Study: Interpretable Models Can Predict Human Randomisation Failures in Games

A recent submission on arXiv (2608.07220) explores if transparent machine learning models can achieve the same predictive accuracy as black-box models when predicting human deviations from mixed strategy equilibrium in game theory. Analyzing 84,060 decisions made by 2,802 pairs, the research centers on O'Neill's zero-sum card game, where players often stray from the i.i.d. standard. The authors compare naive and behavioral models with interpretable and deep learning models, assessing modified EWA specifications from previous studies. They introduce a nested frequency tracking extension using LASSO diagnostics. Findings reveal that players' tendencies to repeat or avoid actions, particularly in managing their recent decisions, largely contribute to the interpretable and strategically exploitable predictive capabilities. The abstract is accessible at the provided URL.