Regime-Switching ESG Portfolio Optimization: Integrating Temporal Fusion Transformers and Reinforcement Learning for Enhanced Risk-Adjusted Returns
Year: 2026
Type: article
Abstract: Integrating ESG criteria into institutional portfolios creates tension with financial objectives under changing market conditions. This paper proposes an architecture combining Temporal Fusion Transformer (TFT)–XGBoost forecasts, Gaussian Hidden Markov Model (HMM) inference, Proximal Policy Optimization (PPO), and constrained portfolio refinement. An executed annual chronological experiment separately evaluates a tractable allocation implementation using Ridge forecasts and an observable prior-return risk indicator. The annual experiment uses 4,837 source firm records, a lagged-data candidate rule, and eight independently solved allocation strategies. Over the retrospective 2023–2024 evaluation, annual adaptive ESG allocation earns 19.00% cumulative net return, compared with 54.29% for constrained equal weight and 62.91% for constrained capitalization weighting. Its mean lagged ESG score is 74.58. On 100 firm-year forecasts across two market dates, Ridge RMSE is 51.25 percentage points, compared with 42.21 for historical means. All 240 allocations across six scenarios satisfy the specified annual constraints within 10−7. These findings establish conditional pipeline feasibility rather than a fully validated TFT–HMM–PPO architectural outperformance or a statistically robust algorithmic edge. ESG Beta remains an exploratory co-movement metric within the framework. Unverified data vintages and low-frequency annual constraints highlight practical deployment bottlenecks for automated financial systems.
Language: English
Field: Decision Sciences
Domain: Social Sciences
SDG: Climate action
Open Access status: diamond
OpenAlex ID: W7224993136