Distinguishing between recurring and nonrecurring components of earnings is a critical task in financial analysis and valuation. Academics and quantitative investors often rely on measures of recurring and nonrecurring components derived from standardized financial databases. We use unobserved components modeling and the Kalman smoother to obtain efficient ex-post estimates of the recurring and nonrecurring components of annual earnings. We then show that popular measures are significantly misspecified and that investors appear to anticipate a significant portion of the misspecification. Finally, we identify certain misclassified items that drive misspecification and provide algorithms to improve their ex-ante classification.