TLDR: A new research paper introduces an agentic AI system where large language models (LLMs) discover stochastic differential equations (SDEs) to model financial time series. These SDEs generate crucial market risk metrics, which, combined with news analysis, guide daily trading decisions. Evaluated through traditional backtests and a synthetic market simulator, the system demonstrates that model-informed trading strategies significantly outperform standard LLM-based agents, leading to improved Sharpe ratios across various equities. The study suggests LLMs can automate complex financial model development and enhance trading profitability.
Large language models, or LLMs, are increasingly being used in advanced AI systems that perform complex tasks by using various tools. While these systems, known as agentic frameworks, have shown great promise in many fields, including finance, they often lack a structured way to build financial models. Instead, they typically rely on simpler methods like analyzing market sentiment or trends.
A new research paper addresses this gap by introducing an agentic system where LLMs work together to discover mathematical models for financial time series. These models are based on what are called stochastic differential equations (SDEs), which are powerful tools for describing how financial prices change over time with an element of randomness.
The core of this system involves LLMs iteratively discovering these SDEs. Once an SDE model is found, it’s used to generate important risk metrics, such as Value-at-Risk (VaR), Conditional Value-at-Risk (CVaR), and Maximum Drawdown (MDD). These metrics provide a deeper understanding of potential market losses and are crucial for making informed trading decisions.
How the Agentic System Works
The system operates using a ‘builder-critic’ pattern. Imagine one set of AI agents, the ‘builder,’ tasked with taking a proposed SDE equation and turning it into a working, calibrated model that can simulate historical price data. This involves implementing the model and estimating its parameters to match real-world observations.
The ‘critic’ agents then evaluate the synthetic data generated by the builder’s model. They calculate various statistical measures, like the mean, variance, and how often extreme events occur, comparing them to historical data. Based on these evaluations, the critic scores the model and suggests improvements or entirely new models for the builder to try. This continuous feedback loop allows the system to refine and discover better models over time.
To ensure the LLMs aren’t just recalling information they’ve seen before, the researchers used a sophisticated market simulator called Simudyne Horizon. This simulator generates synthetic, yet realistic, price paths and news events. This means the trading strategies can be tested in novel scenarios, including ‘black swan’ events (unforeseen, high-impact events) that the LLMs wouldn’t have encountered in their training data. This advanced testing environment helps confirm that the system’s performance is robust and not just a result of memorized historical patterns.
Trading with AI-Discovered Models
Once a suitable SDE model is identified, it’s used to calculate daily risk and trend metrics. These metrics, along with summaries of recent news articles (analyzed by a ‘news analyst’ LLM), are fed to a ‘trader agent.’ The trader agent, also an LLM, then makes daily decisions to buy, sell, or hold a particular stock.
The research evaluated this system using both traditional backtesting with real historical data (for stocks like Apple, NVIDIA, Microsoft, and Ford) and the synthetic market simulator. The results were compelling: trading strategies informed by these AI-discovered models consistently outperformed standard LLM-based agents that only relied on news context. Specifically, the model-informed strategies significantly improved Sharpe ratios, a key measure of risk-adjusted return, across multiple equities.
Also Read:
- The Purple Agent: A Game-Changing Defense Against LLM Jailbreaking
- Navigating Strategic AI Interactions: Introducing the LLM-Nash Game Framework
Key Takeaways
The study demonstrates that LLMs are already capable enough to enhance existing trading decisions with insights derived from sophisticated financial models. The process of model discovery, traditionally requiring deep technical and market-specific expertise, can now be semi-automated in a structured way. While human supervision is still expected to be necessary, this marks a significant step towards more autonomous and explainable AI in finance.
The paper also highlights that the performance of different LLMs within this agentic framework can vary, and a ‘reasoning’ LLM doesn’t always guarantee superior trading performance. This complexity underscores the need for robust testing and careful integration of LLMs into financial analysis workflows. For more in-depth technical details, you can read the full research paper here.
