https://github.com/nixtla logo
#neural-forecast
Title
# neural-forecast
b

Brian Head

10/16/2023, 3:21 PM
When using
predict_insample
I get a warning message for some of the algos (e.g., LSTM, RNN) that the in sample preds may not be accurate due to the scaler (e.g.,
WARNING: Predict insample might not provide accurate predictions for recurrent model LSTM class yet due to scaling.
). Is this message still accurate? Does it only apply to RNN-based models (that's what it looks like from what I'm seeing in the warnings)?
c

Cristian (Nixtla)

10/17/2023, 3:33 PM
Hi @Brian Head, yes it still applies if you are using
scaler_type
to scale the data, and only for RNN methods. It will still return values, but we have seen that because of the way we are scaling, we have seen some performance degradation for the predict insample. If you scale the time series before (with the core class, see this tutorial: https://nixtla.github.io/neuralforecast/examples/time_series_scaling.html ) then this warning does not apply.
❤️ 1
b

Brian Head

10/18/2023, 2:16 PM
Thank you again @Cristian (Nixtla). I'm now having a different issue with
predict_insample
. I started getting an error
Copy code
ValueError: could not broadcast input array from shape (48384,9) into shape (48222,9)
after making the scaling change. It still happens after switching back to what I had. I can't find anything else I've changed. Any ideas?
BTW,
cross_validation
and
fit
both work as expected. It's only a problem with the
predict_insample
.