River summary

A price move in one prediction market can carry information for another: higher inflation odds may reprice a rate-decision contract, and a recession signal in one country may shift growth expectations elsewhere. Statistical tests such as Granger causality can find these lead-lag relationships, but large pairwise searches inevitably surface correlations that are significant by chance or too fragile to survive a regime change.

This paper exploits a property unique to prediction markets: every price series is attached to a natural-language event description. That lets a language model inspect a proposed relationship in economic terms and ask whether there is a coherent reason the follower should reprice when the leader moves.

The framework separates discovery from judgment. Stage one applies Granger-causality tests to daily prices, converted to log odds, and ranks candidate leader-follower pairs by statistical strength. Stage two gives the event titles and descriptions to an LLM, which scores the plausibility of the transmission mechanism and assigns the expected sign. The LLM never searches for new pairs; it re-ranks candidates that already passed the quantitative screen, and both portfolios trade under identical mechanical rules.

Tested on 554 Kalshi Economics markets across 18 out-of-sample periods, the hybrid portfolio improved the win rate from 51.4% to 54.5% and cut the average losing trade by 46.5%. The gains were largest on sharp leader moves, exactly where fragile correlations become expensive.

Key results

54.5%
Hybrid win rate, up from 51.4% for the statistical screen
46.5%
Reduction in the average losing trade, from $649 to $347
+205%
Total experimental P&L, from $4,100 to $12,500
Statistical screen vs. hybrid semantic screen, default configuration
Metric Statistical screen Hybrid semantic screen Change
Win rate 51.4% 54.5% +3.1 pts
Average winning trade $724 $636 -12%
Average losing trade -$649 -$347 -46.5%
Total experimental P&L $4,100 $12,500 +205%

Loss reduction persisted across every tested holding period from one to 21 days, with average loss magnitude falling by 22.6% to 47.2% depending on the horizon.

Authors

Sumin Kim Minjae Kim Jihoon Kwon Yoon Kim Nicole Kagan Joo Won Lee Oscar Levy · River Markets Alejandro Lopez-Lira Yongjae Lee Chanyeol Choi

A collaboration across LinqAlpha, MIT, Kalshi, Arrowpoint Investment Partners, UC Berkeley, the University of Florida, UNIST, and River Markets.

Related River articles

From the Insights blog

Semantic correlations and how to trade them

A readable walkthrough of this paper: why statistically significant trades fail, how the two-stage screen works, and why the LLM is best described as a risk manager.

From the Insights blog

The hidden 2.4% tax on your prediction market trades

A live order-book case study of what execution really costs across venues: depth, fees, routing, and the all-in price.