This message was deleted.

Hi @James Gessel! Yes, all time series should have observations for all timestamps during both training and inference. Imputing missing values is a viable option. Other alternative is to encode missing values (with 0 for example), do not pass them as None or NaN. The

df

input from the

fit

,

cross_validation

and

predict

methods allows for passing an

available_mask

column with 1-0 values to indicate missing data and prevent computing the training loss on this values.

Great, thanks for the excellent reply Christian! I'll try a few things and come back if I'm stuck.