Strategic Planning: Thinking Big

Assortment planning begins with setting goals for each department and product category for unit sales, market share, revenue, and profit.² You’ll increase resources, including store space and purchase dollars, earmarked for some product categories, while decreasing them for others to reflect product, competitive, and fashion trends.

A forecast of future sales helps to decide which categories to expand and which to contract. Figure 2-1 shows Best Buy’s sales of digital and film cameras for 1995 to 2002. A logical forecast for 2003 would be a decrease in the sales of traditional cameras and an increase in sales for digitals, compared with sales in 2002.

FIGURE 2-1

Historical sales of film and digital cameras

Source: Kevin Freeland, chief operating officer, Advance Auto Parts, and former senior vice president for inventory management, Best Buy.

But figure 2-1 also shows the limitation of this approach. A forecast of 2001 film camera sales based on an extrapolation of historical sales would have erred high. A smarter forecast would have factored in intelligence from trade shows, conversations with vendors, observations of competitor moves, and reviews of new technology. With this information, a retailer identifies potential changes in sales for a product or category that might not be apparent from a straightforward extrapolation of sales history. In the case of cameras, these sorts of observations would’ve led you to predict the fast rise in the sales of digitals.

The method we’ll describe relies on recent sales history to identify profitable assortment changes and therefore is most appropriate for product segments where demand patterns change slowly. Many product segments, particularly in grocery and hard lines, satisfy this requirement, but some, such as fashion apparel, do not. Urban Outfitters’ CEO Glen Senk’s remark that “In fashion apparel, there is nothing as boring as last season’s hot seller,” clearly illustrates that for fashion products, history can lead one in exactly the wrong direction. But even in the volatile world of fashion apparel, attributes can help. Urban Outfitters has developed an impressive approach to dynamically adjusting their assortment within a season by having their buyers focus on current sales and identify which product attributes are trending upwards or downwards.

Borders, the bookseller, has an even more advanced approach to strategic assortment planning. Two brothers, Tom and Louis Borders, founded the company. Tom majored in English at the University of Michigan, which explains why they based their chain in Ann Arbor. Louis studied computer science at MIT, which enabled their chain to develop sophisticated information technology in its early days.

Borders segments its books into about nine thousand categories and allows each store to carry a different number of titles in each category.³ To choose the number of titles in each category for each store, it relies on a measure called relative sales per title (RST), which equals the sales in a category in a store over a specified period, divided by the number of titles carried in the category over the same period. If RST is high for a store category in a recent period, then Borders increases the number of titles in that category. If RST is low, it reduces them. It might, for example, divide the one thousand categories in a store into upper, middle, and lower thirds of RST values and then increase the number of titles carried in the upper third by 10 percent and reduce the lower third by 10 percent. (Borders’ process also takes seasonality into account, but that is outside the scope of this chapter.)

The philosophy behind Borders’ approach makes sense—put resources where they can earn the greatest return. Borders’ approach is on the right track, but we can do even better by understanding that as you analyze your company’s assortment, what matters is the marginal return on incremental resources, not the average return of resources already deployed.

Figure 2-2 shows revenue versus SKU count for two hypothetical product categories. The SKUs in each category are ranked by dollar sales rate. Then the revenue is plotted as SKUs are added, one by one, in rank order. Note that the first category has higher average sales per SKU, but just a few high sellers cause this; the weakest SKUs in the category barely sell and add little revenue. By contrast, the second category has lower average sales per SKU, but all of its SKUs are equally productive. It would therefore make sense to reduce the SKU count for category 1 and increase it for category 2. You can delete half the SKUs in category 1 with little loss of revenue. If you then use the resources freed to add a comparable number of SKUs to category 2, their sales rate will probably resemble that of the SKUs currently in the category.

FIGURE 2-2

Enhancing the Borders approach by considering the marginally performing SKU in each category

As you apply these insights in your company, you should evaluate categories by their marginal productivity—that is, you should target categories where the worst-performing SKUs have relatively low sales rates, and delete them to free space to add SKUs to categories where the worst performers have relatively high sales rates, indicating potential for growth.

Operational Planning: Getting Specific

Once a retailer decides how many SKUs to carry in a category, it faces the operational problem of picking particular ones. In flat panel TVs, for example, Best Buy might stock 82 different models. The number of potential SKUs, of course, is much larger, comprising eight diagonal widths, five screen types, seven resolution levels, and nine major vendors, for a total of 2,520 (that is, of 8 × 5 × 7 × 9). Company buyers select the specific SKUs, and they incorporate many factors into this decision, including ensuring that Best Buy offers products from a variety of vendors. That way, the chain can benefit from competition when negotiating on price, and its customers have more choices.

