Introduced on day 24
Direct forecasting
Multi-step forecasting with a separate model for each horizon, each trained to predict the value h steps ahead from features available at the cutoff.
In detail
Direct models avoid compounding errors and can use horizon-specific features, at the price of one model per horizon (or one model with the horizon as a feature). Every lag and window must end at the cutoff, so the shortest lag equals the horizon. Multi-output models are a middle option that return all horizons from one model.