Efficient Capital Markets: II

Sequels are rarely as good as the originals, so I approach this review of the market efficiency literature with trepidation. The task is thornier than it was 20 years ago, when work on efficiency was rather new. The literature is now so large that a full review is impossible, and is not attempted here. Instead, I discuss the work that I find most interesting, and I offer my views on what we have learned from the research on market efficiency. I take the market efficiency hypothesis to be the simple statement that security prices fully reflect all available information. A precondition for this strong version of the hypothesis is that information and trading costs, the costs of getting prices to reflect information, are always 0 (Grossman and Stiglitz (1980)). A weaker and economically more sensible version of the efficiency hypothesis says that prices reflect information to the point where the marginal benefits of acting on information (the profits to be made) do not exceed the marginal costs (Jensen (1978)). Since there are surely positive information and trading costs, the extreme version of the market efficiency hypothesis is surely false. Its advantage, however, is that it is a clean benchmark that allows me to sidestep the messy problem of deciding what are reasonable information and trading costs. I can focus instead on the more interesting task of laying out the evidence on the adjustment of prices to various kinds of information. Each reader is then free to judge the scenarios where market efficiency is a good approximation (that is, deviations from the extreme version of the efficiency hypothesis are within information and trading costs) and those where some other model is a better simplifying view of the world. Ambiguity about information and trading costs is not, however, the main obstacle to inferences about market efficiency. The joint-hypothesis problem is more serious. Thus, market efficiency per se is not testable. It must be tested jointly with some model of equilibrium, an asset-pricing model. This point, the theme of the 1970 review (Fama (1970b)), says that we can only test whether information is properly reflected in prices in the context of a pricing model that defines the meaning of “properly.” As a result, when we find anomalous evidence on the behavior of returns, the way it should be split between market inefficiency or a bad model of market equilibrium is ambiguous. Does the fact that market efficiency must be tested jointly with an equilibrium-pricing model make empirical research on efficiency uninteresting? Does the joint-hypothesis problem make empirical work on asset-pricing models uninteresting? These are, after all, symmetric questions, with the same answer. My answer is an unequivocal no. The empirical literature on efficiency and asset-pricing models passes the acid test of scientific usefulness. It has changed our views about the behavior of returns, across securities and through time. Indeed, academics largely agree on the facts that emerge from the tests, even when they disagree about their implications for efficiency. The empirical work on market efficiency and asset-pricing models has also changed the views and practices of market professionals. As these summary judgements imply, my view, and the theme of this paper, is that the market efficiency literature should be judged on how it improves our ability to describe the time-series and cross-section behavior of security returns. It is a disappointing fact that, because of the joint-hypothesis problem, precise inferences about the degree of market efficiency are likely to remain impossible. Nevertheless, judged on how it has improved our understanding of the behavior of security returns, the past research on market efficiency is among the most successful in empirical economics, with good prospects to remain so in the future. The 1970 review divides work on market efficiency into three categories: (1) weak-form tests (How well do past returns predict future returns?), (2) semi-strong-form tests (How quickly do security prices reflect public information announcements?), and (3) strong-form tests (Do any investors have private information that is not fully reflected in market prices?) At the risk of damning a good thing, I change the categories in this paper. Instead of weak-form tests, which are only concerned with the forecast power of past returns, the first category now covers the more general area of tests for return predictability, which also includes the burgeoning work on forecasting returns with variables like dividend yields and interest rates. Since market efficiency and equilibrium-pricing issues are inseparable, the discussion of predictability also considers the cross-sectional predictability of returns, that is, tests of asset-pricing models and the anomalies (like the size effect) discovered in the tests. Finally, the evidence that there are seasonals in returns (like the January