Conjectured procedure underlying GPT-4 factor generation

Determine whether OpenAI’s GPT-4 follows the six-step procedure for generating financial factors from futures market panel data schema metadata, specifically: (1) select relevant market features such as prices, volume, volatility, basis, and futures premium/discount; (2) compute derived statistics (e.g., moving averages, differences) and apply normalization (e.g., Z-scores); (3) construct composite indicators by combining individual signals via equal or custom weighting; (4) adapt calculations via rolling windows and adjustments for market specifics; (5) ensure analytical objectivity through statistical significance testing and validation against historical data; and (6) perform continuous evaluation and iteration via back-testing and adjustments. Establish whether GPT-4’s internal reasoning and code generation indeed implement these steps and formally characterize the mapping from input feature labels to algorithmic factor construction along this pipeline.

Background

The paper uses GPT-4 to generate 40 novel factors for the Chinese futures market by providing only data schema information (column labels) and asking the model to output Python code. To understand how GPT-4 produced these factors, the authors analyzed all generated code and inferred a structured, multi-step process they believe the model followed when constructing the factors.

The authors then asked GPT-4 to explain its factor-generation approach and judged the explanation to be consistent with their inferred steps. However, this remains a conjecture about the model’s internal reasoning, and verifying or characterizing the actual procedure would clarify how LLMs operationalize quantitative factor construction from schema-level inputs.