Simudyne and Barclays have published “To Trade or Not to Trade: An Agentic Approach to Estimating Market Risk Improves Trading Decisions”, by Dimitrios Emmanoulopoulos, Ollie Olby, Justin Lyon and Namid Stillman.
Most LLM trading agents reason from sentiment or trends and skip the step a quant would insist on: building a model. The paper closes that gap. A builder-critic loop of LLM agents proposes stochastic differential equations for an equity’s price, implements and calibrates them against historical data, tests them, and refines the candidates over successive rounds. The discovered models then produce risk metrics, including value at risk, conditional value at risk and maximum drawdown, which feed the agent’s daily buy, sell or hold decisions alongside trend indicators.
The system was evaluated two ways: in conventional backtests on equity data from September 2024 to April 2025, and inside Simudyne’s market simulator, which generates plausible but fictional price paths and news events so the models cannot lean on anything they saw in pre-training. In both settings, model-informed strategies beat standard LLM agents, improving Sharpe ratios across multiple equities.
The work is an early example of the model-discovery loop that now sits at the centre of Simudyne’s approach: AI that builds and calibrates the model, then tests decisions against futures no historical dataset contains.