Steward AI trader limits live-test loss to 0.3%

Yebo Feng and five co-authors introduced Steward, a nine-agent AI investing system, in an arXiv preprint submitted Oct. 22 and published Oct. 23.

The not-yet-peer-reviewed preprint says Steward held losses to 0.3% during a five-month, 102-trading-day live deployment, despite a 24.6% sector drawdown, a peak-to-trough decline.

The paper addresses a central problem for autonomous financial agents: they must ingest live internet material, act on it and learn from results without letting self-improvement alter their controls.

“LLM agents that act on live Web data must combine heterogeneous evidence, remember what they have done and improve from outcomes without weakening their own safeguards,” the authors wrote in the abstract.

Steward divides work among nine role-based agents. Four analyst agents process news feeds, SEC filings, macroeconomic series and market data, then produce typed reports. A Portfolio Manager sets target positions, while a protected Risk Manager audits them.

The system also uses a seven-layer shared memory built from its own records. According to the preprint, that memory gives each decision information about current positions, the market narrative, calibration and earlier choices by the agent.

Its self-improvement loop is constrained. Code grades trades and forecasts every night, then writes those scores into memory. Each quarter, a reflector proposes prompt lessons, and a deterministic gate can apply them reversibly after replaying recorded inputs.

The authors said those learning tools sit outside an invariant core covering order arithmetic, risk rules and discipline-agent prompts. In plain terms, the system can adjust some instructions, but cannot edit the components that define trading calculations and safeguards.

In replay tests using the recorded inputs, the authors said a single model given the same raw data traded more frequently and lost 1.35 points against the market over five days. The joint Steward design showed no significant loss in that comparison, according to the preprint.

The authors also said no injected instruction was followed during the replay, a relevant claim because financial agents using web content can encounter hostile or misleading text.

A separate 2015-2026 backtest found that the invariant core alone lowered maximum drawdown under all 50 random entry policies, according to the authors. The preprint also says the reflector identified the main failure source and that the gate applied a lesson used on trigger evenings but not on control evenings.

The paper lists the subject areas as artificial intelligence, computation and language, multiagent systems and portfolio management. The authors provided code at a GitHub repository linked from the arXiv page.

The results remain preliminary because arXiv papers are posted before peer review. The source material does not identify an external audit, a production trading account, asset universe details or a regulator-reviewed performance record.

Source: arXiv cs.AI preprint, published Oct. 23, 2026