Regret-Driven Portfolios: LLM-Guided Smart Clustering for Optimal Allocation

Battle Count: 89

Highly relevant to quantitative trading and portfolio management. The paper directly addresses portfolio allocation, risk-adjusted return optimization, and drawdown management—core concerns in quantitative finance. The no-regret learning framework provides a theoretically grounded approach to adaptive allocation. The LLM-guided hedging and sentiment filtering add practical dimensions for systematic strategy construction. The quarterly rebalancing frequency makes it suitable for semi-passive institutional strategies. However, it is not designed for high-frequency trading, and the LLM components introduce some non-determinism. The empirical results (69% higher annualized returns, 119% higher Sharpe ratio vs SPY) are compelling but based on backtesting only.

Moderate to high complexity. Requires: (1) custom backtesting framework with realistic constraints (delisting handling, bond pricing from yield curves, transaction costs), (2) integration of LLM API calls for sentiment interpretation and hedging recommendations, (3) implementation of the no-regret allocation algorithm with action set generation via asset perturbation, (4) sentiment filtering logic using Fear & Greed Index, (5) dynamic clustering with multi-prompt voting logic, (6) handling of non-uniform asset listing durations. The core algorithm is relatively straightforward (greedy FTL), but the full pipeline with LLM integration, multiple modules, and parameter tuning adds complexity. No specific LLM model is named, requiring users to select and configure their own.

Source code is available at an anonymous link (anonymous.4open.science/r/no-regret-paper-62CA). Data sources are specified (Yahoo Finance, investing.com, World Gold Council, GitHub for Fear & Greed Index). Key hyperparameters and design choices are described in detail. However, the LLM used is not explicitly named, and the anonymous code link may limit long-term reproducibility. The backtesting framework is custom-built. Sensitivity analyses are provided for rebalancing periods, look-back windows, and sentiment filters.