RMATS AI trading system cuts drawdown to 9.6% during geopolitical turmoil

RMATS uses four specialized agents—Sentiment (gauging market mood from news), Report (analyzing company filings and macroeconomic reports), Analysis (technical chart patterns and quantitative signals), and Risk (position sizing and stop-loss rules)—all coordinated by a recursive Manager Agent with iterative feedback loops. The system was tested on a 24-asset multi-class universe spanning equities, bonds, commodities, and currencies from January 2023 to March 2025 (561 trading days). Results show a maximum drawdown of just 9.62%, significantly lower than traditional mean-variance optimization (15.49%) and a FinBERT sentiment-based strategy (15.28%). In 3 out of 5 geopolitical stress scenarios (e.g., simulated conflict escalation and sanctions), RMATS exhibited the lowest event-period drawdown.

While RMATS underperforms return-maximizing baselines during sustained bull markets—by design—ablation studies confirm each agent component contributes meaningfully to downside protection. The recursive feedback loop allows the Manager Agent to dynamically adjust trading decisions based on real-time outputs from all four agents, creating a robust hedging mechanism against geopolitical shocks. The paper positions RMATS as a risk-control-oriented architecture for institutional investors prioritizing capital preservation over aggressive growth. Key technical contributions include iterative optimization of agent weights and a multi-agent coordination framework that reduces max drawdown by nearly 38% relative to traditional MVO.