Fine-tuning generative models for short-horizon price prediction
Fine-tuning generative models on Level 3 order-flow data delivers promising sub-second price forecasts, with potential applications in execution and market making.
Paper
Fine-tuning Generative Models for Short-Horizon Price Prediction
Author
Fedor Shabashev
Date
November 11, 2026
Fine-tuning generative models on Level 3 order-flow data offers a promising alternative to conventional approaches to short-horizon price prediction. This paper investigates whether a pretrained generative model can learn useful predictive signals from granular market data.
The author pretrains a 300M-parameter model on order-flow data from two cryptocurrency exchanges before fine-tuning it to predict microprice returns and realised volatility. The resulting model outperforms classical approaches, including gradient boosting on handcrafted features, in predicting microprice returns over sub-second horizons.
More interestingly, the predictive ability transfers zero-shot to assets on which the model has never been fine-tuned. The forecasts are too small to overcome bid-ask spreads and fees through outright trading, pointing instead to applications in execution and market making.