Short‑horizon mid‑price forecasting from limit order book (LOB) data is a cornerstone of algorithmic trading. Most deep LOB forecasters are point predictors that output only a direction or a displacement, without indicating the reliability of each prediction. UQ‑LOB introduces a lightweight, encoder‑agnostic uncertainty quantification (UQ) module that can be attached to any pretrained LOB encoder. Inspired by attentive neural processes, the module conditions each forecast on a context set of recently completed windows whose outcomes are already known.
The UQ‑regression variant produces a calibrated Gaussian distribution over the future tick displacement $\mathcal{N}(\mu,\sigma^2)$ and emits a scalar confidence (predicted signal‑to‑noise ratio). The UQ‑classification variant yields a categorical distribution over down/up/stationary and provides class probabilities as confidence scores. These confidences enable selective prediction, allowing the system to act only on high‑confidence outputs.
Experiments on 5.2 billion LOB events across seven cryptocurrency assets and horizons of 5, 10 and 15 seconds show that UQ‑regression attains roughly 68% interval coverage. Restricting to the top 10% most confident predictions raises macro F1 by 0.11‑0.15 for regression and 0.05‑0.11 for classification at every horizon. For large, economically meaningful moves, the tightest confidence tier reaches directional F1 scores of 0.88 (down) and 0.83 (up) at the 5‑second horizon.
Key benefits include: 1) compatibility with existing encoders without retraining; 2) direct quantification of prediction uncertainty via Gaussian or discrete distributions; 3) confidence scores that can be used for risk management and trade‑filtering.
Review