Market-structure
Prop trading firms including AQR, Susquehanna and Jump Trading are building infrastructure for Kalshi and Polymarket as combined prediction-market volume hit $13.7bn in June, even as a benchmark test found frontier AI models losing 16-30% when trading those markets directly
Kalshi's institutional volume climbed 800% and its annualized volume now exceeds $178bn, drawing serious quant infrastructure. But the Prediction Arena benchmark, which gave six frontier models $10,000 each to trade prediction markets from January to March, found every one of them lost money on Kalshi and roughly broke even on Polymarket.
Prediction markets have crossed from retail novelty to genuine institutional infrastructure this year: combined Kalshi and Polymarket volume reached $13.7 billion in June, Kalshi's annualized volume now exceeds $178 billion, and institutional volume on Kalshi specifically climbed 800%. Named firms building dedicated infrastructure or hiring specialist roles for these markets include AQR, Susquehanna, OKX, Jump Trading, Clear Street and Marex, alongside newer entrants like Propr, founded by a former Credit Suisse quant trader who argues prediction markets suit AI agents well because each contract follows a fixed structure, produces an observable price, and resolves against a clear rule.
That structural argument runs directly into a result this site thinks is worth taking seriously rather than glossing over: the Prediction Arena benchmark gave six frontier AI models $10,000 each to trade live prediction markets between 12 January and 9 March, and every model lost money on Kalshi, with losses ranging from 16% to 30.8%, while performance on Polymarket averaged roughly flat at -1.1%. That's not a small-sample fluke given the spread across six different models and roughly two months of live trading, and it directly undercuts the premise that prediction markets are an easy fit for autonomous AI trading just because the contract structure is simple.
The gap between institutional infrastructure build-out and actual AI-agent trading performance is the real story here, and it maps onto a split this site has now seen play out in equities too: firms deploying AI as a tool inside a broader, professionally-run strategy (the prop firms building infrastructure) are a different proposition from AI systems trading autonomously and directly (the benchmarked models). The former looks like where the institutional money is actually going; the latter, on this evidence, still loses to the house.
Whether that gap closes is an open question the benchmark's authors frame as three scenarios: professional liquidity improving execution for everyone, selective professionalization where only well-resourced firms capture the edge, or concentrated advantage flowing to whoever builds the fastest infrastructure first. Given the 800% institutional volume growth already underway, the third scenario, faster infrastructure concentrating the edge rather than spreading it, looks like the one to watch for as more prop firms follow AQR and Susquehanna into this market.
Read the original: CryptoSlate - Prop trading firms including AQR, Susquehanna and Jump Trading are building infrastructure for Kalshi and Polymarket as combined prediction-market volume hit $13.7bn in June, even as a benchmark test found frontier AI models losing 16-30% when trading those markets directly. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.