Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

Hi! It is not yet available in pip. To have the refit parameter you will need to clone the repo and install it will `pip install .`

<@U03619XRWKD> I see, thanks for this info!

<@U03619XRWKD> with refit equals to True, every window will be retrained with the new cutoff date data.

What's the training data for each window when refit equals to False?

its only the initial training data, before all the test windows

I see, that's quite a huge data not being used for training when the n_windows is large. Thanks for the confirmation Christian! :+1:

The refit should ideally reflect how often you will retrain the model in production (for industry applications), and that varies a lot. In academia, we evaluate models without `refit` on thousands of windows.

Interesting to know that evaluation practice. Really appreciate it!