Cutting the time series during training

Created by: Dreamlone

There is an idea to add a new hyperparameter to the lagged operation

It will allow to cut down the lagged table. Potentially, it could improve two things:

Models on cutted tables fit faster
For non-stationary autocorrelation function time series, this could potentially lead to more accurate forecasts

Of course, cutting the time series can be implemented as a separate operation cut_ts for example

It could be relevant for time series from such paper