Are machines better predictors of insider trading?

This study examines whether machine learning (ML) techniques can improve the prediction of insider trading behaviour compared with traditional linear approaches. Using a comprehensive sample of U.S. insider trading data from 2000 to 2022, we compare ensemble learning methods—random forest and extreme gradient boosting—to logistic regression and least absolute shrinkage and selection operator (LASSO). We implement Bayesian hyperparameter optimisation to improve model tuning and employ Shapley additive explanations (SHAP) values to maintain interpretability and identify the principal economic determinants of insider trading decisions. Additionally, we apply Gaussian Thompson sampling to evaluate competing hypotheses about insiders' market-timing motives. The results demonstrate that ML methods substantially outperform linear models in predicting both the likelihood and magnitude of insider sales, with predictive gains particularly pronounced among female insiders. SHAP analysis indicates that incentive structures play a stronger role for male insiders, whereas female trading behaviour appears more closely associated with private information about future firm performance.