Overreaction Bias in Earnings Forecasts

Overreaction in Earnings Forecasts

Werner F. M. De Bondt and Richard H. Thaler

It has long been part of the conventional wisdom on Wall Street that financial markets overreact. Both casual observation and academic research support irrational exuberance (Shiller, 2000). For instance, Shiller’s survey evidence relating to the 1987 stock market crash reveals that investors were reacting to each other’s behavior rather than to hard economic news.

In prior research, we argued that mean reversion in stock prices is evidence of overreaction. In our 1985 paper, we showed that stocks that were extreme losers over an initial 3- to 5-year period earned excess returns over subsequent years. In the 1987 paper, we showed that these excess returns could not easily be attributed to changes in risk, tax effects, or the small firm anomaly. Rather, we argued that the excess returns to losers might be explained by biased expectations of future earnings. We found that the earnings for losing firms had fallen precipitously while their share prices were lagging market performance, but then rebounded strongly over the next few years. Perhaps, we speculated, “the market” did not correctly anticipate this reversal in earnings. This hypothesis, of excessive pessimism about the future prospects of companies that had done poorly, was suggested by the work of Kahneman and Tversky (1973). They found that people’s intuitive forecasts have a tendency to overweight salient information such as recent news, and underweight less salient data such as long-term averages.

Of course, there are many reasons to be skeptical that actual investors and stock market professionals are subject to the same biases as student subjects in laboratory experiments. Those who follow the stock market for a living are experts in their field, have much at stake, and may be driven out of business if they are prone to systematic error (see Stickel, 1992). Therefore, we present here a study of the expectations of one important group of financial market professionals: security analysts who make periodic forecasts of individual company earnings.

This is an interesting group to study on three counts. First, other investigators have found repeatedly that earnings forecasts (and forecast revisions) have an important influence on stock prices (Brown, Foster & Noreen, 1985). Second, past work suggests that analysts are rather good at what they do. For example, analyst forecasts often outperform time series models (see Conroy & Harris, 1987). Finally, the precision of analyst expectations represents a natural upper bound to the quality of the earnings forecasts of less sophisticated agents. After all, most investors do not have the time or the skill to produce their own predictions and, accordingly, they buy (rather than sell) earnings forecasts. Thus, for all of the above reasons, it is particularly interesting to see whether market professionals display any of the biases discovered in studies of nonexpert judgment.

We test for a type of generalized overreaction, the tendency to make forecasts that are too extreme, given the predictive value of the information available to the forecaster. This tendency is well illustrated by an experiment conducted by Kahneman and Tversky. Subjects were asked to predict the future grade point average (GPA) for each of 10 students on the basis of a percentile score of some predictor. Three predictor variables in percentile form were used: GPA, a test of mental concentration, and a test of sense of humor. Obviously, a percentile measure of GPA is a much better predictor of actual GPA than is a measure of mental concentration that, in turn, is likely to be more reliable than information on sense of humor. Therefore, subjects should give forecasts that are much more regressive in the latter two conditions. That is, the forecasts should deviate less from the mean. The results indicated that people were not nearly sensitive enough to this consideration. Subjects that were given a nearly useless predictor (the sense of humor condition) made predictions that were almost as extreme in variation as those given a nearly perfect predictor (the percentile GPA condition). This pattern leads to a systematic bias: forecasts that diverge the most from the mean tend to be too extreme, implying that forecast errors are predictable.

Our study asks whether security analysts display similar biases. Our focus is on forecasted changes in earnings per share (EPS) for 1- and 2-year time horizons. We study two questions: The first is whether forecast errors in EPS are systematically linked to forecasted changes. In particular, are the forecasts too extreme? Are most forecast revisions “up” (“down”) if the analysts initially projected large declines (rises) in EPS? Clearly, under rationality, neither forecast errors nor forecast revisions should ever be predictable from forecasted changes. The second question is whether the bias in the forecasts gets stronger as uncertainty grows and less is (or objectively can be) known about the future.

To investigate this question, we run regressions of the form AC = α + β FC, where AC is the actual change in earnings and FC is the forecasted change. The null hypothesis of rational expectations is that (α, β) = (0, 1). If α is not equal to zero, there is an average error when the forecasted change is zero: optimism if α < 0, or pessimism if α > 0. The overall average error also depends on the slope and the distribution of forecasts. If β is not equal to 1 then the forecasted changes are either too extreme (if β < 1) or not extreme enough (β > 1). Our primary focus is on the estimates of β because there are numerous explanations for the optimism we (and others) find. The two behavioral hypotheses sketched above make the following predictions for our estimates of β:

HYPOTHESIS 1 Forecasted changes are too extreme, so actual changes are less (in absolute value) than predicted: β < 1*.*

HYPOTHESIS 2 The estimated β for the 2-year forecasts is less than the β for the 1-year forecasts.

