Clarify intended behavior of model.backtest and model.historical_forecast with multiple series and allow training on multiple series per fold when backtesting

[eschibli]

Is your feature request related to a current problem? Please describe. It is unclear from the documentation that model.historical_forecasts and model.backtest will only ever train on a single series at a time if multiple series are provided, in contrast to model.fit which will train on every series provided.

Describe proposed solution Add a Boolean keyword argument to model.backtest and model.historical_forecast to train on all "past" samples in each series for each fold of backtesting.

Describe potential alternatives Clarify the documentation and possibly provide an example of how to achieve the proposed functionality.

[madtoinou]

Hi @eschibli,

The documentation states that "The (or a sequence of) target time series used to successively train and compute the historical forecasts". How would you change the wording to make it clearer that the method is going to process them independently?

Training on multiple series is only supported for global models, I am worried that introducing such an argument might cause a bit of confusion (since an additional argument would be needed to indicate which series should be predicted) but I understand that using historical_forecasts() with retrain=True and a global model trained with several series is a feature that you could expect from the library . WDYT @dennisbader?

[eschibli]

Hi @madtoinou, I would simply state "The (or a sequence of) target time series used to successively train and compute the historical forecasts. (Regardless of whether a model is global, it will only train and infer on a single series at a time.)," or something similar.