Eliciting ESG Preferences for Reinforcement Learning-Based Portfolio Optimization

Battle Count: 67

Highly relevant to quantitative trading as it directly addresses portfolio optimization with multi-objective constraints (financial returns + ESG). The MORL framework with preference-conditioned PPO is directly applicable to algorithmic trading systems. The Sharpe Ratio optimization via differential Sharpe Ratio is a core quantitative trading technique. However, the synthetic ESG data and simplified market assumptions (frictionless trading, 10 stocks) limit immediate practical deployment. The LLM persona evaluation is more of a validation methodology than a trading tool.

High complexity due to: (1) Multi-objective RL with preference-conditioned PPO and FiLM architecture, (2) Gaussian Process preference elicitation with qEUBO acquisition function via BoTorch, (3) LLM persona prompting and evaluation pipeline, (4) Synthetic ESG data generation calibrated to real-world divergence patterns, (5) Multiple training phases (policy training, elicitation, evaluation) with temporal splits. Requires expertise in RL, Bayesian optimization, LLM prompting, and quantitative finance.

The paper provides detailed hyperparameters (Table 1), clear methodology description, and specifies the asset universe (10 DJIA stocks). However, ESG ratings are synthetically generated (not real data), no GitHub repository is mentioned, and the exact LLM prompts are provided but the full code implementation is not referenced. The synthetic data generation process is described but not fully reproducible without code.