This article by Wilson Chan, Founder and CEO at Permutable AI explores LLM trading applications and specifically whether large language models can effectively simulate historical market conditions for trading strategies, aimed at quantitative analysts, macro strategists, and systematic trading teams considering AI integration in their workflows.
There’s a lot of noise out there about the power of large language models. Some say they’re set to reinvent everything from research to reasoning. Others believe they’re black boxes that have no place in trading or investment workflows. But there’s one question we’ve been discussing a lot here at Permutable HQ recently and it’s one I am sure has also been discussed equally by quants, macro strategists and systematic trading teams:
Can LLMs simulate the past — and should we be using them to do so?
As someone building LLM-trading applications for institutional use, I think the question is both the right one and the wrong one. Let me explain.
LLMs don’t “remember” the past — they model relationships
The core appeal of an LLM isn’t that it stores facts like a database. It’s that it learns the structure and relationships between ideas, entities, and events. That makes it incredibly useful for LLM trading applications such as surfacing patterns in how markets respond to macroeconomic events, geopolitical risks or even extreme weather.
But when it comes to simulating the past – for example, recreating sentiment dynamics during the 2008 financial crisis – we need to be cautious.
If your LLM has been trained on post-2008 commentary and news, and you ask it to simulate 2008 sentiment, what you may get is hindsight – not history. That’s not simulation. That’s leakage.
What we’re really after is dynamic understanding
In systematic trading and macro analysis, we don’t just want to know what happened – we want to understand why it moved the way it did. This is where I believe LLM trading applications are extremely powerful, but underutilised.
At Permutable, we don’t use LLMs to reconstruct old data. We use them to extract dynamic relationships across thousands of entities in real time – currencies, commodities, macro indicators, regions – and track the sentiment landscape as it evolves. That way, we’re not guessing. We’re building explainable context around what’s moving and what it might signal.
Why some are getting burned
Perhaps some think they can plug LLMs into their backtesting engines. It seems smart on the surface of things – until you realise the model “knows” things it shouldn’t. The results look good, but they’re meaningless. You’re not testing your strategy. You’re testing the model’s memory.
And while public LLMs are great for general research, in terms of LLM trading applications, they’re often unsuitable for financial simulation or decision support. Fundamentally this is because of four key reasons which are:
Their context window is limited
They lack market-specific grounding
Their output is variable, not deterministic
They hallucinate — especially on unfamiliar or niche domains
For regulated, high-stakes environments, that’s not just a limitation – it’s a liability.
Build vs buy: What the smart teams are doing
Another decision many teams face when considering LLM trading applications is: do you build your own internal model, or work with a trusted partner?
Of course, building internally offers control – but it’s resource-intensive. You need data pipelines, prompt engineering expertise, monitoring systems, and constant tuning. And that’s before even addressing governance or explainability. By contrast, buying off-the-shelf tools can be faster, but often comes with trade-offs in customisation and trust.
At Permutable, we’ve built a middle path for LLM trading applications. We offer LLM-powered insights and multi-entity sentiment intelligence that’s ready for real use via our plug and play solutions — but with full transparency and customisation where needed. It’s the “build-quality” edge, delivered as a service.
The upshot?
Don’t try to use an LLM to re-run the past. Use it to help you interpret what you’re seeing now — and where the signal might lead.
That’s where the edge is. Not in recreating a news feed from ten years ago. But in understanding, in real time, how today’s news fits into the wider economic and market structure – and how that’s evolving. LLMs, when applied responsibly, are exceptional at mapping sentiment shifts, surfacing hidden correlations, and giving context to volatility. But only when you trust the source, the data, and the reasoning.
Looking ahead
The world of finance doesn’t need more noise. It needs clarity. As we move deeper into this era of machine-led intelligence and LLM trading applications, the challenge for every institutional team is the same – how do we harness new technologies like LLMs without compromising on rigour, transparency, or performance?
That’s the question we’re building for every day at Permutable. And if you’re asking it too – let’s talk. Feel free to connect with me on LinkedIn and reach out via DM.