We develop and compare alternative empirical models for the detection of insider abuse and fraud occurring at U.S. commercial banks with the goals of identifying leading indicators of fraud and fraud prediction. Specifically, we use information on enforcement actions taken by bank supervisors to remove bank officers and other employees and to prohibit bank employees from banking during the period 2007–2019 to identify instances of serious insider abuse and fraud. To the best of our knowledge, these data — enforcement actions — have not been used in previous studies of insider abuse and fraud (hereafter, fraud) at banks. The explanatory variables in our models are measures of banks’ capital adequacy, asset quality, management, earnings, liquidity and sensitivity to market risk and other financial measures that previous studies have identified as indicators of fraud. We model fraud using machine learning approaches, each with different strengths and weaknesses — artificial neural network, deep learning autoencoder network, gradient boosting, logit regression and random forest. The models are fitted annually using observed incidences of fraud and four quarters of bank financial data. We use these models to make predictions of the probability of fraud at banks in the following calendar year. Our results indicate that artificial neural network, gradient boosting and random forest are, in general, more accurate in out-of-sample/time fraud forecasts than are deep learning autoencoder network and logit regression models. Finally, in terms of the drivers of fraud the estimates of logit regressions are somewhat mixed but do suggest that earning assets and capitalization are negatively related to fraud while loan loss rates and noncore funding dependence are positively related to fraud.