In an investigation of an asset pricing anomaly, “predictability bias,” stocks were classified into quintiles based on how predictable the company's earnings have been in the past, as measured by analysts' past forecast errors. “Current” forecasts of earnings are excessively optimistic for companies whose earnings were hard to predict in the past; that is, the least predictable companies have much larger, positive forecast errors (forecast > actual) relative to the most predictable companies. Abnormal returns are consistent with the current forecast errors; that is, stocks of the least predictable companies substantially underperform stocks of the most predictable companies. Adjustments for systematic risk, firm size, book-to-price ratio, and industry factors do not eliminate the differential returns between least predictable and most predictable companies.