The data and results below support the behavioral view. Possible sources of the forecast errors are considered afterward.

DATA

The analysts’ earnings forecasts are taken from IBES International. We study forecasts between 1976 and 1984. IBES contacts individual analysts on a regular basis and computes summary data such as means or standard deviations. The data are sold to institutional investors. Updates are available each month. Here, we work with the April and December predictions of EPS for the current as well as the subsequent year. The April forecasts are approximately 1- and 2-year forecasts because we only consider companies with a fiscal year ending in December. For these firms, actually realized earnings are announced during the first few months of the following calendar year.

We match the earnings forecasts with stock returns and accounting numbers. The returns are provided by the Center for Research on Security Prices (CRSP) at the University of Chicago. The accounting data are listed on COMPUSTAT sold by Standard & Poor’s. Because all data sources contain full historical records, no survivorship biases affect the sample selection. We further adjust the data for stock splits and stock dividends so that all current and past returns, earnings figures, and forecasts are expressed on a comparable basis. When necessary, forecasts of fully diluted earnings per share (EPS) are converted to forecasts of primary EPS (excluding extraordinary items).

Although some IBES data are available for approximately 2,300–2,800 companies each year, our annual sample contains many fewer observations. For example, for the 1-year forecasts, the number varies between 461 and 785. This follows from the data selection criteria. Companies only qualify if they have (1) records on IBES, CRSP, and COMPUSTAT; (2) returns for 3 years prior to the forecast month; (3) EPS numbers for 10 years prior to the forecast month; (4) a December fiscal year; and (5) the data needed to compute the variables in the regressions described later.

EMPIRICAL RESULTS

The regressions of actual earnings changes on forecasted changes yield strong evidence for overreaction. The intercepts are significantly negative, consistent with optimism at a forecasted change of zero. More importantly, the slope coefficients are significantly less than 1. For 1-year forecasts, the estimated β is approximately 0.65; for 2-year forecasts, it drops to around 0.46. If changes are expressed as percentages, an additional 1% in the forecasted change corresponds to an increment of 0.65% in the fitted one-year change and 0.46% in the fitted two-year change. Those are slope effects; the full fitted value also includes the intercept. Thus, forecasted changes are systematically too extreme, and the bias intensifies as the horizon lengthens, consistent with both hypotheses. The greater uncertainty associated with longer horizons does not lead analysts to moderate their predictions sufficiently; instead, the forecasts are even more extreme relative to what is warranted.

Further evidence of overreaction comes from the pattern of forecast revisions. Between the April forecasts and the December updates, revisions tend to move the initially extreme projections back toward less extreme values. The reversals amount to roughly 18% of the originally predicted change for the 1-year forecasts and 38% for the 2-year forecasts. Under rationality, such predictable revisions should not exist, yet analysts consistently walk back their earlier overly extreme predictions as new information arrives.

ROBUSTNESS AND ALTERNATIVE INTERPRETATIONS

These results prove robust to controls for data errors and stale forecasts, and they have been replicated in other settings. International data (Capstaff et al., 2001) and forecasts for other variables such as exchange rates and macroeconomic indicators (Froot & Frankel, 1989; Ahlers & Lakonishok, 1983) show similar patterns. Similar patterns across domains are consistent with a shared judgment bias, although they do not by themselves identify its cause.

The primary explanation offered for this bias is that analysts, like the student subjects in Kahneman and Tversky’s experiments, overweight salient recent information (e.g., current earnings trends) and underweight long-term base rates. This leads to forecasts that are too extreme when dramatic recent changes occur—a manifestation of representativeness and extrapolation heuristics. While agency incentives (e.g., incentives to generate optimistic forecasts to please management or attract trading) could also contribute to the average optimism captured in the negative intercepts, they do not easily account for the slope coefficients below 1, nor for the fact that overreaction appears in domains where such strategic incentives are minimal. Additional analyses reveal that forecast errors correlate with market valuation and earnings trend variables, but these factors explain only a small portion of the variation, leaving cognitive overreaction as a proposed explanation.

CONTRAST WITH UNDERREACTION

It is important to distinguish this long-horizon overreaction from a separate phenomenon observed in some studies: underreaction to recent earnings surprises at short horizons. Research by Klein (1990) and Abarbanell and Bernard (1992) finds that analysts’ forecasts are too conservative in updating immediately after an earnings announcement, leading to a delayed incorporation of information. This underreaction stands in contrast to the overreaction documented here for multi-year forecasts. Both biases can coexist: analysts may underweight the latest earnings surprise when forming immediate subsequent forecasts while simultaneously being overinfluenced by longer-run trends in their distant horizons. The direction of the error can therefore depend on the forecast horizon and the information being processed.