What prediction length for one step-ahead forecast?

[mtomic123]

I would like to build a DeepVAR model that predicts only one-step ahead. Given my hourly data, this means I would only like to predict the next hour.

I am not sure how prediction_length works exactly and whether it is advisable to use prediction_length=1 or to build a model with a larger prediction_length but just use the first prediction.

I see three different options for my use case:

1. prediction_length=1
2. prediction_length=1 but context_length > 1
3. larger number such as prediction_length = 24 , and take only first prediction

Any insights?

[jaheba]

If you want to predict just one step ahead, prediction_length should be 1.

I would assume that the results wouldn't differ much compared to prediction_length = 24, but notice that in that case the model tries to optimise for 24 values instead of the one that you are interested in.

I would set context_length to a frequency dependent value, in your case maybe to 12 or 24.