A foundation model for cross-asset market dynamics

Financial time series are conventionally modelled asset by asset, discarding the coupling that binds markets into one system. We ask whether a single model, trained across asset classes, learns a representation of market state general enough to transfer.

Our architecture treats heterogeneous instruments as tokens in a shared sequence space, with a latent state that captures regime, liquidity, and cross-asset flow. Trained on two decades of tick-level data, the model recovers known stylised facts without supervision and surfaces coupling structure that single-asset models cannot represent.

Evaluated out of sample, the shared representation improves risk-adjusted returns and, more importantly, degrades gracefully across regime boundaries where correlation-based systems fail. We read this as evidence that the model has learned mechanism rather than surface statistics.