search

business-science / modeltime.gluonts

GluonTS Deep Learning with Modeltime
https://business-science.github.io/modeltime.gluonts/
Other
39 stars 9 forks source link

Fitted and train error... #33

Open coforfe opened 3 years ago

coforfe commented 3 years ago

Hi Matt,

Thanks for this package, as well as, the whole time series ecosystem you are building!.

Is it a way to get the fitted values when training a model (on training data) in a similar way as with forecast package ?. And evaluate the error?. You can get test error, but sorry I could not find the way to get that for the train set.

I am looking for a way to asses if the model is overfitted.

Thanks! Carlos.

mdancho84 commented 3 years ago

Hi @coforfe , thanks for the kind words. The ecosystem is really coming together.

For GluonTS models, we don't store the fitted values because they aren't stored in the GluonTS model objects.

I don't know if you can predict backwards. I have to see an implementation where the gluonts model predicts the training data. If this is possible, we can then include fitted data inside the modeltime objects.

coforfe commented 3 years ago

Hi Matt,

Thanks for the clarifications!

Well, in the same direction you mention, it is possible to run several models with different _lookbacklength values. It is not optimum, but at least you have some sense of the level of error.

Thanks again!. Carlos.