AI agents are quietly cleaning up the mess in prediction markets.
A paper posted to arXiv this week, "Agentic AI for Clustering, Relationship Discovery, and Semantic Trading in Prediction Markets" (arXiv:2512.02436, updated September 30, 2026, not yet peer-reviewed), describes a system that reads the text of market contracts, not just prices, to find markets asking the same or related questions in different words. The AI first groups markets into topic clusters, then checks pairs of contracts within each cluster for dependency or leader-follower relationships. Tested against a 2026 prediction market dataset, the system's flagged relationships matched actual settled outcomes 62.8% of the time, more than 20 points ahead of a standard natural language inference benchmark's 40.6%. The researchers also report the discovered relationships form an unusually clean signed graph, with a contradiction rate of just 0.324%.
That fragmentation is the real story: overlapping questions, implicit equivalences, and hidden contradictions make it hard for traders to know when two contracts are really betting on the same thing. Automating that cross-market bookkeeping, rather than leaving it to manual arbitrage hunters, is a structural fix, not just another trading bot.
The paper's own trading strategy built on these discovered links posted a 14.12% net ROI after fees over a two-month stretch in 2026, a decent return, but a single backtest on a paper's own benchmark is not the same as a strategy surviving live, adversarial markets.