Synthetic market data needs a fidelity horizon
In short: A synthetic sequence is only supported to the last rollout horizon where its marginal, joint, temporal and decision checks still pass.
Jane Street reported on 30 September 2026 that its first diffusion setup for market data sent 88–95% of generated values more than eight standard deviations from the mean across seven continuous targets. Flow matching and a technique called atom smoothing produced much more realistic next events.
But realism degraded as the model generated further into the future, and Jane Street said the system was not yet accurate enough to be a realistic generator. A synthetic sequence therefore needs a fidelity horizon: the last rollout length where marginal, joint, temporal and decision-relevant checks still pass.
Jane Street improved the next event
The Jane Street experiment used four years of US equities data. Each event combined a discrete action, such as a trade or best-bid-and-offer update, with continuous values such as price and elapsed time.
That mixture was awkward for a model inspired by autoregressive diffusion in continuous image space. Trades were only 8% of events. Timing and price also had sharp spikes: simultaneous events, reaction-time boundaries and one-tick price moves did not behave like smooth pixels.
The team moved from a two-class event head to 20 classes, then used atom smoothing to expose and soften the remaining discontinuities. Its reported one-step evaluation conditioned on 10,240 real events, generated one next event and compared the result with real data using feature marginals and a discriminator.
| Reported check | Scope it supports |
|---|---|
| Per-feature marginal divergence | One generated event's individual feature distributions |
| Real-versus-generated classifier | Joint realism of one candidate next event after real context |
| Visual autoregressive rollout | Qualitative degradation as generated context accumulates |
Jane Street reported these checks; I did not reproduce the model or its results.
One realistic event does not certify a sequence
One-step evaluation gives every prediction a clean history of real events. A rollout does not. Its second event depends on the first generated event, and each later event inherits every earlier miss.
Feature marginals can still look right while relationships break. A plausible price change may arrive at an implausible time. A realistic order type may produce the wrong spread response. Even a discriminator trained on one next event can miss drift that appears only after hundreds of synthetic steps.
The source is unusually direct about this boundary. It says the model's predictions naturally degrade further into the future and leaves “by how much?” as an open question. The article provides no released code or source dataset with which I could measure that curve.
My evaluation rule is to publish the fidelity horizon
This is editorial judgment, not a market-data experiment I ran at Trigent or for a client. I would test the same frozen generator at several rollout lengths and publish the last passing point for four layers: feature marginals, joint event relationships, temporal behaviour and the downstream decision the simulator is meant to support.
| Horizon | Marginal | Joint | Temporal | Decision test | Supported? |
|---|---|---|---|---|---|
| 1 event | pass/fail | pass/fail | n.a. | pass/fail | declare |
| 10 events | pass/fail | pass/fail | pass/fail | pass/fail | declare |
| 100 events | pass/fail | pass/fail | pass/fail | pass/fail | declare |
| 1,000 events | pass/fail | pass/fail | pass/fail | pass/fail | declare |
The numbers above are a test template, not Jane Street results. A use case may need another grid, but it still needs an explicit maximum. If fidelity holds for 10 events and breaks at 100, a 1,000-event simulation remains unsupported even when a few samples look convincing.
What's in it for you
- Evaluate generated context, not only predictions made after real context.
- Report the first failing horizon beside the last passing one.
- Cap downstream use at that boundary until a longer rollout passes the same checks.
A believable next event is a sample; a believable sequence needs a measured horizon.
