Predicting Forecasting Biases and Aggregate Outcomes using Neural Networks

Abstract

We develop a machine learning-based framework to predict the forecast biases of individual analysts based on their past forecast errors. Our results indicate that neural network (NN) models systematically outperform corresponding linear prediction models. Further, unlike linear models, a NN model can predict earnings announcement returns. Examining error predictability as an analyst-specific attribute, we demonstrate that analyst-level errors are persistent. Further, analysts with unpredictable errors have more informative forecasts, are more accurate, and incorporate business cycle news more effectively. Together, these findings suggest that neural network models can capture individual-level biases more accurately, which improves the quality of aggregate forecasts. 

Keywords: analyst biases, earnings forecasts, aggregate analyst sentiment, return predictability, machine learning, neural networks

JEL Classification: C45, G14, G24, G41