Shapley in Context: Explaining Financial Language with Domain Expertise

Battle Count: 64

The paper is primarily focused on explainability of LLMs in financial text applications rather than direct trading strategy development. However, it has significant relevance to quantitative trading through: (1) risk attribution from 10-K filings which can inform portfolio risk management, (2) sentiment analysis for signal generation, (3) model risk management compliance requirements, and (4) understanding model behavior for algorithmic trading applications. The Form 10-K risk factor analysis is directly applicable to fundamental analysis and risk-based trading strategies.

The theoretical framework is mathematically rigorous but the practical implementation is moderate. Key complexity factors: (1) Shapley value computation requires 2^n evaluations of the characteristic function, though feasible for ~5-7 risk headings; (2) Monte Carlo estimation with 30 runs adds computational overhead; (3) API-based interaction with ChatGPT-5.2 requires careful prompt engineering; (4) Handling non-deterministic LLM outputs and ensuring consistency; (5) Data preprocessing for Form 10-K filings (merging headings, removing identical risk factors). The paper notes experiments run on standard CPU hardware without GPUs.

The paper provides detailed prompts for ChatGPT API calls, specifies the experimental setup (Google Colab, 2 vCPUs, 12 GB RAM, Python 3.10), uses publicly available Form 10-K filings from SEC, and reports Monte Carlo results with confidence intervals. All data are publicly available. However, reliance on ChatGPT-5.2 API introduces non-determinism, and no code repository is explicitly mentioned.