Is there method I could use to predict time series of length more than

h

? for example, I train a nbeatsx model which have a predict length

h

of 24, now I want to predict the time series with length of 48, how could I do it?

Hi @Ibfan, For the moment all NeuralForecast architectures use a direct multi-step forecasting strategy. For this reason the decoders of the networks are restricted to output a fixed horizon

h

. What we have done in the past was fit a long enough horizon and trim for other forecasting needs.

Thanks @Kin Gtz. Olivares, are we planning to add new function/method to support more flexible prediction window?