“llms catch spurious correlations that backtests miss, halving average losses in prediction markets”
Proposes a two-stage screener that uses Granger causality to find lead-lag pairs across Kalshi Economics markets, then passes candidates through an LLM that checks whether the proposed direction has a plausible economic transmission mechanism based on event descriptions. The LLM re-ranker barely moves the win rate (51.4% to 54.5%), but it dramatically shrinks the downside — average losing trade drops from $649 to $347 by filtering out statistically fragile links that look good in backtests but break in practice.
Some technical background helpful
Platforms mentioned: Kalshi