Covariates with mixed frequencies.

[mg10011]

Is your feature request related to a current problem? Please describe.
I don't know if this feature is available or not. To my knowledge perusing the documentation, it is not, so I will ask for it. Is it possible to have the main time series be a univariate daily series that has calendar days ('B') as its frequency and many past covariates and future covariates:
-- some of the past covariates are also daily data with the same frequency as the main data series.
-- some of the past covariates are measured and sampled only on a weekly, monthly, or quarterly basis
-- some of the future covariates, like day of the week or day of the month are sampled daily, including non-business days.

Please let me know if your library can handle this.

Describe proposed solution
A way to downsample the covariates (from monthly to daily, for example) where it is impossible to "peak ahead".

[dennisbader]

Hi @mg10011. In Darts, you need to convert all series (past_covariates and future_covariates) to the same frequency as your target series.

prior to creating a time series, you can do some preprocessing (e.g. with pandas):

• for data with daily frequency, you could exclude weekends and non business days.
• for lower frequency data (weekly, monthly, ...) you would have to upsample to "B" frequency
• for converting from higher frequency to lower frequency, we recently added the MIDAS transformer (see Is it possible for the model to make predictions over the training period?)

[dennisbader]

Closing, as MIDAS was added in #1820