3S-Trader: A Multi-LLM Framework for Adaptive Stock Scoring, Strategy, and Selection in Portfolio Optimization

Battle Count: 183

Directly addresses portfolio construction and optimization using multi-modal financial data. Proposes a novel scoring-strategy-selection pipeline with iterative refinement. Achieves 131.83% accumulated return on DJIA with Sharpe ratio 0.31 and Calmar ratio 11.84. The training-free nature and interpretability of strategy adjustments make it highly relevant for practical quantitative trading. However, it is primarily a backtesting study without live validation.

The framework is training-free, eliminating the need for model training and hyperparameter tuning. However, it requires orchestrating six LLM-based agents with specific prompts, managing multi-modal data pipelines (news, fundamentals, technical indicators), implementing weekly rebalancing logic, and maintaining strategy trajectory history. The prompt engineering for each agent and the scoring dimension design require domain expertise. API costs for GPT-4o across multiple stocks weekly could be significant.

The framework uses GPT-4o API (commercial, version-dependent), data from Alpha Vantage (public API), and provides detailed prompt examples in figures. However, no code repository is mentioned, and LLM outputs can be non-deterministic. The experimental setup (weekly rebalancing, top-5 selection, 20% weights) is well-described. Evaluation period (May 2022 - May 2024) and training data (May 2012 - May 2022) are specified.