Forecasting Future Language: Context Design for Mention Markets

Battle Count: 70

The paper is highly relevant to prediction market trading and event-driven strategies. It demonstrates that LLMs can add value to market-implied probabilities, particularly in mid-confidence regimes where markets are uncertain. The MixMCP approach provides a principled framework for combining textual signals with market prices. However, the application is narrow (keyword mention markets) and the improvement over market baseline is modest. The methodology could inform broader NLP-based alpha generation strategies in prediction markets and event-driven trading.

The core methodology is relatively straightforward: construct prompts with market probability and textual context, query an LLM, and apply a convex mixture. The main complexity lies in data pipeline construction (Kalshi API access, SERP API for news retrieval, transcript collection), prompt engineering, and ensuring no information leakage. No model training or fine-tuning is required. The approach is essentially a well-designed prompting strategy with a simple post-processing step.

The paper uses GPT-5.1 (a proprietary API model), Kalshi market data (commercial platform), and SERP API for news retrieval. While the methodology is clearly described with prompt templates in the appendix, the specific market data and API access may limit full reproducibility. The dataset of 856 Kalshi earnings-call mention markets spanning 50 companies is not publicly released. The alpha parameter selection procedure is described but the held-out split details are limited.