effect), and the claim that security prices are too volatile are also considered, but only briefly, under the rubric of return predictability. For the second and third categories, I propose changes in title, not coverage. Instead of semi-strong-form tests of the adjustment of prices to public announcements, I use the now common title, event studies. Instead of strong-form tests of whether specific investors have information not in market prices, I suggest the more descriptive title, tests for private information. Return predictability is considered first, and in the most detail. The detail reflects my interest and the fact that the implications of the evidence on the predictability of returns through time are the most controversial. In brief, the new work says that returns are predictable from past returns, dividend yields, and various term-structure variables. The new tests thus reject the old market efficiency-constant expected returns model that seemed to do well in the early work. This means, however, that the new results run head-on into the joint-hypothesis problem: Does return predictability reflect rational variation through time in expected returns, irrational deviations of price from fundamental value, or some combination of the two? We should also acknowledge that the apparent predictability of returns may be spurious, the result of data-dredging and chance sample-specific conditions. The evidence discussed below, that the variation through time in expected returns is common to corporate bonds and stocks and is related in plausible ways to business conditions, leans me toward the conclusion that it is real and rational. Rationality is not established by the existing tests, however, and the joint-hypothesis problem likely means that it cannot be established. Still, even if we disagree on the market efficiency implications of the new results on return predictability, I think we can agree that the tests enrich our knowledge of the behavior of returns, across securities and through time. Event studies are discussed next, but briefly. Detailed reviews of event studies are already available, and the implications of this research for market efficiency are less controversial. Event studies have, however, been a growth industry during the last 20 years. Moreover, I argue that, because they come closest to allowing a break between market efficiency and equilibrium-pricing issues, event studies give the most direct evidence on efficiency. And the evidence is mostly supportive. Finally, tests for private information are reviewed. The new results clarify earlier evidence that corporate insiders have private information that is not fully reflected in prices. The new evidence on whether professional investment managers (mutual fund and pension fund) have private information is, however, murky, clouded by the joint-hypothesis problem. There is a resurgence of research on the time-series predictability of stock returns, that is, the variation (rational or irrational) of expected returns through time. Unlike the pre-1970 work, which focused on forecasting returns from past returns, recent tests also consider the forecast power of variables like dividend yields D/P), earnings/price ratios (E/P), and term-structure variables. Moreover, the early work concentrated on the predictability of daily, weekly, and monthly returns, but the recent tests also examine the predictability of returns for longer horizons. Among the more striking new results are estimates that the predictable component of returns is a small part of the variance of daily, weekly, and monthly returns, but it grows to as much as 40% of the variance of 2– to 10-year returns. These results have spurred a continuing debate on whether the predictability of long-horizon returns is the result of irrational bubbles in prices or large rational swings in expected returns. I first consider the research on predicting returns from past returns. Next comes the evidence that other variables (D/P, E/P, and term-structure variables) forecast returns. The final step is to discuss the implications of this work for market efficiency. In the pre-1970 literature, the common equilibrium-pricing model in tests of stock market efficiency is the hypothesis that expected returns are constant through time. Market efficiency then implies that returns are unpredictable from past returns or other past variables, and the best forecast of a return is its historical mean. The early tests often find suggestive evidence that daily, weekly, and monthly returns are predictable from past returns. For example, Fama (1965) finds that the first-order autocorrelations of daily returns are positive for 23 of the 30 Dow Jones Industrials and more than 2 standard errors from 0 for 11 of the 30. Fisher's (1966) results suggest that the autocorrelations of monthly returns on diversified portfolios are positive and larger than those for individual stocks. The evidence for predictability in the early work