Beyond Polarity: Multi-Dimensional LLM Sentiment Signals for WTI Crude Oil Futures Return Prediction

Battle Count: 86

Directly addresses commodity futures return prediction using news sentiment, which is highly relevant for energy trading desks and commodity-focused quantitative strategies. The multi-dimensional sentiment framework (especially uncertainty and intensity dimensions) provides actionable signals beyond simple polarity. The weekly horizon is practical for medium-frequency trading strategies. SHAP analysis provides interpretable feature importance useful for signal validation. However, the binary classification approach and lack of transaction cost analysis limit direct implementation without further work.

Moderate complexity: requires API access to GPT-4o (costly), Llama 3.2-3b (open-source but needs GPU), FinBERT (open-source), and AlphaVantage (commercial API). Feature engineering pipeline (weekly aggregation, momentum, dispersion) is straightforward. LightGBM training with Optuna is well-supported. Main challenges: API costs for large-scale sentiment extraction, managing non-deterministic LLM outputs, and the 31-feature engineering pipeline. No end-to-end code provided.

The paper provides detailed methodology including structured prompts (Appendix A), full feature list (Appendix B), model specifications, and hyperparameter tuning procedure. However, no code repository is mentioned, the AlphaVantage API is commercial/paid, and the stratified random sampling seed is not specified. The use of proprietary GPT-4o API introduces non-determinism. Data sources (AlphaVantage, Yahoo Finance) are publicly accessible but require API keys.