I have sales data that obtain high sales peaks in December, that i deseasonalized such that there is no peaks. I then also applied the Holt-Winter method on the sales. I then combined the deseasonlized sales data with the result from Holt - Winter by taking the average between the two. The peaks as expected give me smaller values then my previous actuals for those periods, but i am not too bothered about it because where i have big dips in the sales due to stock outs, i get good forecasts that fix the gaps in the sales. That is what i am looking for and the peaks will be replaced by the highest value in the history for that point.

But i want to know is this method a viable way to do forecasting? Is there something wrong with my method or did someone in the past did something similar. In other words combine decomposed sales data with non-decomposed sales data.

  • $\begingroup$ This question might be more suitable for stats SE (crossvalidated). In general, there are many methods to do this kind of forecasting, ranging from regression methods to machine learning algorithms (like random forests or Bayesian networks). $\endgroup$ – Nameless Jan 24 '14 at 15:49
  • $\begingroup$ Simply post your original data as there is no need to do any upfront analysis as good procedures will extract the combined signal(equation). $\endgroup$ – IrishStat Jul 25 '15 at 18:12

To find the best result from Holt Winter's only you should try to use different alpha beta gamma constants. With some combination you wont find the seasonality in the HW residuals. You don't need to de-seasonalise the data then.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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