Reflexivity as Prompt: Does Awareness of Self-Reinforcing Market Dynamics Improve LLMs as Financial Market Forecasters?

Battle Count: 76

Highly relevant to quantitative trading. The paper directly addresses directional market forecasting during boom-bust cycles, evaluates risk-adjusted returns via Sharpe ratio of an implied long/cash strategy, and identifies practical constraints (context window length, episode type) that affect strategy performance. The finding that reflexivity awareness improves GFC forecasting (Sharpe up to +0.674) but struggles with prolonged manias (dot-com) mirrors real trading challenges. The accuracy-Sharpe divergence is a critical insight for strategy design. However, the paper is more exploratory/diagnostic than prescriptive for building a deployable trading system.

Low implementation complexity. The methodology is entirely zero-shot prompting with no fine-tuning, training, or custom model architecture. The main engineering tasks are: (1) data collection and normalization (straightforward), (2) constructing the four prompt variants (fully specified in Appendix C), (3) calling LLM APIs and parsing structured JSON output, (4) computing directional accuracy and Sharpe ratio. The primary challenge is the cost and latency of running 72 monthly forecasts × 4 conditions × 3 models × 2 context windows × 2 episodes = 3,456 API calls, plus ensuring reproducibility with proprietary models.

The paper provides full verbatim prompt text in Appendix C, detailed block composition in Table 2, and specifies exact data sources (Yahoo Finance for S&P 500 prices, Shiller dataset for EPS). However, it relies on proprietary frontier models (GPT-5, Claude Sonnet 4.6, Gemini 3 Pro) whose internal weights and exact API versions may change. The normalization procedure is fully specified. No code repository is mentioned. Statistical power is limited (n=72 per episode cell).