Skip to content
ML Academy

Introduced on day 25

Rolling-origin evaluation

A backtest that moves the forecast origin forward in steps, refits on data before each origin and scores the forecasts for the following horizon.

In detail

Also called time-series cross-validation. The training window may expand (all history) or slide (the last W points). Each origin yields forecasts at horizons 1 to h, so accuracy can be reported by horizon and across origins. Results at neighbouring origins are correlated, so many origins reduce variance but are not independent trials. Baselines are scored on the same origins.