Question on behavior of forecast function for multi-step horizon

[tszumowski]

Suppose I wish to forecast out 8 weeks, and I have a lagged regressor that is lagged 2 steps back. When I call make_future on the time series, I end up with a bunch of NaN values for the lagged regressor, which makes sense. Because for example, if forecasting week 6, you have lags for weeks 4 and 5 in the future. And those aren't known ahead of time.

However, when this is fed into the forecast, the forecaster (specifically CatBoostMultiSegment) takes all inputs except target and timestamp as features to fit and predict on.

This means there are several future forecast weeks that have a lot of missing data. I was curious why this approach was taken instead of applying a Direct Forecasting method as described in this article.

[Mr-Geekman]

Thank you for the question! We are aware that we don't currently support the full Direct strategy: our approach is simplier. However, we are going to add it in the future using current approach as a starting point.