One challenge in assortment planning is including new products; buyers will have to consider many goods for which Best Buy has no sales experience. The lack of sales histories for these thwarts analytic planning. A buyer will have little, if any, data to use in making decisions.

Be aware that assortment decisions aren’t only about maximizing the profit on every single SKU. Though that’s the focus of this chapter, you might also select a SKU to help sell other goods. An obvious example would be a loss leader that attracts shoppers to your store. And low-priced offerings are not the only way of enticing customers. A broad array in a category, including high-end aspirational goods, can signal that your store is the best place to buy all of the products in that category. Bike shops, for example, will often sell custom-built frames by makers like Seven and Independent Fabrication, even though the vast majority of their sales come from off-the-rack offerings like Trek and Cannondale. The custom frames give shops cachet with bike aficionados, even if many of those customers choose not to pony up that much money for a bike.

Carrying a premium-priced offering can increase the sales of other pricey products by making them look like relative bargains. In a Publix supermarket, one of us noticed a bottle of wine selling for $264—a lofty price anywhere, much less in a grocery store. The natural question for the wine manager was how much of that wine he sold. His response: “None, but we sell a lot of this $67 bottle next to it because we carry the $264 bottle.”

Another reason to carry slow-moving SKUs is to please your best customers. Jim Halpin, former president of BJ’s, tells a story that illustrates this point.⁴ Soon after he became president of BJ’s, he got a sales report and noticed a dessert pie SKU at the bottom of the list: BJ’s was selling a whopping three a week. So he deleted the SKU. That was a no-brainer, he thought—until he got an angry call a few days later from one of his best customers asking why he’d stopped carrying his dessert pie. Those three a month mattered to this customer, and this customer mattered to BJ’s.

Grocery chains, which can obtain data on customers’ baskets of purchases through loyalty cards, are ideally positioned to protect themselves from this pitfall. Before they delete a SKU, they can assess the average size of the basket of goods purchased by customers who typically bought that SKU. If they notice that caviar buyers tend to purchase $500 worth of groceries every time they visit, they should keep the slow-selling fish eggs.

All Retailing Is Local

Traditionally, a retailer would use the same assortment in all of its stores, except perhaps for eliminating less important SKUs in smaller outlets. Recently, some firms have made great efforts to localize their assortments.⁵ They’ll group their stores into clusters with similar demographics or sales mixes and then apply their chain-level assortment planning process to each cluster.

Though it sounds promising—after all, what’s better than catering to your customers?—problems arise with this approach. It’s a ton of work. Manual planning must be done not just once for the chain but for each cluster. And despite all of the toil, localization may not lift sales. One home-fabrics retailer implemented the approach for a category that it thought would benefit, and achieved an 18 percent sales increase. Pleased with the results, it tried a second category, but saw no rise. That ended its enthusiasm for localization.

Another challenge is figuring out the right number of clusters to form and picking one of the many ways of forming them.

Politics complicates localization, just as it complicates many things. If you choose to localize, you also have to decide who gets to decide. That is, you have to figure out how to divide decision-making authority between your headquarters and stores. The most common approach is for headquarters personnel to dictate a single common assortment to be carried by all stores. But a few retailers—Bed Bath & Beyond is one—allow their store managers considerable authority in deciding which SKUs to carry. Usually, the corporate office mandates a portion of the SKUs, and store managers choose the rest for their locations from an approved list of options. Differing offerings in each store end up, in theory, tuned to the tastes of that location’s customers. Delegating a portion of the assortment decision to stores is decentralized localization, but it’s also possible to make all assortment decisions centrally and still localize the assortment. The Borders approach to assortment planning described previously is an excellent example of localized assortments created centrally.

Tasty Techniques in the Snack Cake Aisle

Now let’s turn to Little Debbie and her competitors in the snack cake aisle and use the example of planning an assortment for a convenience store chain to illustrate this approach. The primary input that we’ll use is store-SKU sales over the last six months of 2005. Table 2-1 provides an example of the data for one store. This retailer carried thirty-nine out of the sixty-eight possible brand-size-flavor combinations (two brands × two sizes × seventeen flavors = sixty-eight combinations) at this location. We can’t reveal the actual names of the two brands, so we’ll call them Yummy Cakes and Tiny Tina.

