You Have a Point – But a Point Is Not Enough: The Case for Distributional Forecasts of Earnings

Abstract

Existing forecasts of earnings are typically expressed as point estimates. The future earnings number is ex-ante uncertain, however, and is statistically represented by a probability distribution over all possible earnings outcomes. Thus, a proper forecast of earnings is by nature distributional. We use recent advances in statistical machine learning to produce full distributional forecasts of earnings right before earnings announcements, and investigate the utility of such forecasts in three directions. First, we show that our distributional forecasts are well-calibrated to actual earnings realizations. Second, we use our distributional forecasts to model the probability of beating/missing the consensus analyst forecasts. Going long (short) on stocks in the extreme decile probabilities of beating (missing) the consensus produces hedge returns of about 60 basis points over the three-day earnings announcement window. We also show that using distributional forecasts dominates mean-based approaches in earning abnormal returns. Third, we document that ranges in management and financial analyst forecasts are too narrow, severely underestimating the variability of future earnings. Crucially, since our distributional forecasts are available in real time at the firm-quarter level, they can be proactively used to identify and correct such miscalibration.

Keywords: Earnings distribution, Statistical machine learning, Analyst forecasts, Stock returns

JEL Classification: G14, G15, M41