Hi, I have a question regarding feature engineering. I have time series going from 2015 to today. We use to drop the year 2020 because of COVID and the series during this year don't make any sense. So I've used Statsforecast, drop 2020, use an int frequency to not have problems regarding dates. I wanted to do the same with MLForecast but I wanted to introduce dates features, like months, years, days, weeks and so on... But how can I do it by using feature engineering capabilities of MLForecast and still drop the year 2020? What frequency should I put then? I want to avoid year 2020 while doing a crossval. Thanks

Hey. Are your series daily?

Yes sir, but the same can be applied to monthly dataset

To be honest skipping a year in CV isn't something that's easy to do. You may be better off computing the splits manually. All mlforecast does in CV is just compute the splits and call fit + predict, so the easiest way would be just doing that yourself.

What if I do the mlforecast.cross_validation() but I drop the results for the year... And then do the evaluation without this year? Would it work?

No, the

freq

will be used to build the validation dates in each fold and when it gets there it'll raise an error.

Ok thank you for your help :)