For years in financial services the application of natural language processing (NLP), and more recently with large language models (LLMs), has largely centered on quick insights derived from sentiment analysis or trend following.
By Richard Harmon, vice-president: global financial services industry at Red Hat
While valuable, these approaches often lack a deep, principled understanding of market dynamics. But what if AI could go beyond simply reacting to information and instead discover the underlying mathematical models that govern financial time series?
A groundbreaking new approach, detailed in a paper by quants from Barclays and Simudyne, titled “To Trade or Not to Trade: An Agentic Approach to Estimating Market Risk Improves Trading Decisions,” is doing just that.
This innovative system leverages the power of LLMs within an “agentic” framework, creating a sophisticated multi-agent system that not only informs trading decisions but fundamentally enhances market risk estimation through model discovery.
The core idea: agents working in harmony
At the heart of this system is a collaborative network of specialised AI agents, each with a distinct role, working together to achieve more robust and profitable trading outcomes. Here’s a look at the key players and their synergistic interaction:
- The Risk Analyst: the brains behind model discovery – This is arguably the most pivotal agent in the system. The Risk Analyst is responsible for identifying the most suitable stochastic differential equations (SDEs) to model historical financial price paths. This isn’t a one-and-done process; it’s an iterative, self-improving loop guided by a “Builder-Critic” pattern:
- The Builder: Think of the Builder as the architect and engineer. Given a proposed SDE model, this sub-agent’s role is to implement and simulate it, generating synthetic price paths. It crafts the necessary programs and even suggests initial parameters for calibration.
- The Critic: The Critic acts as the discerning evaluator. It rigorously tests the synthetic data from the Builder against historical market data using a comprehensive suite of statistical metrics (moments, tail metrics, Hurst exponent, etc., as well as direct distribution comparisons like Kolmogorov-Smirnov and Wasserstein distance). Based on these evaluations and a “novelty score” for the proposed equation, the Critic scores the model and, crucially, suggests new and improved SDEs for the Builder to explore in the next iteration. This continuous feedback loop drives the system towards more accurate and representative financial models.
- The News Analyst: providing crucial market context – In the fast-paced world of finance, news impacts markets immediately. The News Analyst agent is designed to keep the system informed about the broader market landscape. It actively gathers recent news headlines for a company and its related entities, summarising the key information, identifying potential pros and cons for investment, and providing crucial context-specific insights to the Trader.
- The Trader: making the daily decisions- With the deep analytical insights from the Risk Analyst and the timely market context from the News Analyst, the Trader agent steps in to make the critical daily trading decisions: whether to buy, sell, or hold. It processes model-informed risk metrics (like value-at-risk, conditional value-at-risk, and maximum drawdown), trend metrics (such as the relative strength index and the SDE’s inherent drift polarity), and the news sentiment to formulate its trading strategy.
How model discovery fuels trading decisions
The workflow is a sophisticated collaboration between these agents:
- Monthly model discovery: At the start of each month (or other periodicity depending upon market conditions), the Risk Analyst (via its Builder-Critic loop) works tirelessly to identify the optimal SDE model based on the past six months of historical price data. This ensures the foundational model is highly relevant and adapted to recent market conditions.
- Daily recalibration and insight generation: Every trading day, the chosen SDE model is recalibrated to the very latest market data. Using this refined model, the system calculates critical market risk metrics (VaR, CVaR, MDD, expected shortfall and even tail risk insights from extreme value theory). Simultaneously, the news analyst fetches and processes the latest news.
- Informed trading decision: All of these inputs – the precise risk metrics, the trend indicators (RSI, drift polarity), and the summarised news sentiment – are then fed to the Trader agent. This comprehensive market context allows the Trader to make a more informed and potentially more profitable decision, moving beyond simple heuristics or superficial trend analysis.
Beyond historical bias: the power of synthetic data
A key innovation in this design is the use of a market simulator (Simudyne Horizon) to generate synthetic but causally plausible price paths and news events.
Why is this important? LLMs are trained on vast amounts of historical data. While powerful, this can lead to biases or a lack of robustness when faced with “black swan” events or market conditions outside their training distribution.
By using synthetic data for backtesting, the system can evaluate its trading strategies under diverse, realistic, yet entirely novel scenarios, ensuring its resilience and adaptability in unprecedented market environments.
The future of AI
This agentic approach to financial trading represents a significant leap forward in developing AI embedded trading strategies. By embedding a principled model-building step into LLM-driven financial systems, it moves beyond typical sentiment analysis towards a deeper, more robust understanding of market dynamics.
The study’s results show that this model-informed approach significantly outperforms standard LLM-based agents, leading to improved Sharpe ratios across multiple equities.
While human supervision remains crucial, this research demonstrates that current LLMs are already capable of augmenting trading decisions with sophisticated, model-informed insights, semi-automating a process previously thought to require extensive human expertise. As LLM capabilities continue to advance, we can expect even greater precision and profitability from these intelligent, collaborative AI agents analysing the financial markets.