1
$\begingroup$

I have a data-set containing 7046 unique dealer codes and their monthly sales data from April 2013-August 2018. The Financial Year for the sales data begins in the month of April for a year and ends in the month of March of the coming year. eg FY14 is April2013-March2014.

I have to forecast the sales dealer-wise. Should I go ahead with forecasting the sales for a single dealer using ARIMA and then doing the same for other dealers as well?

I tried the other queries in this platform as well as others where using multiple regression, random forest, SVM and several other techniques are used due to the presence of other predictors, but in my case, I have data in the following format:

 DealerCode | Apr-13 | May-13 | June-13 |......| Aug-18
     DTC06 | 16.08  | 35.285  | 85.26  |......| 81.504
     DTC10 | 15.315  | 0  | 16.26  |......| 0
     DTC128 | 32.535  | 0  | 34.21  |......| 0 
     DTC141 | 39.805  | 18.175  | 9.58  |......| 20.997

The columns where the value is 0 implies no sale has occurred in that particular month. I have thought of imputing the 0 values with median of the entire column. How should I proceed in terms of the types of techniques that can be applied? Please suggest.

|cite|improve this question
$\endgroup$
  • $\begingroup$ Please note that a 0 is not a missing value: it counts the actual number of sales; nothing needs to be imputed. $\endgroup$ – whuber♦ Oct 22 '18 at 13:39
  • 1
    $\begingroup$ Thank you @whuber for the suggestion. The results are in line with the expectation. $\endgroup$ – Ami Oct 30 '18 at 11:17
0
$\begingroup$

"Should I go ahead with forecasting the sales for a single dealer using ARIMA and then doing the same for other dealers as well? "

Yes ! Making sure that your ARIMA model includes pulses, level/step shifts/ , seasonal pulses and possibly local time trends. In this way you are constructing a robust ARIMA model that captures unspeciied latent factors which should be ultimately investigated and possibly understood/vetted. I would use adspend as a possible predictor series for each of the 7046 items as some of them may be impacted by advertising activity.

|cite|improve this answer
$\endgroup$

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.