Quant-Safe Explainable Artificial Intelligence for Dynamic Portfolio Management A Reproducible, Leakage-Resistant ML + SHAP Framework with Execution-Grade Accounting Logs

Machine learning (ML) backtests in finance frequently overstate performance due to data leakage, non-point-in-time features, and evaluation procedures that inadvertently incorporate future information. This paper proposes a leakage-resistant, reproducible, and deployment-oriented framework, the Quant-Safe architecture, combining (i) point-in-time feature engineering with explicit reporting lags, (ii) walk-forward evaluation with out-of-sample (OOS) explainability, and (iii) a robust portfolio translation layer with transaction- cost modeling and execution-grade accounting logs. We validate the framework on the Dow Jones Industrial Average (DJI) constituent universe over 2015-2025, using gradient-boosted trees and Shapley Additive Explanations (SHAP) to demonstrate that macro regime variables (e.g., interest-rate proxies) become dominant drivers during stress periods. The primary contribution is an engineering methodology enabling other researchers to reproduce, extend, and audit financial ML results with explicit controls against common failure modes.