From Economic Explanation to Performance Improvement in Machine-Learning Portfolios
In briefFrom Economic Explanation to Better Machine-Learning Portfolio Performance
SummaryThis paper proposes a local first-order attribution-propagation framework that links feature exposures in an allocation function to local sensitivities of portfolio performance metrics. The authors relate these attributions to Shapley values defined for portfolio objectives and derive statistical inference for Sharpe-ratio attributions. An attribution-guided exposure intervention can be applied after training without retraining the model. Empirically, the interventions yield statistically significant improvements in net Sharpe ratios across different model classes.