Accuracy & evaluation

Quantile forecast

A quantile forecast predicts a percentile of the future distribution — p50 as the middle, p90 as a high plan — instead of (or in addition to) a single expected value.

Pinball loss is how you train these without pretending errors are Gaussian. You can also take samples from a generative model and read off percentiles.

A mean forecast optimized for MAE is not p50 unless the error is symmetric. If stockouts hurt more than leftovers, ask for a high quantile on purpose. That is a different product than 'the line in the middle.'

Serving has to return the percentiles. A PNG of a fan chart is not a quantile forecast. If the API only gives you a center, you are back to pretending.