often lacks statistical power, however, and the portion of the variance of returns explained by variation in expected returns is so small (less than 1% for individual stocks) that the hypothesis of market efficiency and constant expected returns is typically accepted as a good working model. In recent work, daily data on NYSE and AMEX stocks back to 1962 [from the Center for Research in Security Prices (CRSP)] makes it possible to estimate precisely the autocorrelation in daily and weekly returns. For example, Lo and MacKinlay (1988) find that weekly returns on portfolios of NYSE stocks grouped according to size (stock price times shares outstanding) show reliable positive autocorrelation. The autocorrelation is stronger for portfolios of small stocks. This suggests, however, that the results are due in part to the nonsynchronous trading effect (Fisher 1966). Fisher emphasizes that spurious positive autocorrelation in portfolio returns, induced by nonsynchronous closing trades for securities in the portfolio, is likely to be more important for portfolios tilted toward small stocks. To mitigate the nonsychronous trading problem, Conrad and Kaul (1988) examine the autocorrelation of Wednesday-to-Wednesday returns for size-grouped portfolios of stocks that trade on both Wednesdays. Like Lo and MacKinlay (1988), they find that weekly returns are positively autocorrelated, and more so for portfolios of small stocks. The first-order autocorrelation of weekly returns on the portfolio of the largest decile of NYSE stocks for 1962–1985 is only .09. For the portfolios that include the smallest 40% of NYSE stocks, however, first-order autocorrelations of weekly returns are around .3, and the autocorrelations of weekly returns are reliably positive out to 4 lags. The results of Lo and MacKinlay (1988) and Conrad and Kaul (1988) show that, because of the variance reduction obtained from diversification, portfolios produce stronger indications of time variation in weekly expected returns than individual stocks. Their results also suggest that returns are more predictable for small-stock portfolios. The evidence is, however, clouded by the fact that the predictability of portfolio returns is in part due to nonsynchronous trading effects that, especially for small stocks, are not completely mitigated by using stocks that trade on successive Wednesdays. An eye-opener among recent studies of short-horizon returns is French and Roll (1986). They establish an intriguing fact. Stock prices are more variable when the market is open. On an hourly basis, the variance of price changes is 72 times higher during trading hours than during weekend nontrading hours. Likewise, the hourly variance during trading hours is 13 times the overnight nontrading hourly variance during the trading week. One of the explanations that French and Roll test is a market inefficiency hypothesis popular among academics; specifically, the higher variance of price changes during trading hours is partly transistory, the result of noise trading by uniformed investors (e.g., Black (1986)). Under this hypothesis, pricing errors due to noise trading are eventually reversed, and this induces negative autocorrelation in daily returns. French and Roll find that the first-order autocorrelations of daily returns on the individual stocks of larger (the top three quintiles of) NYSE firms are positive. Otherwise, the autocorrelations of daily returns on individual stocks are indeed negative, to 13 lags. Although reliably negative on a statistical basis, however, the autocorrelations are on average close to 0. Few are below −01. One possibility is that the transitory price variation induced by noise trading only dissipates over longer horizons. To test this hypothesis, French and Roll examine the ratios of variances of N-period returns on individual stocks to the variance of daily returns, for N from 2 days to 6 months. If there is no transitory price variation induced by noise trading (specifically, if price changes are i.i.d.), the N-period variance should grow like N, and the variance ratios (standardized by N) should be close to 1. On the other hand, with transitory price variation, the N-period variance should grow less than in proportion to N, and the variance ratios should be less than 1. For horizons (N) beyond a week, the variance ratios are more than 2 standard errors below 1, except for the largest quintile of NYSE stocks. But the fractions of daily return variances due to transitory price variation are apparently small. French and Roll estimate that for the average NYSE stock, the upper bound on the transitory portion of the daily variance is 11.7%. Adjusted for the spurious negative autocorrelation of daily returns due to bid-ask effects (Roll (1984)), the estimate of the transitory portion drops to 4.1%. The smallest quintile of NYSE stocks produces the largest estimate of the transitory portion of price variation, an upper bound of 26.9%. After correction for