I'm trying to determine which micro-economic/econometrics concepts, models, and/or tools are appropriate for an analysis of promotions.
- Describe the problem in general terms
- Give illustrative details
- Describe my approach thusfar
General Description of the Problem
As is common in microeconometrics, I have a situation wherein a decision maker seeks to analyze returns from promotions (promotions meaning various forms of discounts, not advertising, in this case) with the objective of selecting a promotion strategy which maximizes profits.
Given several practical constraints, in terms of
- data -- which is on the customer-level, going ranging 2.5 years to present
- time -- I need to conduct this relatively quickly, with very little time for literature review, etc (though I read this tangentially related paper by MIT and Oracle which used Integer Linear Programming to optimize grocery store discounts)
- scope -- which requires me to consider the following
- retention -- the idea that promotions attract customers who will remain active for shorter periods of time
- consider the possibility of no promotion as a best solution
- consider the possibility of 1 past promotion type as optimal
- consider the possibility of some optimal combination of past promotions types as optimal
- there are around 10 different types of promotion (10% off for yoga classes, 6 weeks free during Summer, 20% off for military students, etc)
- each type of promotion can be offered to a segment of customers or to all customers
- promotions can be temporary (seasonal) or ongoing
This analysis is being conducted on the prospect and customer level (within-firm), meaning that we have data on customers and people who we've interacted with who may or may not become customers ("prospects").
Specific Details for Illustrative Purposes
A close proxy for the specific type of firm at hand would be a chain of mixed martial arts studios.
In addition to basic student demographics, we have the follow dimensions which may be considered:
- Kung Fu
- Tae Kwon Do
- Women's self defense
- Access to all classes
Student/class skill levels:
- Beginner (white belt - green belt)
- Intermediate (blue belt - red belt)
- Advanced (brown belt - black belt)
- 1 class per week
- 3 classes per week
My Approach Thusfar
Without much time to find an appropriate theoretical framework or empirical model from the literature, I began by creating an ad hoc demand model using Poission regression, wherein
The independent variable, Y = the number of enrollments in a given time period The independent variables, X included:
- a vector of prospect/student demographics
- the type of class, class load, and skill level a prospect/student considered or enrolled in
- the promotions being offered
- lag and lead effects to capture the "cannibalization" of promotion from one term on enrollments in adjacent terms
This gave me some rough estimates of the effects of promotion on leads. However, it did not directly answer questions of optimization, segmentation, or changes in revenue per student from changes in retention.
My 2nd step was to conduct an analysis that estimated the change in revenue per student (over the maximum time range available). I utilized Propensity Score Matching to balance quasi-treatment and quasi-control groups (those who were prospects during time periods where a promotion was offered versus those during a time period with no promotion).
The results showed the decline in revenue per student (broken down by class type, class load, and skill level). The decline revenue comes in part from the cost of the promotion (i.e. the reduction in price) and in part from the change in retention (i.e. students remain active for shorter periods of time, often quitting after their discount ends, etc).
The decision makers would like to know which segments of students respond best/worst to promotion, what's the optimal type of promotion or mix of promotions, etc.
Before going further with ad hoc statistical analyses, I wanted to see what guidance economic/econometric theory can provide to this research project.