"Beating a Random Walk? Earnings Forecasts and the Delisting Bias"
Existing cross-sectional earnings forecast models are based on samples of Compustat firms that survive for at least two years. But a substantial portion of observations disappear from the Compustat data each year. This portion is greater for firms that report losses in their last year in the Compustat data. We simulate a cross-section and time-series of earnings data using a random walk model and we remove the lowest (most negative) earnings in each year. We examine the performance of earnings forecasting models within these simulated data. We show that an earnings forecasting model that allows the autocorrelation of earnings to be different for loss firms than for positive earnings firms provides forecasts that outperform random walk forecasts even though the underlying data are generated via the random walk model. That is, the removal of the extreme loss observations (simulating what happens when firms disappear from Compustat) is the reason why the forecasting model out-performs a random walk model. We then turn our analysis to real data. We provide evidence that echoes the results of the simulation - extant models beat a random walk because of the effects of disappearance from the Compustat data set. Finally, we offer suggestions for future researchers interested in evaluating the efficacy of their earnings forecast models.