Abstract. This research re-examines whether there are differences in the forecast accuracy of financial analysts through a comparison of their annual earnings per share forecasts. The comparison of analyst forecast accuracy is made on both an ex post (within sample) and an ex ante (out of sample) basis. Early examinations of this issue by Richards (1976), Brown and Rozeff (1980), O'Brien (1987), Coggin and Hunter (1989), O'Brien (1990), and Butler and Lang (1991) were ex post and suggest the absence of analysts who can provide relatively more accurate forecasts over multiple years.
Contrary to the results of prior research and consistent with the belief in the popular press, we document that differences do exist in financial analysts' ex post forecast accuracy. We show that the previous studies failed to find differences in forecast accuracy due to inadequate (or no) control for differences in the recency of forecasts issued by the analysts. It has been well documented in the literature that forecast recency is positively related to forecast accuracy (Crichfield, Dyckman, and Lakonishok 1978; O'Brien 1988; Brown 1991). Thus, failure to control for forecast recency may reduce the power of tests, making it difficult to reject the null hypothesis of no differences in forecast accuracy even if they do exist.
In our analysis, we control for the differences in recency of analysts' forecasts using two different approaches. First, we use an estimated generalized least squares estimation procedure that captures the recency-induced effects in the residuals of the model. Second, we use a matched-pair design whereby we measure the relative forecast accuracy of an analyst by comparing his/her forecast error to the forecast error of another randomly selected analyst making forecasts for the same firm in the same year on or around the same date. Using both approaches, we find that differential forecast accuracy does exist amongst analysts, especially in samples with minimum forecast horizons of five and 60 trading days. We show that these differences are not attributable to differences in the forecast issuance frequency of the financial analysts. In sum, after controlling for firm, year, forecast recency, and forecast issuance frequency of individual analysts, the analyst effect persists.
To validate our findings, we examine whether the differences in the forecast accuracy of financial analysts persist in holdout periods. Analysts were assigned a “superior” (“inferior”) status for a firm-year in the estimation sample using percentile rankings on the distribution of absolute forecast errors for that firm-year. We use estimation samples of one- to four-year duration, and consider two different definitions of analyst forecast superiority. Analysts were classified as firm-specific “superior” if they maintained a “superior” status in every year of the estimation sample. Furthermore, they were classified as industry-specific “superior” if they were deemed firm-specific “superior” with respect to at least two firms and firm-specific “inferior” with respect to no firm in that industry. Using either definition, we find that analysts classified as “superior” in estimation samples generally remain superior in holdout periods. In contrast, we find that analysts identified as “inferior” in estimation samples do not remain inferior in holdout periods.
Our results suggest that some analysts' earnings forecasts should be weighted higher than others when formulating composite earnings expectations. This suggestion is predicated on the assumption that capital markets distinguish between analysts who are ex ante superior, and that they utilize this information when formulating stock prices. Our study provides an ex ante framework for identifying those analysts who appear to be superior. When constructing weighted forecasts, a one-year estimation period should be used because we obtain the strongest results of persistence in this case.