ChatGPT as a Time Capsule: The Limits of Price Discovery

Battle Count: 142

Directly relevant to quantitative trading: demonstrates a novel alpha signal (LLM outlook score) with statistically robust cross-sectional return predictability (t=4.25 single-model, t=6.02 pooled panel) after controlling for standard factors and market-implied valuations. Provides portfolio construction application with mean-variance optimization achieving Sharpe ratio of 2.31 (gross of costs) and maximum drawdown of -3.7% over 30 months. The signal is implementable via API calls to frozen model checkpoints. Key caveats: in-sample results, substantial overlap with profitability factor, only 6 unique time periods, and the anomaly may attenuate with adoption. The narrative-congestion mechanism suggests the signal may persist as disclosure complexity grows.

Moderate complexity: requires OpenAI API access to specific model versions (some may be deprecated), standard financial data feeds (Refinitiv, Compustat, I/B/E/S), Axioma risk model for portfolio optimization, and standard econometric software for Fama-MacBeth regressions and panel analysis. The prompt engineering is straightforward (single fixed template). Main challenges: maintaining access to frozen model checkpoints, handling API rate limits for ~7,000 firms per checkpoint, and constructing market-implied valuation metrics (ICE-GLS requires numerical optimization). No code repository provided.

High reproducibility: deterministic API calls (temperature=0, top_p=1), fixed prompt template provided in Appendix A, standard data sources (Refinitiv, Compustat, I/B/E/S, Kenneth French data library), sector-neutral z-score normalization clearly defined. Missingness rate below 1%. However, API access to specific model versions may become unavailable over time, and the OpenAI Researcher Access Program dependency limits independent replication.