Out-of-Sample Predictability of Firm-Specific Stock Price Crashes: A Machine Learning Approach

Abstract

We use machine learning methods to predict firm-specific stock price crashes and evaluate the out-of-sample prediction performance of various methods compared to traditional regression approaches. Using financial and textual da-ta from 10-K filings, our results show that a logistic regression with financial data inputs performs reasonably well and sometimes outperforms newer classifiers such as random forests and neural networks. However, we find that a stochastic gradient boosting model systematically outperforms the logistic regression, and forecasts using suitable combinations of financial and textual data inputs yield significantly higher prediction performance. Overall, the evidence suggests that machine learning methods can help predict stock price crashes.

Keywords: Stock Price Crash Risk; Textual Disclosures; Machine Learning; Natural Language Processing

JEL Classification: G11, G14, G32, M41