From Explainable Forecasts to Portfolio Weights: SHAP-Based Quantile Machine Learning for Dynamic S&P 500 Sector Allocation

This study investigates the integration of explainable artificial intelligence (XAI) and quantile-based machine learning techniques in dynamic portfolio management under changing market conditions. While machine learning models have demonstrated strong predictive capabilities in financial applications, their limited interpretability often constrains practical adoption in portfolio decision-making. To address this challenge, we develop an explainable portfolio allocation framework that combines gradient boosting prediction models (XGBoost and LightGBM) with SHAP (SHapley Additive exPlanations) values to generate transparent sector-level forecasts and dynamic asset allocation decisions. Using daily data for the S&P 500 and eleven GICS (Global Industry Classification Standard) sector indices from January 2017 to June 2025, two forecasting approaches are examined: conventional mean-based gradient boosting models and quantile gradient boosting models designed to capture distributional asymmetries and tail risks. Quantile-specific SHAP values are transformed into portfolio weights through a Sharpe-inspired, attribution-based allocation rule, and performance is evaluated against the S&P 500 benchmark and a tail-risk-oriented portfolio during two contrasting market episodes: the COVID-19 market crisis of 2020 and the first half of 2025, a period of lower systemic stress. The empirical results indicate that the quantile-based framework captures distributional information beyond point forecasts, although the raw tail quantiles are imperfectly calibrated out of sample. SHAP analyses reveal substantial shifts in sector importance across market regimes, with Information Technology, Communication Services, Financials, and Real Estate emerging as key drivers of portfolio decisions at different stages of the sample. The XAI-driven allocation process exhibits pronounced adaptability, dynamically adjusting sector exposures in response to evolving market conditions. Portfolio performance results indicate that the explainability-enhanced strategy is capable of generating economically meaningful allocation signals and, in the two examined mar-ket windows, records higher risk-adjusted ratios than the passive benchmark. The findings suggest that combining quantile machine learning models with explainable AI offers a promising framework for transparent, adaptive, and risk-aware portfolio management. The reported Sharpe-ratio advantage is statistically supported for the 2025 window, whereas the 2020 comparison does not reach conventional significance; the study does not isolate the incremental return contribution of SHAP from matched forecast-only allocations.