Beyond ESG Scores: Learning Dynamic Constraints for Sequential Portfolio Optimization

Battle Count: 64

Highly relevant to quantitative trading as it addresses ESG-aware portfolio optimization using deep reinforcement learning with constrained optimization. The MACF-X framework is directly applicable to portfolio construction, risk management, and sustainable investing strategies. The paper provides concrete implementations for PPO, CRPO, TRPO, and CPO backbones with detailed update rules. However, it focuses on daily rebalancing of large-cap equities rather than high-frequency trading, and the ESG constraint integration is the primary contribution rather than pure alpha generation.

The framework involves multiple components: (1) a point-in-time multimodal ESG data pipeline with LLM-assisted semantic labeling, (2) a three-head MLP constraint field (MACF) with weak supervision, (3) portfolio-level cost aggregation with softplus gating, (4) budget calibration from warm-start rollouts, and (5) optimizer-specific adapters (MACF-PPO, MACF-CRPO, MACF-TRPO, MACF-CPO) with shared pressure layer. The data pipeline is particularly complex, requiring news deduplication, ticker-date alignment, semantic labeling, and multi-level aggregation. However, the MACF-X adapters are modular and attach to existing DRL backbones without architectural changes.

Code, processed data, and run configurations are provided at an anonymized repository (https://anonymous.4open.science/r/SHUFHOSPNCUS). Detailed appendices cover data construction pipeline, optimizer update rules, universe construction, reward protocol, and MACF training. However, the paper is a preprint and some data sources (Yahoo Finance, Google News, GDELT, SEC, EEOC) require external access. The weak-supervision labeling functions use fixed coefficients described in the repository.