A Transformer trained to forecast stock order books turns out to double as a market-impact simulator, according to a new arXiv paper.
The study, posted to arXiv as paper 2609.16930, tests a model-agnostic method for measuring how a trained limit-order-book forecaster's predictions shift when researchers inject counterfactual but mechanically valid order messages - the digital equivalent of asking what if this trade had happened instead. Applied to a Transformer-based forecaster, the method recovered the correct ranking of simulated scenarios with a Spearman correlation of 0.99 against real historical outcomes. It also matched the direction of actual market moves 97.2% of the time among scenarios that were not neutral. The paper also reports that the model's estimated impact tracked the specific sequence of events, not just the scenario's broad category, meaning it captured more than a coarse label.
That distinction matters because most market-impact tools are purpose-built and narrowly calibrated, while this result suggests a general forecasting model can be repointed at counterfactual analysis without retraining. For trading desks and risk teams, that could mean one existing model doing double duty - forecasting the book and stress-testing hypothetical orders - instead of maintaining separate systems for each job.
The numbers are strong, but they come from a single architecture and dataset in one paper; whether the approach holds up across venues and volatile regimes is still untested, which is exactly the kind of claim that deserves scrutiny before it ends up powering a live risk desk.