Part II: How do I model cannibalization between similar skus, items or products?
This is a follow-up to our blog Demand Forecasting: Part I: How much do I order given point-of-sales, promotion history and seasonality patterns. In this write-up we discuss ways to model cannibalization – the effect of promotion of one item on a substitute item.
The way we have defined cannibalization in itself suggests two parts to deal with problem – identifying substitute items and modeling the effect of the promotion. Cannibalization is a fairly advanced topic in Demand Forecasting that entailed a blog of its own. Moreover, in discussions with many fellow practitioners and retailers, we have found wide-ranging interest on the topic.
In our literature review, we have come across articles which mostly discuss the second stage – modeling the effect of promotion, of cannibalization, assuming that the identification of similar items is already taken care of. In our experience on working with our clients, we did not encounter any such file available off-the-shelf for modeling. Given the large number of items a retailer carries, the task of arranging “likely-to-be” cannibalized items is hardly trivial. A good first step, in our experience, is to use some variation of machine learning based text analytics (on product description) or curve recognition (on sales/demand curves) algorithms to group similar items within a product category. However, such algorithms identify similar items, not substitute ones. Using manual intervention at this stage becomes somewhat easier and also very necessary. Let’s consider this through an example of likely cannibalization between Coke and Pepsi. A can of Coke on sale will certainly cannibalize a can of Pepsi and maybe even a 20 oz. bottle or a 6-pack set of Coke in a convenience store setting. However, the promotion on the can of Coke may not cannibalize a large bottle of Coke in the convenience store. Such subtleties are highly subjective and are best resolved by merchandise managers or assortment planners. We propose another method in identifying substitutes in our next blog in the series – Demand Forecasting Part III: Important considerations in Forecasting.
After the items are grouped, we then model the effect of promotion on the sales. The regression parameters, which represent uplifts from promotional activities, from forecasting process could be used as a guide to develop a promotion score. Greater the score more will be the effect of cannibalization. Once again, this step will best be implemented with close collaboration with marketers to maintain a sense-check of whether the regression parameters reflect their experience.
Lastly, we compare the forecast with and without cannibalization to analyze the improvements. We were able to attain marginal improvement in forecast using this approach without any intervention from merchandise planners due to time constraints.
By: Sagar Kalyankar