Inferring Latent Market Forces: Evaluating LLM Detection of Gamma Exposure Patterns via Obfuscation Testing

Battle Count: 52

Highly relevant to quantitative trading in several dimensions: (1) Validates LLMs as tools for detecting structural market constraints (dealer hedging flows) that directly impact price dynamics; (2) The 91.2% materialization rate of detected patterns provides actionable signal quality; (3) The separation of structural detection from profitability (stable detection vs. declining alpha) is crucial for understanding when patterns are tradeable; (4) Granger causality validation (p=0.042 at 2-day lag) confirms predictive utility; (5) The obfuscation methodology provides a framework for validating AI-generated trading signals against overfitting; (6) Directly applicable to options market making, volatility trading, and 0DTE strategy development. However, the paper focuses on validation methodology rather than direct trading strategy implementation, and the net alpha is modest (+5.6 bps full year).

Moderate complexity. The core pipeline involves: (1) GEX calculation from options data (standard but requires complete options chain data), (2) Data obfuscation (straightforward text transformation), (3) LLM API calls with structured prompts (requires API access and prompt engineering), (4) Forward return verification (standard financial data processing), (5) Statistical validation (binomial tests, Granger causality, bootstrap - all standard). The main challenges are: obtaining complete SPY options data with all strikes/expirations, designing effective obfuscation protocols, implementing the WHO→WHOM→WHAT structured output schema, and running comprehensive statistical validation. The GitHub repository provides code, reducing implementation burden. No model training or fine-tuning is required.

Code and validation pipeline are publicly available on GitHub (https://github.com/iAmGiG/gex-llm-patterns). The methodology is well-documented with clear obfuscation protocols, prompt templates, and validation criteria. However, the specific LLM used is not explicitly named in the extract, and the exact API configuration parameters are not fully detailed. The data source (SPY options data) is commercially available. Statistical tests (Granger causality, binomial tests, bootstrap) are standard and reproducible. The paper is a conference version with an extended journal version in preparation.