bid-ask effects, however, the estimate drops to 4.7%-hardly a number on which to conclude that noise trading results in substantial market inefficiency. French and Roll (1986, p. 23) conclude, “pricing errors … have a trivial effect on the difference between trading and non-trading variances. We conclude that this difference is caused by differences in the flow of information during trading and non-trading hours.” In short, with the CRSP daily data back to 1962, recent research is able to show confidently that daily and weekly returns are predictable from past returns. The work thus rejects the old market efficiency-constant expected returns model on a statistical basis. The new results, however, tend to confirm the conclusion of the early work that, at least for individual stocks, variation in daily and weekly expected returns is a small part of the variance of returns. The more striking, but less powerful, recent evidence on the predictability of returns from past returns comes from long-horizon returns. The early literature does not interpret the autocorrelation in daily and weekly returns as important evidence against the joint hypothesis of market efficiency and constant expected returns. The argument is that, even when the autocorrelations deviate reliably from 0 (as they do in the recent tests), they are close to 0 and thus economically insignificant. The view that autocorrelations of short-horizon returns close to 0 imply economic insignificance is challenged by Shiller (1984) and Summers (1986). They present simple models in which stock prices take large slowly decaying swings away from fundamental values (fads, or irrational bubbles), but short-horizon returns have little autocorrelation. In the Shiller-Summers model, the market is highly inefficient, but in a way that is missed in tests on short-horizon returns. To illustrate the point, suppose the fundamental value of a stock is constant and the unconditional mean of the stock price is its fundamental value. Suppose daily prices are a first-order autoregression (AR1) with slope less than but close to 1. All variation in the price then results from long mean-reverting swings away from the constant fundamental value. Over short horizons, however, an AR1 slope close to 1 means that the price looks like a random walk and returns have little autocorrelation. Thus in tests on short-horizon returns, all price changes seem to be permanent when fundamental value is in fact constant and all deviations of price from fundamental value are temporary. In his comment on Summers (1986), Stambaugh (1986) points out that although the Shiller-Summers model can explain autocorrelations of short-horizon returns that are close to 0, the long swings away from fundamental value proposed in the model imply that long-horizon returns have strong negative autocorrelation. (In the example above, where the price is a stationary AR1, the autocorrelations of long-horizon returns approach −0.5.) Intuitively, since the swings away from fundamental value are temporary, over long horizons they tend to be reversed. Another implication of the negative autocorrelation induced by temporary price movements is that the variance of returns should grow less than in proportion to the return horizon. The Shiller-Summers challenge spawned a series of papers on the predictability of long-horizon returns from past returns. The evidence at first seemed striking, but the tests turn out to be largely fruitless. Thus, Fama and French (1988a) find that the autocorrelations of returns on diversified portfolios of NYSE stocks for the 1926–1985 period have the pattern predicted by the Shiller-Summers model. The autocorrelations are close to 0 at short horizons, but they become strongly negative, around − 0.25 to − 0.4, for 3– to 5-year returns. Even with 60 years of data, however, the tests on long-horizon returns imply small sample sizes and low power. More telling, when Fama and French delete the 1926–1940 period from the tests, the evidence of strong negative autocorrelation in 3– to 5-year returns disappears. Similarly, Poterba and Summers (1988) find that, for N from 2 to 8 years, the variance of N-year returns on diversified portfolios grows much less than in proportion to N. This is consistent with the hypothesis that there is negative autocorrelation in returns induced by temporary price swings. Even with 115 years (1871–1985) of data, however, the variance tests for long-horizon returns provide weak statistical evidence against the hypothesis that returns have no autocorrelation and prices are random walks. Finally, Fama and French (1988a) emphasize that temporary swings in stock prices do not necessarily imply the irrational bubbles of the Shiller-Summers model. Suppose (1) rational pricing implies an expected return that is highly autocorrelated but mean-reverting, and (2) shocks to expected returns are uncorrelated with shocks to expected dividends. In this situation, expected-return shocks have