Using Machine Learning to Measure Conservatism

Abstract

Machine learning can improve empirical proxies of conservatism by detecting patterns beyond linear regression techniques assumed in prior literature. Using a neural network approach, we show that measures based on machine-learning exhibit (a) better fit adjusted for degrees of freedom, (b) fewer economically anomalous observations, (c) less unexplained year-over-year instability, (d) a secular decline in conservatism, and (e) more robust associations with periods post restatements. In simulations, proxies based on machine learning methods are the most robust to specification error and reduce the incidence of false negatives. Our approach shows that, separate from their usefulness in predictive analytics, methods from machine learning can be used to capture more informative variation than existing measures.

Keywords: machine learning, neural network, accounting, conservatism, measure, proxy

JEL Classification: C1, D2, M4