, medium, and high groups correspond to the bottom three, middle four, and top three deciles. We then construct nine portfolios from the intersection of the dividend and payout yield groups and compute valueweighted average returns for each portfolio. Our dividend yield factor is computed as the average return across the three high dividend yield groups minus the average return across the three low dividend groups. The (net) payout yield factor is constructed in a similar manner. This approach mirrors Fama and French's (1993) method for forming size and booktomarket factors and, as such, aids in purging the correlation between our yield factors. The result of this procedure is four monthly timeseries: DYHML (corresponding to the dividend yield factor), PYCFHML (corresponding to the cash flow–based payout yield factor), PYTSHML (corresponding to the Treasury stock–based payout yield factor), and NPYHML (corresponding to the net payout yield factor).
The analysis that we perform is standard and based on the original portfolio regressions performed by Fama and French (1993). We begin by merging monthly data for the riskfree return, excess market return, SMB factor return, and HML factor return (all of which are obtained from Ken French's website) with our yield factors discussed above. These three time series, in addition to one of our yield factors, form the design matrix in our factor regressions. The dependent variables consist of monthly excess stock returns for three sets of 25 portfolios: beta/payout yield, size/payout yield, and booktomarket/payout yield.^{23} For comparison with the existing evidence, these portfolios include zeroyield stocks, though the construction of our factor returns do not. For the beta/payout yield portfolios, we sort NYSE stocks in June of each year t into beta and (independently) payout yield quintiles.^{24} We then construct 25 portfolios from the intersection of the quintiles and compute a valueweighted monthly return. Next we regress monthly excess portfolio returns on an intercept, the excess market return, SMB, HML, and either DYHML, PYCFHML, PYTSHML, or NPYHML. Panel A of Table V presents the estimated intercepts and yield coefficients, as well as the corresponding tstatistics, for the booktomarket/yield portfolios.^{25}
Table V. Factor Regressions The sample consists of all nonfinancial firms in the intersection of CRSP and Compustat with data for dividends paid to common shareholders, repurchases of common stock, and sales of common stock. The dividend (payout) [net payout] yield is the ratio of common dividends (dividends plus common share repurchases) [dividends plus repurchases minus common share issuances] in year t to yearend market capitalization. To mitigate the effect of outliers, we trim the upper 5% of the dividend and payout yield distributions, and the upper and lower 2.5% of the net payout yield distribution. We also require that firms have at least 2 years worth of historical return data available on CRSP. The regression equation is The regressand is monthly excess portfolio returns, R_{t}−Rf_{t}, from July 1984 to December 2003, and the regressors are the market excess return (RM_{t}−Rf_{t}), the small minus big factor return (SMB_{t}), the high minus low factor return (HML_{t}), and the high minus low yield factor return for the dividend yields, payout yields, and net payout yields. The first three regressors are obtained from Ken French's website. YIELDHML corresponds to one of the four yield factors: dividend (DYHML), cash flow–based payout (PYCFHML), Treasury stock–based payout (PYTSHML), or net payout (NPYHML). The table presents intercept and yield factor slope coefficient estimates (and corresponding tstatistics) for 25 portfolios formed on booktomarket and payout yield (panel A). Panel B presents a summary of statistical tests of intercept significance for five different model specifications and the three aforementioned sets of portfolios. χ^{2} is the test statistic corresponding to a Wald test of the joint hypothesis that all of the intercepts are equal to zero. One asterisk (two asterisks) correspond to statistical significance at the 5% (1%) level. Panel A: BooktoMarket Equity/Payout Yield Portfolios 

