A Hybrid Architecture for Options Wheel Strategy Decisions: LLM-Generated Bayesian Networks for Transparent Trading

Battle Count: 68

Highly relevant to quantitative trading. The paper directly addresses algorithmic options trading (wheel strategy) with a novel hybrid architecture combining LLMs and Bayesian Networks. It provides comprehensive empirical validation over 18.75 years with 8,919 trades, strict out-of-sample testing, transaction cost modeling, and multiple baseline comparisons. The focus on risk-adjusted performance (Sharpe 1.08 vs 0.62), drawdown management (-8.2% vs -60%), and full decision transparency (27 factors per trade) makes it directly applicable to production trading systems. The model-first paradigm addresses critical concerns about LLM reliability in quantitative domains.

Moderately high complexity. Requires integration of LLM API calls (GPT-4), Bayesian network construction and inference (pgmpy library), historical data management (8,919 trades with 27 factors), DAG validation, conditional probability table population, feedback loop implementation, and real-time market data processing. The multi-stage pipeline (context analysis → network construction → data selection → probability population → inference → decision → feedback) adds architectural complexity. However, the paper provides detailed code listings for all components, reducing implementation barriers. Requires expertise in both NLP/LLM engineering and probabilistic graphical models.

The paper provides detailed implementation code listings (LLM configuration, prompt templates, parsing pipeline, DAG validation, integration pipeline, error handling), model specifications (GPT-4-0613 with temperature 0.1, top_p 0.9), and complete temporal validation protocol. However, no GitHub repository is provided, the historical dataset (8,919 trades) is not publicly available, and the specific LLM prompts and fine-tuning details for optimal performance are partially described. The data source (Yahoo Finance API) is specified but the exact preprocessing pipeline is not fully detailed.