An intriguing anomaly in recent market-based accounting research is that both the market and analysts appear not to recognize properly the time-series properties of quarterly earnings shocks. Bernard and Thomas (1990) and Freeman and Tse (1989) present evidence that the market underestimates the implications of previous period earnings for future earnings. Mendenhall (1991) and Abarbanell and Bernard (1991) find that analysts do not seem to utilize time-series information about earnings correctly when setting their forecasts. Specifically, these last two studies document a positive serial correlation in analysts' quarterly forecast errors and interpret this finding as analysts systematically underestimating the persistence of past earnings forecast errors in forecasting future earnings. The purpose of this article is to examine whether analysts properly recognize the time-series properties of annual earnings when setting their estimates of future earnings. Givoly (1985) investigates this issue and finds that analysts' forecasts of annual earnings per share (EPS) are unbiased and that prediction errors are not serially correlated. He concludes that forecasts are formed in an efficient manner. In contrast, this study finds that, on average, analysts set overly optimistic estimates of the next period's annual EPS and that forecast errors display significantly positive serial correlation. These results hold for short-term as well as for longer-term IBES consensus forecasts. Bias and positive serial correlation in forecast errors suggest that analysts do not properly recognize the time-series properties of earnings when setting expectations of future earnings. Studies show that, for any given year, earnings shocks have both permanent and transitory (i.e., mean-reverting) components (see, e.g., Brooks and Buckmaster 1976; Ou 1990; Ou and Penman 1989b) and that the level of earnings persistence varies across firm-years. We examine whether analysts are aware of the differences between permanent and temporary components in the previous year's earnings when predicting future earnings. The findings show that analysts are able to differentiate partially between permanent and temporary components in previous period earnings. We also find that the overestimation bias in forecasts is most pronounced for firms that recently experienced negative earnings. In contrast, the positive serial correlation is most evident for firms that had predominantly permanent earnings. These results suggest that the overestimation bias and serial correlation are not uniform across firms. To address the economic significance of our findings, estimates of bias and serial correlation in forecast errors are used to adjust existing forecasts. The accuracy of the adjusted forecasts are then compared to the unadjusted forecasts. A significant improvement in forecasting ability is found for the adjusted longer-term forecasts, as evidenced by a 12 percent improvement in the mean squared error. Further, the bias and serial correlation in the adjusted forecasts are significantly less than in the actual forecasts for both longer and short terms. These results further support the view that analysts do not utilize available information efficiently when setting forecasts.
