This blog post is part of the HBR Online Forum The Future of Retail.
Retailers periodically update their product assortments, deleting slow sellers and adding new products in response to shifts in consumer demand or to accommodate new offerings from suppliers. Assortment-planning processes vary greatly across retailers and product segments but have one thing in common: They rely too much on human judgment and not enough on hard data that might allow a retailer to predict how customers will react to a change in the assortment.
This is the indisputable finding of research (pdf) that Ramnath Vaidyanathan of McGill University and I have conducted. Moreover, techniques that we have developed with several retailers over the last few years show that analytics are providing retailers a tremendous opportunity to improve revenues and profits.
Retailers who rely heavily on human judgment to make assortment decisions are flying blind, which leads to tales of woe like these:
- Walmart introduced Project Impact in 2008, an effort to 'declutter' stores by removing 15% of the SKUs they carried. It saw an immediate decline in sales and eventually had to roll back most of the changes.
- A grocery retailer deleted 20% of its dry-grocery SKUs to allow for expansion of fresh product offering. Sales declined 40% and the retailer is in bankruptcy. While all of the deleted SKUs had low sales, when customers couldn't find them, they elected to shop elsewhere.
- A retailer of items for the home sought to localize its assortments by store. Out of the 35 categories it carried, it chose fashion bedding, designed localized assortments for five store clusters and was thrilled to see an 18% increase in revenues. It next applied the same process to fashion bath, got no revenue lift, and abandoned its localization effort.
These examples illustrate the need to answer several questions when revising assortments:
- How will sales change if we increase or decrease the number of products carried in an assortment?
- If customers don't find their ideal product, what is the likelihood they will buy a substitute product?
- What's the likely benefit of localizing a category? How many store clusters should we have? What are the right metrics to use in clustering stores?
- What are the likely sales of items we are considering adding to our assortment?
The technique that Ramnath Vaidyanathan and I developed answers these questions. We identify a few attributes for each SKU that are meaningful for customers, use sales of existing SKUs to estimate the demand for attribute levels, and then use the estimates to forecast the demand for any combination of attributes, including those that correspond to new products that might be added to the assortment. We have made this approach formulaic so you can sic a computer on your sales data.
This technique lets you discover new products that have a high chance of selling well. For example, if an auto-parts retailer sees that one of its stores sells many parts for Hondas Accords and lots of brake pads, but doesn't carry pads for Accords, then it would seem like a good idea for the store to add Accord brake pads to its assortment.
For food products, the attributes could be flavor, brand, and package size. For flat panel TVs, they could be screen size, screen types, resolutions, and vendor. And across many product segments, price/value is a relevant attribute.
Our approach also challenges common practices like this: A retailer thinks customers don't want to buy product type X, so offers a limited amount of it and thus doesn't sell much of it, thereby confirming its assumption that customers didn't want X. But that may not be the case. In a study of tires that we conducted, where the attributes were size, brand, and mileage warranty, the lowest-priced brand had only a 10% share of the retailer's sales but the retailer only offered it in a limited number of sizes (nine). We discovered that in those nine sizes, the brand outsold the next-highest-price brand 40 to 1 and had a 61% share.
This retailer offered a limited selection of the cheapest tire because it thought it could get customers to trade up to a higher-priced brand, and the data showed that 45% of the time it could. But the 55% of the time it couldn't, times the 61% share of that category, showed it was losing a third of potential sales. We used our estimates to revise the tire retailer's assortments. The result: a 5.8% revenue increase.
A study done for another auto-parts retailer focused on appearance chemicals, a category that includes products for washing and waxing cars, shining tires, and polishing and protecting glass. This retailer was eager to understand how demand patterns differed across stores but felt it would not be feasible to have more than five different assortments corresponding to five store clusters. We showed that most of the gain could be achieved with just two store clusters. The interesting pattern we observed was that appearance chemicals applied to tires sold many times better in stores that had an urban/bilingual demography.
As these examples illustrate, the era of analytics has arrived in retailing. While there will always be room for intuition, it should be tested by analyzing the data.