4
$\begingroup$

Suppose products A, B and C are sold on markets 1, 2 and 3. I have access to three years worth of monthly historical data. I'd like to detect/foresee possible shifts in sales mix for the three products over the three markets (e.g. I'd like to know if market 1 is buying more of product A than it used to, or if product B is cannibalizing the others on market 3).

Does a 5/10/15% increase mean that I'm facing a shift in consumer behaviour? Does it depend on the intrinsic variability of the behavioural pattern? Also how do I choose the optimal time-frame on which to perform such analysis (i.e. how do I know if comparing semester over semester increase makes more sense than comparing quarter over quarter increase)?

Would t-test/ANOVA work here? Or maybe a control chart? Or maybe some intervention detection method to check if there was a level shift in my time series? If so, how do I go about it?

Any input is greatly appreciated!

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ The answers to the following question seem relevant: stats.stackexchange.com/questions/17623/… $\endgroup$
    – Gala
    Commented Mar 7, 2012 at 13:02
  • $\begingroup$ Yeah, I thought intervention detection "a la Tsay" would be an option (which reminds me I need to have a chat with IrishStat =D) so +1 for that. Does anyone else have any other suggestion? $\endgroup$
    – Bruder
    Commented Mar 7, 2012 at 13:48
  • 1
    $\begingroup$ @Bruder I await your contact . You might check out my contact info. I don't necessarily see a role for intervention detection but rather incorporating possible canibilizers as predictor series in a Transfer Function. Annheuser-Busch used our software to relate sales of a product in a particular size as it was impacted by sales of the same product but in different sizes . They also included the price of the possible canibalizers in order to assess the effect. $\endgroup$
    – IrishStat
    Commented Mar 7, 2012 at 21:45

2 Answers 2

Reset to default
1
$\begingroup$

I'd go with control charts. They're not sexy because they're kind of low tech -- but they work. They're also visual, that is, you can spot trends before any numeric flags go up.

I keep recommending this book because of its almost universal applicability.

Improve this answer
$\endgroup$
3
  • $\begingroup$ thank you for your link. I actually own that book, and that's exactly where I got the idea of control charts from. It's good to see we're on the same page. Litteraly :) $\endgroup$
    – Bruder
    Commented Mar 7, 2012 at 19:33
  • $\begingroup$ perhaps the numeric flags you are considering are not "good enough" $\endgroup$
    – IrishStat
    Commented Mar 8, 2012 at 11:30
  • $\begingroup$ @Bruder I actually went out and bought the book but found it lame in many regards. $\endgroup$
    – IrishStat
    Commented Apr 6, 2012 at 21:34
1
$\begingroup$

I think because this is a time series intervention analysis as mentioned earlier may be the best approach. Tom and Dave Reilly who have the software product called Autobox calim that they can do this automatically with their software.

Improve this answer
$\endgroup$

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.