https://github.com/nixtla logo
d

Daniel Falbel

04/28/2022, 4:36 PM
What’s your take on normalization strategies for neural models? Specially when predicting for multiple time series with different scales? I’ve been normalizing per group, but seems that pytorch-forecasting encourages normalizing per observation? Have you experimented with different approaches?
m

Max (Nixtla)

04/28/2022, 7:49 PM
@Kin Gtz. Olivares?
c

Cristian (Nixtla)

04/28/2022, 11:15 PM
In our experience normalizing per time-series is the best alternative
although we haven't tried grouping time-series and then normalize separately each group
d

Daniel Falbel

04/28/2022, 11:20 PM
I think what I meant is the same as you mean with ‘per time series’. Ie, if you have sales per store/department, you’ll normalize per store/department right?
k

Kin Gtz. Olivares

04/29/2022, 12:09 AM
@Daniel Falbel : You might one to take a look to the (N-BEATSx normalization)[https://github.com/cchallu/nbeatsx/blob/main/src/utils/experiment/utils_experiment.py] strategies. We have had very good experience using the Mean Absolute Deviation (MAD) instead of the classic std normalization. MAD is robust to outliers.
c

Cristian (Nixtla)

04/29/2022, 12:17 AM
Yes, its the same
d

Daniel Falbel

04/29/2022, 5:04 PM
Cool! Thanks for the pointers @Kin Gtz. Olivares ! Will try that too