Large Language Models and Futures Price Factors in China

Battle Count: 215

Highly relevant to quantitative trading. The paper directly addresses factor construction for futures markets, portfolio construction (long-short and long-only), risk-adjusted performance evaluation, and alpha generation. It provides a practical framework for using LLMs to generate trading signals in data-scarce emerging markets. The IPCA benchmark comparison and comprehensive backtesting methodology are directly applicable to quantitative trading strategy development. The focus on Chinese futures (104 contracts including stock index and commodity futures) is particularly relevant for practitioners in Asian markets.

Moderate complexity. The core methodology (prompting GPT to generate factor code) is conceptually simple, but practical implementation requires: (1) careful prompt engineering to avoid forward-looking bias, (2) manual code review and validation of GPT outputs, (3) handling technical issues in generated code, (4) constructing proper backtesting frameworks with in-sample/out-of-sample splits, (5) implementing IPCA benchmark, (6) managing 104 futures contracts with varying data availability, and (7) computing multiple performance metrics. The zero-shot approach eliminates the need for model training but requires significant human oversight of generated code.

The paper provides detailed prompts, factor descriptions (40 factors in Appendix B), data specifications (104 futures contracts from Wind database, 2010-2023), and methodology steps. However, GPT outputs are inherently non-deterministic, and the exact prompts and code for all 40 factors are not fully provided. The IPCA benchmark implementation details are referenced but not fully specified. Transaction fee assumptions (0.025%) are stated. The paper acknowledges variability from temperature parameter settings.