In this paper I estimate the risk of CEO dismissal using a variety of machine learning algorithms.I show that linear regression tree methods significantly outperform the logit and
linear models used in prior literature, as well as other algorithms, notably neural networks,
who perform surprisingly poorly in this setting. Taking these superior predictions to applications
from prior studies, I find that relationships change. Peters and Wagner (2014) found
that increases in forced turnover risk were related to a material increase in CEO pay, the more
accurate risk estimate remains statistically significant, but becomes less than a percent of its
previous size. As well, decreases in pay-performance-sensitivity found in Bushman, Dai, and
Wang (2010) are no longer economically significant. Furthermore, using the likelihood of dismissal
to address sample selection bias in Huson, Malatesta, and Parrino (2004), I find even
stronger evidence of a positive link between CEO firing and future firm-level performance.