Nixtla's Open Source Time Series Ecosystem.

Nixtla Community

image.png

*Hi everyone!*

I've been reading the <https://nixtlaverse.nixtla.io/mlforecast/docs/how-to-guides/analyzing_models.html|documentation> on SHAP values, and something caught my attention: when I transform my target variable, the output of the waterfall plot becomes harder to interpret.
How can I properly adjust the output while still keeping the target transformation for the model?

You need to inverse transform manually the shap values per series
You can find the stddev and mean per series here:

# Create dictionary for stdscaler
 scaler_dict = {
     unique_id: [
         scaler_stats[0],  # Mean mu
         scaler_stats[1],  # Std deviation sigma
     ]
     for unique_id, scaler_stats in zip(
         fcst.ts.uids, 
         fcst.ts.target_transforms[0].scaler_.stats_,
     )
 }

Then the base value should be rescaled with inverse formula (X*sigma + mu)
The shap values shoudl be rescaled only with X*sigma

It would be awesome to have a method included in nixtla library by the way

Oh also, this works only with linear scaling (e.g. for logscaling or boxcox should not work)

It's clear. Thanks a lot :slightly_smiling_face: