Autonomous AI trading is easy to demonstrate and hard to evaluate. NOVAQ designed a test that could fail.
The design
The AI trader runs on a simulation account behind the same gate as other agents. In September it entered a formal trial with five criteria: positive net result, at least 30 trades, a drawdown limit, no single day dominating profit and no rule violations.
What was measured
A counterfactual replayed 203 trades without the agent's active management and found an effect indistinguishable from zero, which located the losses in entries. Slippage in points exceeded edge at the tight stops the agent favoured. Its confidence score did not predict results.
The verdict
On 12 October, after 168 trades in 17 sessions, the result was edge not demonstrated, with a recommendation not to deploy real capital. The trial was extended and its conditions changed during its course, so it is best read as a verdict and not as a clean experiment.
What the work produced
NOVAQ now has infrastructure to run an AI agent under real rules with a full decision record, and a measurement set that shows why it wins or loses. The experiment continues with simpler rules.
Limits
The test concerns one agent, one configuration and one period, and a simulation account differs from live execution.