Glossary Entry

Quantile Loss

An asymmetric regression loss (also called pinball loss) whose minimiser is a chosen quantile of the target distribution rather than the mean, making it the natural loss for asymmetric error costs.

Metrics Optimization

Also called: pinball loss, quantile regression loss

Seed source: scikit-learn documentation

The quantile (pinball) loss at level tau charges tau times the error for under-prediction and (1 minus tau) times the error for over-prediction. Tilt the charges and you tilt the target: minimising it yields the tau-quantile of the conditional distribution, with tau = 0.5 recovering the median (and absolute error). This is exactly what asymmetric-cost problems need, since the cost-minimising quantile is under-cost divided by total cost, the classic newsvendor fractile.

Beyond point forecasts, training the same model at several tau values produces prediction intervals, which is how probabilistic forecasting competitions like M5 Uncertainty scored uncertainty: a weighted, scaled pinball loss across nine quantiles. Gradient boosting libraries ship it natively (LightGBM’s objective="quantile", scikit-learn’s QuantileRegressor).