Behavioral Machine Learning? Computer Predictions of Corporate Earnings also Overreact

Abstract

Machine learning algorithms often predict more accurately than do humans. But, do they
satisfy rational expectations under standard tests? We show that the answer is, no. Like stock
analysts, machine predictions of corporate earnings overreact, albeit less strongly. Machine
overreaction can be reduced, but that also reduces average accuracy. Human stock analysts
with technical training overreact less than do other analysts. Human stock analyst predictions
contain information not otherwise available to the algorithms. A model is provided showing
the impact of increased analyst machine learning training on equity market equilibrium.

Keywords: stock analysts, machine learning, behavioral, overreaction

JEL Classification: G10, G20, G30