Fit model before historical forecast method?

[ETTAN93]

Do I actually need to fit the model before implementing the historical forecast method? From what I understand, even if you do not fit the model, if retrain is set to False, the historical forecast method already fits the model once.

However, if I do not fit the model before historical forecast, I notice that setting retrain = False or setting retrain = True makes no difference in the speed of the method. If I do fit the model before, then it makes a difference.

Is that expected?

[madtoinou]

Hi,

This is a duplicate of #1617, it is the expected behavior introduced by #1465 to avoid mutating the underlying model object. You indeed should fit() the model before using historical_forecasts.

The model gets retrained because it was not fit before (Darts requires the models to be fit before prediction), regardless of the retrain argument value.

[ETTAN93]

@madtoinou thanks for letting me know. just to be sure I got the methodology right, do I fit on the entire dataset that I have before doing historical forecast or do I only fit on the training set, i.e. between start_date and split_date, that is used by the historical forecast?

[dennisbader]

If you split your data into train and test split with my_series.split_before(split_point=split_date), then use split_date as your start in historical forecasts. This will perform historical forecasts only on your validation set (the train set is ignored).

# split series into train and test set
train_series, test_series = my_series.split_before(split_point=split_date)

# fit your model on train set
model.fit(train_series)

# perform historical forecasts on val set by setting `start=split_date` (use the original series)
model.historical_forecasts(series=my_series, ..., start=split_date, retrain=False)