0

I am working with a time series data (date column and a value column) and I have extracted date features(weekday, weekend etc) and rolling means or the prediction.

For features to be given equal importance, I used min-max scaling i.e., all integer features are converted to range (0,1). My question is Do I also need to normalize the value column? and somehow after prediction since prediction would lie in range of (0,1) Can I de-normalize it back?

bella_pa
  • 103

1 Answers1

1

You can scale your time series if you want to, fit a model, forecast, and scale back. The back-transformation is trivial, since it simply inverts an affine linear transformation.

However, most standard time series forecasting algorithms don't require scaling. Most likely, forecasts on unscaled observations and forecasts after scaling-forecasting-backscaling will be identical. Try it both ways and compare!

Stephan Kolassa
  • 123,354
  • 1
    In my understanding, LightGBM should not require scaling. Just try all four ways (all variables unscaled; only the target variable scaled; only the predictors scaled; all variables scaled) and see whether the forecasts differ. – Stephan Kolassa Nov 17 '22 at 16:36