UQ-LOB: Uncertainty-Aware Limit Order Book Mid-Price Forecasting
The paper adds calibrated confidence scores to short-horizon order-book forecasts.
UQ-LOB attaches an uncertainty module to pretrained limit order book encoders, using recent completed windows as context. It has regression and classification variants, producing either a Gaussian tick-displacement forecast or down/up/stationary probabilities. Tested on 5.2 billion crypto order-book events, the regression version reached near-nominal 68% interval coverage. Filtering to the most confident 10% improved directional macro F1 across 5, 10, and 15 second horizons. ArXiv · AI/CL/LG's note
UQ-LOB attaches an uncertainty module to pretrained limit order book encoders, using recent completed windows as context. It has regression and classification variants, producing either a Gaussian tick-displacement forecast or down/up/stationary probabilities. Tested on 5.2 billion crypto order-book events, the regression version reached near-nominal 68% interval coverage. Filtering to the most confident 10% improved directional macro F1 across 5, 10, and 15 second horizons. ArXiv · AI/CL/LG's note
score 3