Skip to content
Quantoros

Article

A market that never happened

TradeFM from J.P. Morgan AI Research generates a stream of market events for a simulator. What it reproduces, why quants should care, and why matching statistics isn't enough.

Perhaps the most useful neural network for trading won't be the one that looks for entry points, but the one that builds a market where other algorithms learn to make decisions. That's a more interesting framing than yet another attempt to guess the next candle.

In February, researchers at J.P. Morgan AI Research introduced TradeFM. It's a 524M-parameter transformer trained on billions of trading events across more than 9,000 US stocks. Its job isn't to produce trading signals but to generate a realistic stream of market events.

The mechanics resemble text generation, except that instead of the next word, the model predicts the attributes of the next event, including side, size, and arrival time. The simulator then processes that event, updates the order book, and feeds the resulting state back to the model. The result is a closed loop in which every action changes the conditions for the next one.

Why should quants care?

Historical market data shows what happened given the specific actions participants took, but on its own it can't tell you what would have happened after your large order. Maybe liquidity would have dried up sooner, execution would have been more expensive, and subsequent trades would have printed at different prices. Studying effects like these requires interactive simulators. This line of work existed before TradeFM, for example in the ABIDES family of projects.

What it manages to reproduce

In the experiments, TradeFM combined with the simulator reproduces heavy tails in the return distribution and volatility clustering. In other words, sharp moves don't become implausibly rare, and periods of high volatility tend to cluster together. What's being tested is specific statistical properties, not how pretty the chart looks.

Potentially, an environment like this lets you study trade execution, run stress tests, and train trading agents. But the authors explicitly stress that TradeFM's practical usefulness for these tasks has yet to be demonstrated. For now, it's a research result, not a ready-made substitute for the real market.

Similarity is not a response to intervention

Statistical similarity doesn't yet mean a faithful response to intervention.

Two environments can have similar return distributions and respond completely differently to the very same order. So the next important question isn't how convincingly the simulator imitates the market, but which decisions it lets you transfer from the experiment to reality. That, to me, is the main point of this work. An attempt to build a lab for trading decisions. Its value will be measured by how many bad ideas it helps filter out before they meet the real thing.

Further reading

New articles by email

A market that never happened | Quantoros