Forecasting Earnings Using k-Nearest Neighbor Matching

Abstract

We use the k-nearest neighbors (i.e., k-NN) algorithm to forecast a firm’s annual earnings by matching its recent trend in annual earnings to historical earnings sequences of “neighbor” firms. Our forecasts are more accurate than forecasts obtained from the random walk, the regression model developed by Hou, van Dijk and Zhang (2012), other regression models and the matching approach described in Blouin, Core and Guay (2010). The k-NN model is superior to these alternative models both when analysts’ forecasts are available and when they are not. Further, for firm-years with I/B/E/S earnings data available, the accuracy of k-NN forecasts of I/B/E/S earnings is similar to the accuracy of analysts’ forecasts. The k-NN model is also superior to a random forest classifier that we use to choose the best model ex-ante. Finally, we find that our forecasts of earnings changes have a positive association with future stock returns.

Keywords: earnings, forecasting, machine learning

JEL Classification: C21, C53, G17, M41