A new research paper adds a confidence gauge to algorithmic trading's blackest of black boxes: short-term price prediction.
Researchers built UQ-LOB, a module that bolts onto any pretrained limit order book (LOB) forecaster without retraining it from scratch. It borrows an idea from neural process models: each new forecast gets compared against a set of recent windows whose outcomes are already known, then outputs either a calibrated probability range for the next price tick or a probability split between the price going up, down, or staying flat. Either way, the model also outputs a confidence score. Tested against 5.2 billion order book events across seven cryptocurrencies at 5, 10 and 15 second horizons, the regression version's confidence intervals lined up with reality about as often as advertised - a predicted 68% range held true roughly 68% of the time.
That confidence score is the actual product here. Filtering out the bottom 90% of forecasts and keeping only the most confident 10% pushed directional F1 scores up by 0.11 to 0.15 points for the regression model and 0.05 to 0.11 for the classifier, at every horizon tested. On the biggest, most tradeable price swings, the top confidence tier hit directional F1 scores of 0.88 for downward moves and 0.83 for upward ones at the 5-second mark.
None of this tells a trading bot what to do - it just tells it when to shut up, which for a field full of overconfident price predictions is arguably the more useful skill.