1
$\begingroup$

I have to examine historical sales data in order to figure out which calendar events have an influence. I would like to ask for some feedback if it makes sense or what I could do better.

What I have:

  • Historical daily sales data of the last last five years

Here is my plan (so far):

  1. Take every year and transform it into a seasonal adjusted time series. (Not sure yet which of the two methods I should use: X-12-ARIMA or TRAMO/SEATS)
  2. When I subtract the seasonal adjusted graph of a year from the raw data of that year I get a graph consisting of "seasonality + calendar effects", right?
  3. After I have done 1. and 2. for each year I should have 5 graphs which I can compare. Is there any way I can divide calendar effects and seasonality?

I would like to figure out for example what influence Christmas, Chinese new year or Easter has on the sales.

Improve this question
$\endgroup$
4
  • $\begingroup$ What language/tools are you using? They can help the form of the answer to be more relevant. $\endgroup$
    – EngrStudent
    Commented May 5, 2015 at 16:10
  • $\begingroup$ That is completely up to me. So far i had look at Demetra+, Gretl and R. But as i said in the beginning i am not really familiar with the field and happy about any advices! $\endgroup$
    – RandomDude
    Commented May 5, 2015 at 17:43
  • $\begingroup$ check out the following link in "R". You may find it helpful in this context. rdatamining.wordpress.com/2011/08/23/… $\endgroup$
    – EngrStudent
    Commented May 5, 2015 at 18:26
  • $\begingroup$ Engr, there doesn't seem to be any mention of handling of holidays effects in the article... $\endgroup$
    – Tom Reilly
    Commented May 6, 2015 at 6:15

1 Answer 1

Reset to default
1
$\begingroup$

Why don't you use a regression approach with day of the week dummies, monthly dummies, dummies for the holidays? Post your data to dropbox.com and we can take a look. Just specify the beginning date of the data and what country it is for.

Improve this answer
$\endgroup$
7
  • $\begingroup$ Hi Tom, unfortunately i don't have the data yet. I have to come up with a draft for a method "how i want to tackle the data". Could you explain me your suggestion a bit more in detail / give me some keywords (i couldn't find anything with google that made a lot of sense to me considering my task). $\endgroup$
    – RandomDude
    Commented May 5, 2015 at 17:46
  • $\begingroup$ Hi. This post will explain it. Have you used a dummy variable in a regression before? 0/1 stats.stackexchange.com/questions/58657/… $\endgroup$
    – Tom Reilly
    Commented May 5, 2015 at 18:28
  • $\begingroup$ Unfortunately not. So i am kind of starting from scratch... As far as i understood you are saying that i should keep the time series as one part (5years) and use dummies for each day of the week, month and for each relevant holiday? $\endgroup$
    – RandomDude
    Commented May 5, 2015 at 19:33
  • $\begingroup$ That is correct. As you can see from the other older post, there might also be a trend or a change in the trend or a change in the day of the week pattern or outliers. You need to tidentify those and then you can better identify the impact of a holiday. You might also have special days of the month and lead and lags around the holiday too. Take a look at the back of this ppt on slide 45 and read through to see a full breakdown of an example. bit.ly/1EVr48V $\endgroup$
    – Tom Reilly
    Commented May 5, 2015 at 19:49
  • $\begingroup$ Hi Tom, i was digging deeper into the subject and i found "RegARIMA" models that as far as i understand combine regression and ARIMA in the way that you can use dummy variables for holidays etc and the trend, cycle and seasonal components are modelled with ARIMA. Is that something that would make sense for my case? $\endgroup$
    – RandomDude
    Commented May 7, 2015 at 9:27

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.