Payout Yield Quintiles 

 Low Yield  2  3  4  High Yield  Low Yield  2  3  1  High Yield 



B/M Quintiles  YIELDHML= Dividend Yield 

 α  t(α) 
Small  0.04  −0.02  0.30  0.55  0.53  0.33  −0.16  2.11  3.04  2.45 
2  −0.15  −0.12  0.09  0.05  0.16  −1.12  −0.75  0.51  0.26  0.77 
3  −0.32  −0.33  −0.15  0.11  0.16  −1.82  −1.78  −0.79  0.61  0.82 
4  −0.36  −0.09  0.03  −0.28  0.02  −2.07  −0.43  0.16  −1.81  0.13 
Big  −0.37  −0.41  −0.36  −0.26  0.11  −1.94  −1.69  −1.37  −1.42  0.69 
 β_{4}  t(β_{4}) 
Small  −0.24  −0.16  −0.15  0.01  −0.30  −5.07  −2.81  −2.71  0.15  −3.62 
2  −0.14  −0.20  −0.06  0.06  −0.09  −2.71  −3.11  −0.86  0.91  −1.05 
3  −0.24  −0.10  0.14  0.31  0.11  −3.49  −1.34  1.97  4.44  1.48 
4  −0.10  −0.18  −0.01  0.36  0.23  −1.57  −2.24  −0.16  5.92  3.41 
Big  −0.02  −0.17  0.00  0.21  0.30  −0.30  −1.82  0.03  2.91  5.19 

YIELDHML= Cash Flow–Based Payout Yield 
 α  t(α) 
Small  0.17  0.06  0.33  0.43  0.33  1.34  0.38  2.25  2.37  1.53 
2  −0.15  −0.12  0.17  0.01  −0.03  −1.04  −0.67  0.94  0.05  −0.17 
3  −0.24  −0.28  −0.14  0.04  0.05  −1.28  −1.45  −0.75  0.21  0.27 
4  −0.31  0.00  0.04  −0.31  −0.13  −1.78  0.01  0.17  −1.82  −0.73 
Big  −0.28  −0.26  −0.27  −0.32  0.01  −1.45  −1.06  −1.03  −1.66  0.05 

B/M Quintiles  YIELDHML= Cash Flow–Based Payout Yield 

 β_{4}  t(β_{4}) 
Small  −0.26  −0.17  −0.06  0.26  0.42  −5.18  −2.77  −1.02  3.64  4.95 
2  −0.01  −0.01  −0.16  0.07  0.41  −0.13  −0.18  −2.25  1.03  5.09 
3  −0.18  −0.12  −0.01  0.14  0.22  −2.44  −1.56  −0.15  1.81  2.79 
4  −0.09  −0.18  −0.01  0.05  0.31  −1.27  −2.21  −0.11  0.71  4.42 
Big  −0.19  −0.32  −0.18  0.11  0.19  −2.54  −3.26  −1.67  1.40  3.03 

 YIELDHML= Treasury Stock–Based Payout Yield 

 α  t(α) 
Small  0.11  0.02  0.32  0.51  0.49  0.86  0.13  2.23  2.83  2.24 
2  −0.13  −0.11  0.12  0.04  0.13  −0.93  −0.65  0.68  0.21  0.61 
3  −0.28  −0.33  −0.16  0.05  0.13  −1.52  −1.77  −0.85  0.27  0.66 
4  −0.35  −0.08  0.05  −0.30  −0.04  −2.03  −0.37  0.22  −1.81  −0.25 
Big  −0.33  −0.37  −0.37  −0.29  0.02  −1.74  −1.50  −1.39  −1.51  0.13 
 β_{4}  t(β_{4}) 
Small  −0.19  −0.12  −0.06  0.12  0.10  −4.68  −2.60  −1.32  2.05  1.43 
2  −0.07  −0.03  −0.09  0.02  0.10  −1.49  −0.62  −1.55  0.41  1.50 
3  −0.13  0.01  0.03  0.16  0.08  −2.22  0.10  0.55  2.71  1.26 
4  −0.01  −0.02  −0.04  0.05  0.19  −0.13  −0.36  −0.64  0.96  3.34 
Big  −0.11  −0.13  0.02  0.06  0.24  −1.84  −1.60  0.24  0.93  4.87 

