Fine-Tuning an Order-Flow Foundation Model for Sub-Second Forecasting
Shabashev
A 300m-parameter generative model pretrained on Level-3 crypto order flow and then fine-tuned beats gradient boosting at sub-second microprice forecasts and transfers to unseen assets.
Source published · In the Daily Brief of 13 November 2026
Filed under Market microstructure, Quantitative finance
The author pretrains an autoregressive generative model of 300m parameters on Level-3 order-flow data from two crypto exchanges, Bitstamp and Bitfinex, and then fine-tunes it to predict the microprice return and its realised volatility. Earlier work had pretrained such models but, the paper says, none had fine-tuned one for downstream tasks.
The fine-tuned model beats classical baselines, including gradient boosting on handcrafted features, at predicting microprice returns over sub-second horizons, and its predictive ability transfers zero-shot to assets it was not fine-tuned on. The predicted moves are smaller than the bid-ask spread plus fees, so the applications are in execution and market making rather than directional trading.