Autoregressive transformers trained on limit order book (LOB) message streams are gaining attention as foundation models for market microstructure. These models are designed to capture the dynamics of financial markets from sequences of order events. However, because the true generative process of real markets is unknown, it is difficult to tell what patterns the models are picking up.

This paper tackles that problem by training and evaluating such models on synthetic limit order books. The synthetic setting gives the researchers full control over the data-generating process, making it possible to compare what the model learns against a known ground truth. The study is framed as a controlled investigation into the representations and behaviors of these models.

The abstract does not include specific results, so the findings are not yet available from the source. Still, the methodological choice is notable: using synthetic data to isolate what a market microstructure foundation model learns could offer a clearer window into its inner workings than real-world data alone.