Putting the Right Products on the Shelves

Robert Rooderkerk creates a methodology for optimizing retail assortments.

If you think the selection of cereal or laundry detergent at your regular supermarket is overwhelming, consider this: the store manager picked them from dozens if not hundreds of options. These are not insignificant choices. Studies have shown that a retailer’s assortment is a key differentiator from its competition and plays an important role in sales. Stores use a variety of methods to decide what to stock; some use software programs, others outsource the job to product manufacturers, and many just go on gut feeling. Each method has its weaknesses, but in general they oversimplify an incredibly complex problem, using restrictive assumptions and incomplete data.

Robert Rooderkerk, a visiting assistant professor of business administration at Tuck, and colleagues Harald van Heerde from Massey University and Tammo Bijmolt from the University of Groningen, have found a better way. In their paper, “Optimizing Retail Assortments,” which was recently published and highlighted in the journal, Marketing Science, they developed a scalable assortment optimization model that can describe consumer demand and predict how that demand would change given different product assortments.

The essential challenge of product assortment is a version of the “knapsack problem.” Just as a backpacker must decide what to bring on a camping trip based on the utility of the items and their size and weight, a store must choose what to stock based on the products’ profitability and how much shelf space they take up. What complicates things is that demand for a given product is not fixed; it varies according to the availability of similar products nearby, and with in-store marketing efforts such as price promotion, and amount of shelf space.

To account for these and other parameters, but avoid the information overload of innumerable assortment variations, Rooderkerk and colleagues built an attribute-based model. Instead of directly comparing products themselves, it compares them as bundles of attributes such as brand, size, and fragrance. “The goal was to have a maximum understanding of real-world complexity with a minimum of moving parts,” Rooderkerk said. This model was designed to explain consumer demand given a particular product assortment.

The second goal was to use the demand model to optimize the product assortment. For this, Rooderkerk and colleagues needed a way to efficiently navigate the enormous amount of potential selections. They decided to use a very large neighborhood search heuristic, which would allow them to test how substituting a product with another product in its neighborhood would affect consumer demand. For example, if a store has an assortment of soda, the model could replace the 12-ounce Coke Classic with a 20-ounce Pepsi and determine the change in demand and profitability. The authors developed another neighborhood search heuristic for price optimization that works in a similar way.

They applied this model to three years of weekly store-level scanner data on 61 different laundry detergents sold at a French supermarket chain. The results were encouraging. Using the model to optimize just the assortment caused the expected profits to increase by 37 percent. The price optimization alone led to a profit increase of 7.9 percent. Jointly, the optimization model resulted in a 43.7-percent rise in profits, which is significant.

“The next step should be to test the model more,” Rooderkerk said, “and it would be great to see it implemented to establish how well it actually works.”