Autogluon: Time-based crossvalidation in AutoGluon Tabular?

[schinto]

Is time-based-cross-validation supported in AutoGluon Tabular?

[Innixma]

Not at present, however this is an active area where we are investigating to expand AutoGluon's functionality to time-series problems.

As a clarification, this cross-validation technique would not produce predictions on all training rows since the earliest in time rows would never be part of a validation set during cross-val, is that correct? Thus, with 5-fold cross-val only 4 folds (80%) of the data will receive out-of-fold predictions. That is the primary difficulty, as with multi-layer stacking, future layers would have to remove this data.

[kmedved]

That's correct, yes @Innixma - the earliest in time rows would never be included in the validation data. Depending on the size of your training data, you can do 10 folds to ensure that 90% of your data receives a prediction, etc..., but to my knowledge there's no way around the core issue you raise.