YIELDHML= Net Payout Yield 
 α  t(α) 
Small  0.20  −0.11  0.24  0.24  0.32  1.54  −0.74  1.63  1.45  1.45 
2  −0.08  −0.19  0.02  −0.07  −0.08  −0.55  −1.08  0.09  −0.40  −0.38 
3  −0.25  −0.28  −0.06  −0.04  0.02  −1.33  −1.44  −0.31  −0.22  0.08 
4  −0.28  −0.16  0.04  −0.33  −0.14  −1.59  −0.75  0.17  −1.94  −0.76 
Big  −0.25  −0.36  −0.25  −0.26  0.03  −1.28  −1.41  −0.94  −1.32  0.16 
 β_{4}  t(β_{4}) 
Small  −0.26  0.16  0.10  0.55  0.37  −5.22  2.58  1.81  8.46  4.24 
2  −0.13  0.11  0.13  0.20  0.42  −2.28  1.61  1.81  2.86  5.16 
3  −0.13  −0.10  −0.15  0.26  0.25  −1.71  −1.25  −2.05  3.42  3.13 
4  −0.13  0.13  −0.01  0.08  0.27  −1.82  1.51  −0.09  1.16  3.80 
Big  −0.21  −0.10  −0.18  −0.02  0.13  −2.82  −0.96  −1.73  −0.24  2.00 
Panel B: Tests of Intercept Significance 

Portfolio Set  FF 3Factor  FF 3Factor + DYHML  FF 3Factor + PYCFHML  FF 3Factor + PYTSHML  FF 3Factor + NPYHML 

Significant α  χ^{2}  Significant α  χ^{2}  Significant α  χ^{2}  Significant α  χ^{2}  Significant α  χ^{2} 

