Interpretable ESG–sentiment hybrid deep learning for asset return forecasting with quantified interactions and latency-aware deployment

Accurate forecasting of financial time series increasingly relies on alternative data such as environmental, social and governance (ESG) scores and news-based sentiment, yet the way these signals interact and when they actually improve forecasts is still poorly understood. We introduce an interpretable hybrid framework for asset return forecasting that combines a Temporal Fusion Transformer (TFT) with a lightweight Support Vector Regression (SVR) residual corrector and an explicit gated late fusion of ESG features with aspect-based financial sentiment (FinBERT-based ABSA). The gating mechanism learns when to emphasize sustainability versus sentiment signals, while SHAP interaction values and Friedman’s H quantify ESG–sentiment interactions across assets and regimes. A finance-grade, leak-proof walk-forward protocol (252 trading days train / 10 days test, within-fold scaling, ABSA items strictly before 16:00 ET; ESG effective T+3; macro T+1, HAC-robust Diebold–Mariano tests) is applied to US large-cap technology equities, major global indices, and BTC/ETH over 2020–2024. Across $$n=5$$ independent seeds, the hybrid achieves aggregate mean absolute error of 2.77× 10⁻³ and RMSE of 5.18× 10⁻³ on next-day log returns, with directional accuracy $$94.5%$$ , IC 0.39, and ICIR 0.82, significantly outperforming tuned deep-learning and machine-learning baselines (HAC-robust per-asset Diebold–Mariano tests with BH-FDR $$q=0.05$$ ; Fisher aggregation yields $$p<0.01$$ ). Simple long-only, thresholded simulations indicate higher risk-adjusted performance and lower maximum drawdown under conservative transaction-cost assumptions. Ablation studies show that removing either ESG or sentiment features yields the largest degradations, and that the SVR corrector stabilizes errors under regime shifts. To directly address market-cycle sensitivity, we evaluate stability across event-defined stress windows (COVID-19 crash, 2022 tightening cycle, and 2023 banking stress) and volatility-defined regimes using terciles of 20-day realized volatility. We report regime-split forecasting and strategy metrics with block-bootstrap confidence intervals, HAC-robust Diebold–Mariano tests within each regime, and residual-stabilization diagnostics that quantify the SVR variance and skewness reduction under stress. ESG–sentiment interactions are statistically non-zero and regime-dependent, with sentiment gaining importance in turbulent periods and ESG in calmer markets. A latency-optimized variant that removes auxiliary BiLSTMs retains over $$90%$$ of the accuracy gains while reducing inference time by approximately $$55%$$ of the full model (i.e., a reduction of about $$45%$$ ), supporting near-real-time deployment.