I am trying to decompose a time series using Holt Winters method and use it for forecast. I am trying to do this for weekly data of last 25-26 months.

The challenge is that the dates of the seasonal points keep on changing. E.g. Diwali was on 13th Nov in 2012, 3rd Nov in 2013 and 23rd Oct in 2014. This may not give me a very close prediction, especially around this seasonal point. In addition, there are other seasonal points that occur on same date, e.g. Christmas, Valentine Day, etc.

How do I handle this problem? Do I need to shift my time series for the seasonal points to coincide?

  • $\begingroup$ It is not really clear what kind of two time series are you taking? Could you edit your question to add additional information? $\endgroup$
    – Tim
    Commented Mar 17, 2015 at 8:24
  • $\begingroup$ @Tim I am talking about the yearly time series. So, if I have data from 2012-2015, there will be three series 2012-13, 2013-14, 2014-15. In these 3 series the same seasonal point can appear in different positions $\endgroup$
    – naka
    Commented Mar 17, 2015 at 8:54
  • $\begingroup$ What's crucial is exactly when your non-calendar years start and end. If you have a choice then problems such as this can be made minor. All the specific examples you mention are accommodated quite well by years starting on 1 July. If you have no choice, then spikes such as two occurrences of Diwali in one non-calendar year may be unavoidable, but how much difference will that really make to annual totals or means? $\endgroup$
    – Nick Cox
    Commented Mar 17, 2015 at 9:13
  • 5
    $\begingroup$ If you have, or did have, weekly data, then your question is different and the answer is different. But you know when Diwali occurs, as I understand it, so the simplest remedy for forecasting is to build in an indicator variable for a week being Diwali to your model. $\endgroup$
    – Nick Cox
    Commented Mar 17, 2015 at 9:55
  • 1
    $\begingroup$ as @NickCox points out the simplest and a straight forward approach is to use a indicator variable i.e., 1 when Diwali else 0. Unfortunately there is no way to include indicator variables in exponential smoothing. so you are left with regression based approaches like ARIMA, linear regression, UCM etc., $\endgroup$
    – forecaster
    Commented Mar 17, 2015 at 14:46

1 Answer 1


The easiest way is to align your weeks to the current year. So if the week of Diwali I do not know is in week 20 this year but in 2014 it was week 19 then ween 201419 becomes 201420(A) , in this way you construct your alligned serrie.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.