2

I need to do sales forecasting.My historical data shows stationary pattern & present of trend,Seasonality & cyclic pattern. I would like to check with you that how to select between Holt Winters and ARIMA to decide the method to be used for future prediction?

Thanks

Michael R. Chernick
  • 42,857
R Learner
  • 529
  • 5
  • 13
  • 1
    The Holt Winter's forecasting method is simple exponential smoothing which is a special case of ARIMA models, namely the IMA(0,1,1) model. – Michael R. Chernick Feb 14 '17 at 14:20
  • 3
    To be more precise, I would say either ARIMA(0,1,1) or IMA(1,1) since the zero in IMA(0,1,1) indexes the AR order that is excluded from the model name. R Learner, check also earlier threads on the subject; similar questions have been asked before. E.g. this would be a duplicate, but it does not have a good answer. Hmm, so probably we still need an answer to this one. – Richard Hardy Feb 14 '17 at 19:45
  • @RichardHardy: do you want to post your comment(s) as an answer? Better to have a short answer than no answer at all. Anyone who has a better answer can post it. – Stephan Kolassa Sep 18 '19 at 11:33
  • @StephanKolassa, I do not think my comment constitutes an answer to the actual question (about model selection for forecasting). Rather, it is more of a clarification regarding terminology. – Richard Hardy Sep 18 '19 at 11:46
  • @RichardHardy: I think it's as close as we will get to an answer to this question... I'll post a short answer. – Stephan Kolassa Sep 18 '19 at 11:50

1 Answers1

2

As Michael Chernick and Richard Hardy point out, there are relationships between Exponential Smoothing and ARIMA models that essentially mean that the forecasts won't differ by much (differences mainly driven by choices in the initialization).

I would recommend that you run a holdout test and choose the method that performs best on a holdout sample. This may be helpful.

(See here for a motivation for short answers. Longer answers are always welcome.)

Stephan Kolassa
  • 123,354