Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

I am training a LSTM model with cross validation. I want non-overlapping forecasts but I get the error "Recurrent models do not support step_size > 1". How can I overcome this? Also, has the 'refit' parameter from the cross-validation been removed?

Hello! By default, LSTM is a recurrent model, meaning it predicts one step at a time until the horizon is complete.

So, in your case, you probably have to use a non-recurrent model, something like N-HiTS or PatchTST, which are models that output the sequence in one shot.

And the `refit` parameter is still in the cross-validation function.

I hope this helps!

For the refit parameter, I get the following error:
```TypeError: TimeSeriesDataModule.__init__() got an unexpected keyword argument 'refit'```


And, I assume that once I put 'refit=1' in cross validation, a recurrent model with step size &gt; 1 shouldn't be a problem cause refit=1 implies the model will be retrained for every predicting every window?

Can you make sure that your are using v.1.7.1?

Okay yes I was infact not using the latest version. I updated it to v.1.7.1 and everything works perfectly! Thank you so much!!