Table VI-Continued

Table VI-Continued

return not associated with earnings news is associated with more persistent drifts in future returns.

One possible explanation for the larger return spreads associated with price momentum, compared with earnings momentum, is as follows. Our earnings momentum strategies are based on the performance of near-term income: the innovations in quarterly earnings, or analysts forecasts of earnings for the current fiscal year. In comparison, when we select a stock on the basis of high or low prior returns, we isolate cases where the market has made very large revisions in its expectations of the firms future outlook. Table II confirms that the highest-ranked portfolio in our price momentum strategy rose in price by roughly 70 percent on average, while the lowest-ranked portfolio fell in price by about 30 percent on average, over the previous six months. It is unlikely that changes of this magnitude arise solely from quarter-to-quarter news in earnings. The corresponding past six-month returns of the portfolio ranked highest (lowest) by analyst revisions, for example, is about 25 percent (-7 percent). Since there has been a larger reappraisal of market beliefs for the price momentum portfolios, and given that the markets adjustment is not immediate, it is perhaps not surprising that the spread in future returns continues to be larger for the price momentum strategy.

In a similar vein, the difference in the persistence of the two strategies has some intuitive basis. The uncertainty underlying the short-horizon measures of profitability used in the earnings momentum strategies is resolved relatively quickly. Prior returns, on the other hand, reflect a broad set of market expectations not limited to near-term profitability. On this basis, we conjecture that it may take longer for the new information to be played out in stock prices for the price momentum strategy.

Panels D and E pit our measures of earnings surprise against each other. In general, each measure of surprise has incremental predictive power for returns and they give rise to similar spreads in average returns. Holding SUE fixed, for example, portfolios sorted by analyst revisions generate average spreads in six-month returns of 3.23 percent; classifying by SUE while holding fixed analyst revisions yields average spreads of 3.37 percent in six-month returns. Similarly, in Panel E, the sorts by REV6 and ABR yield average spreads of 4.90 and 2.70 percent, respectively, in six-month returns. No single measure of the

news in earnings wins the contest; instead, they each add separate pieces of information, as noted in our introduction.

B. Cross-Sectional Regressions

We use Fama-MacBeth (1973) cross-sectional regressions as another way to disentangle price and earnings momentum. Every month, we fit a cross-sectional regression of individual stock returns on the prior six-month return and various measures of the most recent past earnings surprise (SUE, ABR, and REV6). We also include firm size as a catch-all variable for other influences on the cross-section of returns. To account for possible nonlinearities in the relation, in the monthly regressions we first express each explanatory variable in terms of its ordinal ranking and then scale it to lie between zero and one. This has the added benefit of expressing all the explanatory variables on a common scale, so that their coefficients can be directly compared. The dependent variable is either the buy-and-hold return over the subsequent six months or over the first postformation year. Table VII reports the time-series averages of the slope coefficients, and their -statistics. Since the dependent variable in each monthly regression is a return measured over overlapping intervals, the -statistics are corrected for autocorrelation. The standard error of the time series of coefficients from the regression for six-month (twelvemonth) returns is adjusted for a fifth-order (eleventh-order) moving average process.

Prior return and earnings surprise, taken separately, are each strongly and positively related to future six-month returns (Panel A). The average slope from the regressions of returns on prior return alone is 5.7 percent, which is 4.1 times its standard error. In comparison, using either SUE or REV6 as the predictor variable gives very similar average slopes (6 percent), while the average slope for ABR is smaller (3.7 percent). In all cases, the coefficients are large relative to their standard errors.

The regression with all three measures of earnings surprise yields average slopes that are reliably different from zero, confirming our earlier impression that each adds information not contained in the other two. All four momentum variables are considered simultaneously in the last regression. Earnings surprises rob past return of some, but not all, of its predictive power. The coefficient for prior return falls from 5.7 percent when it is the only momentum variable to 2.9 percent in the full regression model. In this latter equation, past standardized unexpected earnings and revisions in analyst forecasts, with average coefficients of 3.2 and 3.1 percent, respectively, are just as important as prior return in predicting returns over the following six months.

The results from regressions for twelve-month returns are reported in Panel В of Table VII. When past return is the only momentum variable, its average slope is 10.3 percent. Introducing earnings surprises into the equation knocks the estimated effect down to 7.6 percent. Nonetheless, the average slope on past return is large not only relative to its standard error, but also compared to the slopes on the other earnings surprise variables in the last regression.