Beta/payout yield  7  44.77**  2  46.26**  2  36.59  1  41.94*  0  42.03* 
Size/payout yield  4  82.93**  5  82.59**  4  80.53**  5  83.85**  6  81.84** 
Booktomarket/payout yield  8  46.54**  4  48.36**  2  38.25*  4  43.98*  0  33.97 
Before commenting on the results containing the yield factors, it is worthwhile documenting the findings for a conventional threefactor model estimated on the three sets of portfolios described above (beta/payout yield, size/payout yield, and booktomarket/payout yield). Panel B of Table V summarizes the test results. In terms of the number of significant alphas, we find 7, 4, and 8 out of 25, respectively. Of course, these alphas may be correlated, which calls for a joint test. We look at the standard Wald test that the alphas are all equal to zero. The Wald tests produce test statistics of 44.77, 82.93, and 46.54, respectively, all of which are asymptotically distributed χ^{2}(25) and highly statistically significant. The finding that these tests reject the joint hypothesis that all of the intercepts are zero is potentially important. While it is not the first rejection of the FamaFrench model (see Davis, Fama, and French (2000) and Cremers (2003), among others), it does suggest that portfolio returns sorted in some way on payout yield cannot be explained crosssectionally solely by the FamaFrench factors.
To this point, there is some evidence that the payout yield may be a factor in describing expected returns. Across all three cross sections of portfolios sorted on the payout yield and the other factors (i.e., beta, size, booktomarket), the alphas tend to be statistically indistinguishable from zero (see Panel A for the booktomarket/yield portfolios in Table V as a representative sample). For example, Panel B shows that relative to the threefactor model, for portfolios sorted on beta, size, and booktomarket, the number of significant alphas substantially declines. For the dividend yield factor there are 2, 5, and 4, respectively, significant alphas compared to those for the cash flow–based payout yield factor (2, 4, and 2), Treasury stock–based payout yield (1, 5, and 4), and net payout yield factor (0, 6, and 0). While the alphas are probably correlated, suggesting a joint hypothesis, the evidence presented here is suggestive of the importance of a yield factor, albeit it distinguishes somewhat less between the dividend yield and the payout yield relative to previous evidence.
To complete the analysis, we perform a Wald test analogous to the one described above, the results of which are presented in Panel B of Table V. We find a negligible difference when we include the dividend yield factor in the specification. However, when we replace the dividend yield with the payout yield, the test statistics fall uniformly across the portfolios for the Treasury stock–based payout measure and in all but the beta portfolio for the cash flow–based payout measure. A further decline in test statistics obtains when we include the net payout yield factor. In sum, excess returns are driven to zero, or generally closer in the case of size portfolios, as we progress from the FamaFrench threefactor model to a model that includes the dividend yield, then payout yield, and, finally, net payout yield.
In terms of the coefficients on the payout yield, between onethird and onehalf of them are significant in the regressions, which suggests that they have useful information for describing crosssectional variation that is above and beyond the usual factors. Moreover, the coefficients follow sensible patterns, such as a positive correlation between the yield factor coefficient and payoutsorted portfolios. For each (net) payout measure, the estimated slope coefficients in the low yield portfolio are all negative while those in the highest yield portfolio are all positive. Thus, independent of the booktomarket portfolio, the coefficients tend to increase across the yield portfolios. This finding is consistent with the results of Table III on the relation between average returns and payout yields, and shows that it carries through even in the presence of the welldocumented threefactor model of Fama and French. Also consistent with our previous results, the strength of the association between the estimated yield coefficients and the yield portfolios appears to strengthen as we progress from the dividend yield to the payout yield to the net payout yield.
In concert, this crosssectional evidence suggests that including repurchases has additional explanatory power for expected returns, and that these yields generally outperform dividend yields, which supports the measurement issue. The results of Tables IV and V also suggest that investing in highyield stocks, especially when the yield measure includes repurchases, results in higher returns than investing in yieldneutral portfolios. These findings ultimately beg the question: How does the strategy of investing in high yield portfolios perform over time?
To illustrate the applicability of this analysis, we analyze the performance of various yield portfolios. We consider the popular Dogs of the Dow trading strategy and variations of that strategy based on our discussion. In its simplest form, this strategy amounts to buying highyield Dow Jones Index stocks (say a third). In Malkiel's (2003) wellknown book, A Random Walk Down Wall Street, he describes how this strategy historically outperformed the Dow Jones by 2% to 3% per annum. Malkiel goes on to say, however, that once this strategy became popular, the returns disappeared—in his language, “the dogs no longer hunt” (p. 246). Our analysis suggests an alternative explanation.
Panel A of Table VI describes the monthly returns to buying portfolios of stocks formed on various yield measures during the period July 1984 to December 2003. Specifically, in June of each year, stocks are sorted into deciles based on their yield from the previous year. The High (Low) Yield portfolio consists of those stocks falling in the upper (lower) 30% of the nonzero yield distribution. While the average monthly return of holding the market over this period is 1.12%, the return corresponding to the top dividend yield portfolio is 1.35%, a 20 basis point difference. However, when we consider the high (net) payout yield portfolios, we see average returns of 1.57%, 1.53%, and 1.59% to the cash flow–based payout yield, Treasury stock–based payout yield, and net payout yield, respectively. These returns effectively double the spread over the market return exhibited by the dividend yield portfolio. Turning to the risk characteristics of these portfolios, Panel A also presents their factor loadings, which reveal positive loadings on each of the factors, but statistically significant intercepts that range from 59 basis points per month for the dividend yield portfolio to 80 basis points per month for the net payout yield portfolio.
Table VI. Monthly Return Summary Statistics for Yield Portfolios The sample consists of all nonfinancial firms in the intersection of CRSP and Compustat with data for dividends paid to common shareholders, repurchases of common stock, and sales of common stock. The dividend (payout) [net payout] yield is the ratio of common dividends (dividends plus common share repurchases) [dividends plus repurchases minus common share issuances] in year t to yearend market capitalization. To mitigate the effect of outliers, we trim the upper 5% of the dividend and payout yield distributions, and the upper and lower 2.5% of the net payout yield distribution. We also require that firms have at least 2 years worth of historical return data available on CRSP. All stocks with nonmissing yield values are then allocated to 10 yield portfolios using NYSE breakpoints based on positive yields (nonzero yields in the case of net payout). Each portfolio's monthly equalweighted return for July of year t to June of year t+ 1 is calculated, and then the portfolios are reformed in July of year t+ 1. Both panels present results over the period July 1984 to December 2003. Panel A presents high yield portfolios formed from the top 30% of the yield distribution. Panel B presents high minus low yield portfolios formed by subtracting the returns to the bottom 30% of the yield distribution (excluding zero yield stocks) from the top 30% of the yield distribution. Both panels present average monthly returns during the period July 1984 to December 2003. Additionally, factor loadings and tstatistics are presented from timeseries regressions of portfolio returns on the market excess return (RM_{t}−Rf_{t}), the small minus big factor return (SMB_{t}), and the high minus low factor return (HML_{t}), all obtained from Ken French's website. tstatistics are presented in parentheses. Panel A: High Yield Portfolios 

