Forecasting Future Language: Context Design for Mention Markets

River summary

An earnings-call mention market resolves on whether a specified term appears verbatim in the official transcript. A YES contract priced at 60 cents implies a probability of roughly 60%. The task looks tailor-made for a language model, but a naive setup asks the model to forecast from scratch and throws away the strongest available signal: the market price itself.

The paper studies how to design the model's input context. Across 856 Kalshi mention markets covering 50 companies and 70 earnings events, forecasts were made seven days before each call using only information available at that cutoff: the previous quarter's transcript, up to 100 company-related news items, and the market-implied probability. All experiments used the same underlying model without fine-tuning; only the context and instruction changed.

The central contribution is Market-Conditioned Prompting (MCP): instead of showing the model the market price as one more fact, MCP instructs it to treat that price as a prior, evaluate the transcript and news as evidence, and return an updated probability. With identical information, this instruction alone improved every reported metric versus plain-context prompting.

MCP alone still did not beat the market in every regime, so the paper's final method, MixMCP, shrinks the model's update back toward the market with a 70/30 weighting. That conservative mixture produced the best overall result, edging out the market baseline itself.

Key results

MCP helped most where the market was uncertain: on disagreements it beat the market in 56.7% of cases when the market probability was between 50% and 60%, and in 62.5% of cases between 60% and 70%. At high market confidence, the market was already difficult to improve.

Authors

Related River articles