Live LLM Trading Fleet Posts 41% Win Rate vs Retail's 50%

Autonomous LLM trading agents running on live capital lost more roundtrips than a matched retail benchmark. Over roughly six months, one fleet analyzed in a new arXiv study posted a "41% vs. 50% roundtrip win rate" against retail traders, across a dataset covering "7.5M single-model invocations with about 300K onchain actions" and more than 14,500 trades on two platforms.

The paper's sharper point sits in the machinery around the model rather than the model itself. The authors report that "Agent fixed effects accounted for 60% of variance in outcomes," and that a single control setting produced 11% of positions but 62% of liquidations. Median leverage stayed pinned near 5.0x regardless of market volatility. Moving the platform's risk slider one level added roughly +0.425 to leverage on average, and leaderboard visibility skewed which agents users picked, with a 1.75x jump at the top-3 boundary.

Even the wins tended to melt. "43.2% of positions saw at least +300 bps of favorable excursion within 24h," the authors write, but "49.3% of those closed with a negative trade return." A mechanical bracket strategy, run for comparison, recovered +39.0 bps per position that the live agents left on the table.