difference in returns is also obtained when sorting on past analyst revisions. The one-year spread in this case is 7.6 percent (compared to 9.7 percent for the entire sample in Table V). When past SUE or past announcement return is the ranking variable, the one-year spreads are 2.9 and 4.3 percent, respectively (the corresponding spreads based on the entire sample are 7.5 and 8.3 percent).

Panel В of Table VIII replicates our two-way sorts on the larger stocks. Compared to the entire sample, the large-stock sample displays smaller differences in returns between the highest-ranked and lowest-ranked portfolios. Nonetheless, the spreads remain large: 8.4 percent for the two-way classification based on prior return and announcement return, 7.7 percent based on prior return and SUE, and 8.5 percent based on prior return and revisions in consensus estimates. Although sorting by prior return conditional on past earnings news gives rise to larger differences in subsequent returns, earnings surprises still have some marginal explanatory power. For example, the average one-year spread across prior return ranks, holding fixed the rank by standardized unexpected earnings, is 5.7 percent. The average spread associated with standardized unexpected earnings, conditional on prior return, is 2.1 percent. Earnings news have a lesser impact on the returns of large companies because there are numerous additional sources of information about the outlook for these companies.

Table IX fits cross-sectional regressions to future twelve-month returns for the large-stock sample. The regressions support the results from the earlier panels in the table. In the univariate regressions, for example, each momentum variable is statistically significant. When they are considered together in the last regression, the most important variable is the prior six-month return; its average coefficient is 6.4 percent which is more than two standard errors away from zero.

B. Adjusting for Size and Book-to-Market Factors

Our earlier results in Tables II to V raise the possibility that the predictive power of prior returns or prior earnings surprises may be confounded with the effects of book-to-market or firm size. In this section we investigate whether the behavior of returns on our different momentum portfolios can be explained by factors related to size and book-to-market. This is done in the context of the Fama-French (1993) three-factor model, given by time series regressions of the form

rpt - rft = ap + bp{rmt - rft) + spSMB, + /*PHML, + spt. (4)

Here rpt is the return on portfolio p in month t; rft and rmt are the Treasury bill rate and the return on the value-weighted market index, respectively; SMB is the return on the mimicking portfolio for size; and HML is the return on the mimicking portfolio for book-to-market.11 If the momentum strategies

11 We thank Eugene Fama for providing the data on the mimicking portfolio returns.

i i

Table VIII

Mean Returns for Portfolios Based on Large Firms

The sample includes all New York Stock Exchange (NYSE), American Stock Exchange (AMEX), and Nasdaq domestic primary issues with coverage on the Center for Research in Security Prices (CRSP) and COMPUSTAT, and with beginning-of-month market value of equity above the median market capitalization of NYSE issues. Eligible stocks are ranked and grouped into portfolios on the basis of one classification variable (Panel A) or two classification variables (Panel B). Portfolios are formed at the beginning of every month from January 1977 to January 1993. The assignment of stocks to portfolios uses breakpointsbased on NYSE issues only. All stocks are equally-weighted in a portfolio, and average buy-and-hold returns are reported for the first year after portfolio formation. In Panel A the classification variable is either the stocks compound return over the prior six months (R6), standardized unexpected earnings (SUE, the change in most recently announced quarterly earnings per share from its value four quarters ago, divided by the standard deviation of unexpected earnings over the last eight quarters), abnormal returns relative to the equally-weighted market index cumulated from two days before to one day after the date of the most recent past earnings announcement (ABR), or a moving average of the prior six months percentage revisions relative to the beginning-of-month stock price in mean I/B/E/S estimates of current fiscal-year earnings per share (REV6). In Panel B, portfolios are formed from the intersections of independent sorts by prior return and by one measure of earnings surprise (standardized unexpected earnings, cumulative abnormal return around earnings announcement or moving average of analysts revisions).

Panel A: Mean Return in First Postformation Year from One-Way Classifications

Panel B: Mean Return in First Postformation Year from Two-Way Classifications

and prior return

performance is just a manifestation of size and book-to-market effects, then the intercept of the equation, ap, should not be significantly different from zero.

Fama and French (1996) use equation (4) to analyze the performance of portfolios sorted by prior return. Here we examine the evidence when earnings momentum is brought into the picture as well.12 In particular, we focus on the double-sort portfolios based on prior return and revisions in consensus esti-

12 Fama and French (1994) report that the portfolio of losers, compared to the portfolio of winners, loads more heavily on the size and book-to-market factors. The difference in intercepts between the top and bottom deciles is 1.74 percent per month. We find quite similar results.

Table IX

Monthly Cross-Sectional Regressions of Returns on Prior Return and Prior Earnings Surprises, Using Large Firms Only

The sample includes all New York Stock Exchange (NYSE), American Stock Exchange (AMEX), and Nasdaq domestic primary issues with coverage on the Center for Research in Security Prices (CRSP) and COMPUSTAT, and with beginning-of-month market value of equity above the median market capitalization of NYSE issues. Cross-sectional regressions are estimated each month from January 1977 to January 1993. The dependent variable is each stocks one-year buy-and-hold return. The explanatory variables are firm size and the following. R6 is the stocks compound return over the prior six months, SUE is the change in most recently announced quarterly earnings per share from its value four quarters ago, divided by the standard deviation of unexpected earnings over the last eight quarters, ABR is the abnormal return relative to the equally-weighted market index cumulated from two days before to one day after the date of the most recent past earnings announcement, and REV6 is a moving average of the prior six months percentage revisions relative to the beginning-of-month stock price in mean I/B/E/S estimates of current fiscal-year earnings per share. The reported statistics are the means of the time series of coefficients from the month-by-month regressions, and in parentheses the -statistics relative to the autocorrelation-adjusted standard error of the mean.

mates. Table X reports summary statistics of the time series regressions for the highest- and lowest-ranked portfolios (portfolios (3,3) and (1,1) respectively in Panel С of Table VI). We track the monthly returns from a strategy of buying each portfolio and holding it for six months, when a new portfolio is formed and the process repeated. Table X also reports results for the arbitrage portfolio formed by buying the highest-ranked portfolio, or the winners, and selling the lowest-ranked portfolio, or the losers.

The portfolios of winners and losers have very similar market risk exposures (bp). In other respects, the results in Table X generally confirm our earlier findings. Both portfolios load significantly on size. The portfolio of winners concentrates more heavily on glamour stocks, so it loads negatively on the book-to-market factor, while the portfolio of losers is more oriented towards value stocks, and so loads positively on the book-to-market factor. The main conclusion from Table X is that adjusting for size and book-to-market does not change the observed pattern in returns. The intercept for the loser portfolio (-0.953 percent per month) is especially eye-catching. This poor performance