From Economic Explanation to Performance Improvement in Machine-Learning Portfolios

01 ABSTRACT

The authors propose a local first-order attribution-propagation framework that maps feature exposures in the allocation function to local sensitivities of sequential portfolio performance metrics, relating the propagated attributions to Shapley values defined directly for portfolio objectives. They also derive statistical inference for Sharpe-ratio attributions and design attribution-guided feature-exposure interventions that enable post-training adaptation without retraining the underlying allocation model. The authors report that such interventions yield statistically significant improvements in net Sharpe ratios across different model classes. These improvements are the authors' empirical findings, and the framework's economic explanatory value and intervention effects still require independent verification.

02 KEY FINDINGS

1. A local first-order attribution-propagation framework is proposed to map allocation-function feature exposures to local sensitivities of sequential portfolio performance metrics and relate them to Shapley values defined directly for portfolio objectives.
2. Statistical inference for Sharpe-ratio attributions is derived, and attribution-guided feature-exposure interventions are designed to translate portfolio-metric explanations into out-of-sample decisions without retraining the allocation model.
3. Temporal stabilisation and cross-model averaging are introduced as complementary regularisation mechanisms, and statistically significant improvements in net Sharpe ratios are reported across different model classes.