The research team introduced the Signature-Informed Transformer (SIT), a model designed to capture the evolution of market prices and asset correlations rather than simple end-point forecasting. Tested across equity markets in the United States and China, the SIT framework delivered superior risk-adjusted returns compared to traditional forecasting methods. This approach emphasizes decision-focused outcomes, allowing the model to account for volatility and risk factors that standard tools often ignore.

Beyond model architecture, the researchers scrutinized 164 finance-focused large language model studies published between 2023 and 2025. They identified systemic flaws, such as survivor bias, the accidental inclusion of future data, and the omission of transaction costs, which frequently inflate performance metrics. To combat these issues, the team proposed a Structural Validity Framework—a rigorous checklist designed to ensure that AI performance claims hold up under real-world economic constraints. By treating financial AI like a flight simulator, the researchers aim to move the field toward transparent, stress-tested systems that function reliably before real capital is deployed.