 Market  High Dividend Yield  High Total Payout [CF] Yield  High Total Payout [TS] Yield  High Total Net Payout [CF] Yield 

Mean monthly return  1.12  1.35  1.57  1.53  1.59 

Factor Loadings and tStatistics 
Intercept   0.59  0.79  0.77  0.80 
(8.07)  (9.85)  (10.10)  (10.87) 
RM_{t}−Rf_{t}   0.79  0.84  0.80  0.84 
(43.83)  (41.70)  (42.15)  (45.41) 
SMB_{t}   0.32  0.53  0.46  0.52 
(14.13)  (21.20)  (19.53)  (22.68) 
HML_{t}   0.53  0.47  0.49  0.47 
(19.60)  (15.66)  (16.97)  (17.14) 
Panel B: Long High Yield Stocks and Short Low Yield Stocks 

 Market  Dividend Yield (High – Low)  Payout Yield [CF] (High – Low)  Payout Yield [TS] (High – Low)  Net Payout Yield [CF] (High – Low) 

Mean monthly return  1.12  0.18  0.28  0.25  0.37 

Factor Loadings and tStatistics 
Intercept   0.31  0.38  0.37  0.43 
(3.15)  (3.92)  (10.10)  (2.83) 
RM_{t}−Rf_{t}   −0.25  −0.24  −0.26  −0.27 
(−10.18)  (−9.98)  (42.15)  (−7.28) 
SMB_{t}   −0.15  −0.19  −0.23  −0.49 
(−4.81)  (−6.51)  (19.53)  (−10.64) 
HML_{t}   0.14  0.23  0.20  0.44 
(3.68)  (6.27)  (16.97)  (7.74) 
Panel B of Table VI presents a similar analysis for portfolios that simultaneously go long in the high yield portfolio and short in the low yield portfolio. Several observations are worth mentioning. First, the payout and the net payout yield have significantly higher returns than the dividend yield strategy. Second, this zerofinance strategy has a positive alpha, which is highest for the payout and net payout yield. Third, regardless of how we measure the yield, this strategy has negative loadings on the market and size factors, and a positive loading on the booktomarket factor. Finally, the payout yield and net payout yield portfolios exhibit substantial improvements in performance relative to the dividend yield portfolio. Figure 3 presents a graphical view of the performance of these portfolios over time and illustrates both the evolution of the yield factor and the viability of our findings as a trading strategy. We note that in most years the dividend, payout (cash flow and Treasury stock measures), and the net payout strategies were profitable (in 13 out of 19, 14 out of 19, 13 out of 18, and 12 out of 19 years, respectively).
Perhaps the most glaring result is the large negative return to the net payout yield portfolio in 2000, which stands in stark contrast to the other two portfolio returns. Closer examination of this result reveals that it is due primarily to a subset of firms that issued equity during 1998 (i.e., at just the right time) and realized significantly large subsequent returns from July of 1999 to June of 2000. These firms fall predominantly in the hightech and biotech industries (SIC codes 7372, 7373, 7370, 2834, 2835, and 2836).
Overall, the results indicate that even after controlling for alternative risk factors, these strategies appear to earn abnormal returns, as measured by the significantly positive intercepts. These returns are higher when repurchases and issuances are accounted for. At the same time, the analysis illustrates that while following this strategy is profitable, it is not an arbitrage, as evidenced by significant losses in a few years.