no permanent effect on expected dividends, discount rates, or prices. A positive shock to expected returns generates a price decline (a discount rate effect) that is eventually erased by the temporarily higher expected returns. In short, a ubiquitous problem in time-series tests of market efficiency, with no clear solution, is that irrational bubbles in stock prices are indistinguishable from rational time-varying expected returns. DeBondt and Thaler (1985, 1987) mount an aggressive empirical attack on market efficiency, directed at unmasking irrational bubbles. They find that the NYSE stocks identified as the most extreme losers over a 3– to 5-year period tend to have strong returns relative to the market during the following years, expecially in January of the following years. Conversely, the stocks identified as extreme winners tend to have weak returns relative to the market in subsequent years. They attribute these results to market overreaction to extreme bad or good news about firms. Chan (1988) and Ball and Kothari (1989) argue that the winner-loser results are due to failure to risk-adjust returns. (DeBondt and Thaler (1987) disagree.) Zarowin (1989) finds no evidence for the DeBondt-Thaler hypothesis that the winner-loser results are due to overreaction to extreme changes in earnings. He argues that the winner-loser effect is related to the size effect of Banz (1981); that is, small stocks, often losers, have higher expected returns than large stocks. Another explanation, consistent with an efficient market, is that there is a risk factor associated with the relative economic performance of firms (a distressed-firm effect) that is compensated in a rational equilibrium-pricing model (Chan and Chen (1991)). We may never be able to say which explanation of the return behavior of extreme winners and losers is correct, but the results of DeBondt and Thaler and their critics are nevertheless interesting. (See also Jagedeesh (1990), Lehmann (1990), and Lo and MacKinlay (1990), who find reversal behavior in the weekly and monthly returns of extreme winners and losers. Lehmann's weekly reversals seem to lack economic significance. When he accounts for spurious reversals due to bouncing between bid and ask prices, trading costs of 0.2% per turnaround transaction suffice to make the profits from his reversal trading rules close to 0. It is also worth noting that the short-term reversal evidence of Jegadeesh, Lehmann, and Lo and MacKinlay may to some extent be due to CRSP data errors, which would tend to show up as price reversals.) The univariate tests on long-horizon returns of Fama and French (1988a) and Poterba and Summers (1988) are a statistical power failure. Still, they provide suggestive material to spur the search for more powerful tests of the hypothesis that slowly decaying irrational bubbles, or rational time-varying expected returns, are important in the long-term variation of prices. There is a simple way to see the power problem. An autocorrelation is the slope in a regression of the current return on a past return. Since variation through time in expected returns is only part of the variation in returns, tests based on autocorrelations lack power because past realized returns are noisy measures of expected returns. Power in tests for return predictability can be enhanced if one can identify forecasting variables that are less noisy proxies for expected returns that past returns. There is no lack of old evidence that short-horizon returns are predictable from other variables. A puzzle of the 1970's was to explain why monthly stock returns are negatively related to expected inflation (Bodie (1976), Nelson (1976), Jaffe and Mandelker (1976), Fama (1981)) and the level of short-term interest rates (Fama and Schwert (1977)). Like the autocorrelation tests, however, the early work on forecasts of short-horizon returns from expected inflation and interest rates suggests that the implied variation in expected returns is a small part of the variance of returns—less than 3% for monthly returns. The recent tests suggest, however, that for long-horizon returns, predictable variation is a larger part of return variances. Thus, following evidence (Rozeff (1984), Shiller (1984)) that dividend yields (D/P) forecast short-horizon stock returns, Fama and French (1988b) use D/P to forecast returns on the value-weighted and equally weighted portfolios of NYSE stocks for horizons from 1 month to 5 years. As in earlier work, D/P explains small fractions of monthly and quarterly return variances. Fractions of variance explained grow with the return horizon, however, and are around 25% for 2– to 4-year returns. Campbell and Shiller (1988b) find that E/P ratios, especially when past earnings (E) are averaged over 10–30 years, have reliable forecast power that also increases with the return horizon. Unlike the long-horizon autocorrelations in Fama and French (1988a), the long-horizon forecast power of D/P and E/P is reliable