Predicting M&A Announcement Returns with Machine Learning: Firm Size, Model Complexity, and the Role of Time-Consistent Validation

This study examines whether machine learning methods can improve the prediction of abnormal stock returns surrounding merger and acquisition (M&A) announcements. Six machine learning algorithms, including Support Vector Machines, Random Forests, Gradient Boosting, Logistic Regression, K-Nearest Neighbors, and Multi-Layer Perceptrons, were evaluated using both time-ordered and random sampling validation frameworks. The results of the current study indicate that greater model complexity does not consistently improve predictive performance. Although several advanced models demonstrated moderate predictive capability, the models did not reliably outperform simpler and more easily interpreted methods. In the study, predictive accuracy also varied substantially by firm size, with smaller firms exhibiting significantly higher predictability than larger firms. In addition, random sampling produced significantly higher performance estimates than time-consistent validation, suggesting that improper evaluation procedures may overstate true out-of-sample predictive performance. Overall, the findings of the study emphasize the importance of evaluation design and suggest that predictive gains in M&A forecasting depend more on data characteristics and firm heterogeneity than on model complexity.