Exploration of Econometrics Modeling to study Promotion Profitability in the CVS Front Store

Sucheta Jawalkar, Duygu Yasar, Haotian Wu, Tarjani Shah, Duygu Dagli

We use econometric modeling and basket analyses to study the challenging problem of measuring the true impact of promotion profitability faced by retailers like CVS. The work was inspired by Ailawadi et al (2006), who used CVS data and claimed that about 45% of promotions [yellow tags] do not generate the expected margin. We complemented this work by reviewing the updated methods from the 2018 master’s thesis by Pesonen.

When an item (SKU) is promoted, it generally stimulates demand temporarily and generates a unit lift. Measuring the performance of a promotion based on this lift only provides a partial view. There are three more major effects to consider. These are:

1. Halo
2. Cannibalization via Switching
3. Cannibalization via Stockpiling

Halo refers to the item being placed on promotion in one category and as a result there are other purchases across the store as the customer takes advantage of that promotion. Concretely, if there is a promotion on peanut butter, the accompanying halo will include bread.

Cannibalization via Switching refers to an item placed on promotion transferring demand from non-promoted items in the same category. For example, if CVS brand Ibuprofen is on promotion, demand could be transferred from name-brand Ibuprofen.

Cannibalization via Stockpiling refers to the impact of the customer buying large quantity of items that are on promotion and then no longer needing that item for a long time in the future. In the CVS Front store context, an example would be promotions on Cereal that lead to stockpiling of Cheerios.

As a start we focused our study on Halo and Cannibalization via Switching.

Halo

To study Halo, we conducted a basket analysis over one year to study the most purchased categories together in customer transactions. At the fortune 4 scale, this represents a significant computational exercise, and we were able to get significant code optimization enhancements by switching our tech stack from pandas to NumPy to get the basket analysis computed in under 5 mins for the entire CVS assortment.

We saw a clear trend of consumables purchased in the same customer basket demonstrating the association needed for the Halo effect.

Cannibalization via Switching

For the case of cannibalization, we built a regression model leveraging the CVS Front Store Product Hierarchy I.e. Category, Sub-category, Segment, SKUs. An example of this product hierarchy is Personal Cleansing Category has a Sub-category for Body Wash Products which has a Segment for Bodywash Men’s. We assumed that the switching would happen at the segment level.

At CVS, planned promotions generally cover a group of products (SKUs) called an ad block. For example, a Buy 1 Get 1 50% off Vitamin promotion will impact multiple vitamin SKUs. These SKUs will generally be promoted together throughout the year. As the SKUs in the same ad-block will be on or off promotion at the same time, we built our model at the segment — ad-block level:

The left-hand side of the model equation is the segment-level units sold, in week t and the right-hand side of the equation is the ad block level gross lifts for each ad block in geography i, in week t in that respective segment. The Adblock Lift is defined as the incremental unit sales generated by the promotion:

During promotion periods, we only observe sales while the promotion is active and do not know what the baseline sales would have been without the promotion. To predict baseline sales, we used our proprietary promotion forecasting models.

As CVS has over 8000 stores with local demand variations, we controlled for geographic effects. We interpreted the significant coefficients associated with each ad-block in the segment.

The Switching percentage, (1 — βn ) for each ad block n is interpreted as the percentage of demand transferred from non-promoted SKUs. Our study indicated that over 10% of gross lift from promotions in the CVS Front store showed demand transference with statistical significance.

We recommended an actionable insight by creating a new margin calculation by including the negative impact of demand transference on the yellow tag promotions to best inform our business partners in setting promotion strategy for the front store.

We deprioritized cannibalization via stockpiling effects as our literature review indicated that was more applicable to larger grocery style retailers as compared to the CVS Front Store convenience retail model. Additionally, the methodology has certain limitations as it does not capture the stockpiling between different sizes of the same item.

This is only the beginning for using modeling methods for cannibalization and halo in promotion decisions. We strongly recommend retailers re-reevaluate their promotion strategies considering the margin impact caused by promotion secondary effects.

We are grateful to our engineering partners and business partners for supporting this work.

References:

1. Ailawadi Kusum, Harlam Bari, César Jacques, Trounce David. (2006). Promotion Profitability for a Retailer: The Role of Promotion, Brand, Category, and Store Characteristics. Journal of Marketing Research — J MARKET RES-CHICAGO. 43. 518–535. 10.1509/jmkr.43.4.518.
2. Pesonen, Aleksi. (2018). Quantifying Promotion Effectiveness for a Retailer. Master’s Thesis, Aalto University School of Business, Spring 2018.

© 2024 CVS Health and/or one of its affiliates. All rights reserved.