for periods after 1940. Fama and French (1988b) argue that dividend yields track highly autocorrelated variation in expected stock returns that becomes a larger fraction of return variation for longer return horizons. The increasing fraction of the variance of long-horizon returns explained by D/P is thus due in large part to the slow mean reversion of expected returns. Examining the forecast power of variables like D/P and E/P over a range of return horizons nevertheless gives striking perspective on the implications of slow-moving expected returns for the variation of returns. The predictability of stock returns from dividend yields (or E/P) is not in itself evidence for or against market efficiency. In an efficient market, the forecast power of D/P says that prices are high relative to dividends when discount rates and expected returns are low, and vice versa. On the other hand, in a world of irrational bubbles, low D/P signals irrationally high stock prices that will move predictably back toward fundamental values. To judge whether the forecast power of dividend yields is the result of rational variation in expected returns or irrational bubbles, other information must be used. As always, even with such information, the issue is ambiguous. For example, Fama and French (1988b) show that low dividend yields imply low expected returns, but their regressions rarely forecast negative returns for the value and equally weighted portfolios of NYSE stocks. In their data, return forecasts more than 2 standard errors below 0 are never observed, and more than 50% of the forecasts are more than 2 standard errors above 0. Thus there is no evidence that low D/P signals bursting bubbles, that is, negative expected stock returns. A bubbles fan can argue, however, that because the unconditional means of stock returns are high, a bursting bubble may well imply low but not negative expected returns. Conversely, if there were evidence of negative expected returns, an efficient-markets type could argue that asset-pricing models do not say that rational expected returns are always positive. Fama and French (1989) suggest a different way to judge the implications of return predictability for market efficiency. They argue that if variation in expected returns is common to different securities, then it is probably a rational result of variation in tastes for current versus future consumption or in the investment opportunities of firms. They show that the dividend yield on the NYSE value-weighted portfolio indeed forecasts the returns on corporate bonds as well as common stocks. Moreover, two term-structure variables, (1) the default spread (the difference between the yields on lower-grade and Aaa long-term corporate bonds) and (2) the term spread (the difference between the long-term Aaa yield and the yield on 1-month Treasury bills), forecast returns on the value and equally weighted portfolios of NYSE stocks as well as on portfolios of bonds in different (Moodys) rating groups. Keim and Stambaugh (1986) and Campbell (1987) also find that stock and bond returns are predictable from a common set of stock market and term-structure variables. Harvey (1991) finds that the dividend yield on the S&P 500 portfolio and U.S. term-structure variables forecast the returns on portfolios of foreign common stocks, as well as the S&P return. Thus the variation in expected returns tracked by the U.S. dividend yield and term-structure variables is apparently international. Ferson and Harvey (1991) formally test the common expected returns hypothesis. Using the asset-pricing models of Merton (1973) and Ross (1976), they try to link the time-series variation in expected returns, captured by dividend yields and term-structure variables, to the common factors in returns that determine the cross-section of expected returns. They estimate that the common variation in expected returns is about 80% of the predictable time-series variation in the returns on Government bonds, corporate bonds, and common-stock portfolios formed on industry and size. They can't reject the hypothesis that all the time-series variation in expected returns is common. Fama and French (1989) push the common expected returns argument for market efficiency one step further. They argue that there are systematic patterns in the variation of expected returns through time that suggest that it is rational. They find that the variation in expected returns tracked by D/P and the default spread (the slopes in the regressions of returns on D/P or the default spread) increase from high-grade bonds to low-grade bonds, from bonds to stocks, and from large stocks to small stocks. This ordering corresponds to intuition about the risks of the securities. On the other hand, the variation in expected returns tracked by the term spread is similar for all long-term securities (bonds and stocks), which suggests that it reflects variation in a common premium for maturity risks. Finally, Fama and Frenc