Quant Library · The machine layer · Football
Residual Gradient Booster
A gradient booster that starts from your model's own log probabilities and learns only what it missed.
Multiclass gradient boosting with the fitted model's log probabilities as the base margin, Newton leaf values, early stopping, and the option to pool leagues so a single division cannot over-fit itself.
Inputs
4What it takes
| Name | Type | Units | Where it comes from |
|---|---|---|---|
| base_probabilities | dict per match | probability | any match model |
| features | array per match | mixed | rest, congestion, recent shots, form, the season so far |
| results | list | - | - |
| depth, rounds, learning rate optional | ints and float | - | your choice; depth 2 is what survives out of sample |
Outputs
3What it gives back
| Name | Type | Units | Notes |
|---|---|---|---|
| boosted probabilities | dict per match | probability | |
| importances | dict | share of split gain | |
| held_out_log_loss | float | nats | beside the base model's and the market's |
Method
How it works
- Set the base margin to log p from the fitted model, so round zero already prices the match.
- 3 further steps in the licensed specification
Assumptions
- The base model is already close to right, so the trees are learning a small correction rather than the signal itself.
Limits
- Per league on a few thousand matches the honest gain runs from about −4% to +1% of log loss; pooled across thirteen leagues it is a few tenths of a percent, positive.
- 1 further limits in the licensed specification
In the package
- The full specification: inputs, method, assumptions and limits
- The validation panel as measured on the day of purchase
- A reference implementation extracted from the running source
- Perpetual commercial use for one entity
Reference modules:
- 2 module(s), named in the licensed specification and shipped in the package
Read this before you buy. The platform's own walk-forward tests say the closing line forecasts football better than these models do. What they give you is a coherent price for every market and match, before the market opens, with the band and the working attached. Any model here that has been beaten by something simpler says so in its limits, and the validation panel above is generated from the platform's result files rather than typed in.