arXiv Releases Empirical Study on Rule-Following AI Trading Agents

Saad Naja, Nasser Odtallah, Jake Van-Praag, and Sarah May published an empirical research paper testing the degree to which 4 AI trading agents adhere to their own strategies and risk rules, and provided an accompanying agent evaluation toolkit to distinguish whether trading results stem from luck or rule compliance before granting agents autonomous permissions or funds. The study is based on 2627 closed trades from February 12 to August 27, 2026, conducted uniformly under the same asset range and execution conditions; 18% of these are Hyperliquid live orders, with the rest being trades simulated at real-time prices, uniformly charged a 0.09% round-trip fee. Each trade requires pre-declaration of strategy, stop-loss level, and target level, with each agent holding a maximum of 5 open positions simultaneously. The paper explicitly states that the trade-level results obtained in this study are not equivalent to account-level returns, nor do they prove that agents have long-term stable trading advantages.

AI-generated from reports · 3 hr agoUpdated

Not enough consecutive observations yet; trend is not drawn.