Predicting the Accuracy of Earnings Forecasts Using Machine Learning

Abstract

Previous studies present models that predict the accuracy of earnings forecasts and provide evidence consistent with investors using these models when evaluating forecasts. In this study I mirror the framework of past studies but take a machine learning approach to predicting forecast accuracy. I evaluate the out-of-sample performance of random forest, neural network and k-nearest neighbors compared to linear regression. I find random forest and neural network have modestly stronger out-of-sample predictive performance than linear regression for annual and quarterly earnings periods. Then, I test whether the accuracy predictions from these models help explain investor reactions to forecasts. I find the predictions from each model help explain investor reactions to forecasts – consistent with the intuition that investors use these models when predicting forecast accuracy.