Different lags for each past_covariate in past_covariates

[BrandonJenkins0]

From what I understand, the lags_past_covariates argument takes a list or integer and applies the same lags to all past covariates. Is it possible to lag each past_covariate differently when using a regression model? For example, I have the following past_covariates feature1, feature2, feature3 and I want these lags for each:

{'feature1': [-2, -4] 'feature2:' [-1, -7] 'feature3': [-2, -3]}

Any recommendations on how I could accomplish this?

Really enjoying the package by the way!

[hrzn]

Hi, thanks for raising. At the moment this is not possible. However, this would be a nice feature, which we already had in mind for some time. I'll put this issue in our backlog, so we have this somewhere in the development roadmap :)

[feedFloward]

A possible workaround is to lag your features with pandas shift() method individually then turn them to timeseries object and pass to fit.
You need to pass [0] as lags_future_covariates or past_future_covariates since you already lagged them and darts should treat them as they are.

[JanFidor]

Hi @hrzn, what do you think about using something similar to add_encoders param? So basically a nested dictionary where the first key could only have one of 3 values: ["series", "past_covariates", "future_covariates"], and covariates could either be further indexed by their names to get unique lags or have shared ones.

[tsobolev]

After reading all the documentation I had the same question. Perhaps it would be convenient to have different sets of lags for groups of covariates features.

[wojnarabc]

Would be great to have it configurable on per covariate (group) basis, eventually option to exclude some for which we do not wish to create any lags altogether. Especially that calendar features created automatically through add_encoders parameter (any RegressionModel based model types), seem to be also lagged automatically as well.