Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hi <@U06BD0YKH44>! The predict method of the `nf` class is not single-step, but rather single-window. It will use the model to forecast the next `h` values. Is there any reason why you set the horizon to `h=1`? The `NHITS` and most models in the library (except for `DeepAR` ) rely on the direct forecasting strategy to forecast all elements simultanously.

Predicting 20,480 values ahead is also much larger than what the usual literature considers long-horizon, it will be an extremely difficult task unless the temporal patterns are basically trivial.

For instance, ETTm datasets have thousands of datapoints in the test set, but the horizon is at most 720, and the predictions are rolled using the new latest data. I am curious on your setting. What type of series and what frequency do they have?