This publication uses this formula to fit a model to predict the demand of a product A:

Log_Demand_A = constant + b1*log_Price_A + b2*log_Price_B + b3*Promo_1 + b4*Promo_2 + b5*log_Price_A*Promo_1 + b6*log_Price_A*Promo_2

I have 2 questions.

  1. What is done in practise when businesses sell hundreds of products? Could one fit several models for product A (e.g. own for own elasticity and let's say we have 2 other products B and C, another one for the interaction A->B and another one for the interaction A->C and the combine the demand prediction using some kensemble method).

  2. Are the 2 promos in the above formula specific to product A (i.e. do not necesaarily apply to product B)?


1 Answer 1


this comes from a standard cob-douglas utilty function such as $U=A^\alpha\cdot B^\beta$ w.r.t a budget constraint. If you take a monotonically increasing function of U such as a log which leads to $ \ln (U) = \ln (A^\alpha\cdot B^\beta) -> \alpha \cdot \ln (A) + \beta \cdot \ln (B)$ by solving this via a method like subbing in the budget constraint to get the indirect utility form and adding some dummy and interactions to control for the promotions you get the form above. You can see that this is founded purely from first principals but runs into the "too many parameters" problem when you have many products as you correctly anticipated. Imagine we have 100 products and we need to calculate the elasticity for every product pair. Well then we would have 100 equations with 100 elasticity parameters + 100 intercepts and whatever promotional dummies = $ 100 \cdot 100 + 100 10,100$ minimum parameters. As you might not even have this many data points and not enough to measure efficiently you run into problems and need to use other methods that impose more assumptions on the data.

  • $\begingroup$ @csetzkorn Most common approach is multinomial logit. That approach has only 1 price parameter, so all own and cross-price effects are determined by that parameter and market shares. There are generalizations of the approach which are more flexible. Another approach is using a representative consumer and the CES utility function. Solves the problem, but at a cost of flexibility. Again, can be made more flexible. The two approaches are actually closely linked. $\endgroup$
    – Bill
    Jan 5, 2017 at 15:15
  • $\begingroup$ @Bill - thanks could you please be so kind and add some links to these subjects? Thanks! $\endgroup$
    – cs0815
    Jan 5, 2017 at 16:39
  • $\begingroup$ @csetzkorn Sure. Handbook of Econometrics, vol 6A, chap 63. Anderson, de Palma, and Thisse Discrete Choice Theory of Product Differentiation. Kenneth Train has a complete free course including a free text book on his web site: eml.berkeley.edu//~train I would start with the last one, probably, then the first. $\endgroup$
    – Bill
    Jan 5, 2017 at 17:10
  • $\begingroup$ @csetzkorn There is also theigc.org/wp-content/uploads/2016/06/READINGS-Aviv-Nevo.pdf . It has a supplement with code examples here: faculty.wcas.northwestern.edu/~ane686/supplements/…. $\endgroup$
    – Bill
    Jan 5, 2017 at 17:11
  • $\begingroup$ @csetzkorn There is also a large literature on this in Marketing. I don't really know that literature, though, so I can't point you to a good review of it, and a brief google did not bring much up. $\endgroup$
    – Bill
    Jan 5, 2017 at 17:13

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.