Out of the box the bot uses literature-derived weights for each ensemble model. To squeeze the last percent of edge, you can let it learn the residuals on your machine.
The calibration module fits a per-station bias correction using the last 90 days of resolved markets. It's not a black-box neural net — it's a Bayesian update on a small set of interpretable parameters: model weight, station bias and ensemble variance scaling. After ~30 calibration runs the per-station hit rate typically climbs 2-4 percentage points.
Run it weekly via the included calibrate.py script. The learned parameters are stored in calibration.json and reloaded on every bot restart. No GPU required, no training set needed — the bot's own historical predictions are the dataset.
