Combine Direct and Recursive Forecasting Strategies allowing to produce longer prediction then training horizon

[anthonygiorgio97]

Description

It would be great to have a new feature that combines both direct and recursive forecasting strategies within the neuralforecast library. Currently, direct forecasting models like NHITS and NBEATS generate all steps in the forecast horizon simultaneously, while recursive forecasting models predict one step ahead and use that prediction for the next step. However, there is no built-in functionality that allows combining these two strategies

Use case

This feature would be particularly useful in situations where the user needs to train a model using a specific forecast horizon (h) and generate a forecast of length n*h, without having to run recursive predictions manually. For example, in other libraries such as statsforecast and mlforecast, users can specify a different forecast horizon for the .forecast() method

[elephaint]

Thanks for the suggestion! I understand the request; from my side I need to think about how this could be integrated in NF.