TABLE 2-1

July 1–December 31, 2005, snack cake sales in a convenience store^a

These thirty-nine SKUs generated revenue of $82,665 during the six-month period. A crucial question is whether a different set of thirty-nine SKUs might generate more sales in the future. (Prices varied slightly across the different flavors and the two brands within a size, but for simplicity, we’ll use the average prices for single serving and family size of $1.29 and $3.34.)

Not surprisingly, most retailers do better at identifying poorly performing SKUs and deleting them than at identifying promising new ones and adding them. As a result, many retailers follow a “gin rummy” strategy. They discard the worst-selling SKUs from their current hand of products and randomly draw from the deck of potential new offerings. They hope that, over time, they’ll identify a productive assortment.

But what if you could have an accurate sales forecast for any assortment your store might offer? Instead of guessing which products might sell, you could then choose the exact assortment that would maximize revenue. That’s what we’ll show you how to do now.

Our challenge in working with the convenience store chain was to use the data in table 2-1 to estimate the demand for brand-size flavors not offered. We started by estimating the shares of the four brand-sizes. We could have just used the sales shares shown in table 2-1 to estimate demand shares of the four brand-sizes. But these sales shares had been influenced by the assortment offered and thus didn’t represent true demand. The chain had previously offered Yummy Cakes in more single-serving flavors than it did Tiny Tina. That inflated the sales share of Yummy Cakes: in some instances, customers really wanted Tiny Tina in a particular flavor but had to settle for Yummy Cakes. We sidestepped this problem by focusing on the two flavors where the chain had offered all four brand-sizes—chocolate and cinnamon—and where sales thus represented true demand because no customers had been forced to substitute. Table 2-2 shows the calculation of demand shares for these two flavors. Notice that, as predicted, the demand share of Yummy Cakes is lower, and that of Tiny Tina is higher, than their sales shares in single serving. This shows the impact of the assortment on sales. Demand shares for family size also differ for the same reason.

TABLE 2-2

Calculating brand-size demand shares using the flavors where all four brand-sizes are offered

Next we used these brand-share estimates to estimate the demand for the SKUs not offered. How much, for example, would the store sell if it offered Yummy Cakes in family-size butter flavor? It’s tempting to use the 6 percent and 5 percent shares for Yummy Cakes and Tiny Tina in family size, observe that Yummy Cakes sells 20 percent more than Tiny Tina, and reason that since the store sold 331 in Tiny Tina family-size butter, it might expect to sell about 397 of Yummy Cakes.

But this logic ignores product substitution. If customers don’t find what they really want, they may buy another product instead. Thus the 331 units sold of Tiny Tina might have included customers who wanted Yummy Cakes but bought Tiny Tina when they couldn’t find their first choice. Studies of convenience store customers have shown that they place a much higher weight on flavor than brand in their purchase decisions for snack cakes and that they are unlikely to jump between family size and single serving. Someone seeking a midmorning snack doesn’t want to lug a box of Tastykakes or Hostess Ho Hos around in his backpack for the rest of the day. We assumed that the likelihood of substitution across flavors and sizes was zero, and focused on estimating the percentage of customers who substituted between brands within a size and flavor.

How can we estimate the likelihood of substitution between brands? Notice that for four of the flavors shown in table 2-1—butter, cheese, raspberry, and vanilla—the chain carried Tiny Tina in single and family sizes and Yummy Cakes in single size. In those flavors, it didn’t offer any Yummy Cakes in the family size. With this information, you can estimate the fraction of customers who substituted Tiny Tina in family size for Yummy Cakes in that size. Table 2-3 shows the calculation.

Here’s the logic behind table 2-3. The sales of Yummy Cakes and Tiny Tina in single serving represent true demand since the store offered both brands, and customers didn’t have to substitute. These sales totaled 10,915 (6,364 + 4,551), and, as we calculated in table 2-2, they represented 89 percent of total demand for these four flavors. This means that the total demand for the four flavors was 12,264 (10,915 / 0.89). Thus the demand for Yummy Cakes and Tiny Tina in family size can be calculated as 6 percent and 5 percent of this total value, which are the values 736 and 613 shown in table 2-3. If 613 was the true demand for Tiny Tina in family size, then you could estimate that 635 (1,248–613) customers preferred Yummy Cakes but substituted Tiny Tina when their first choice wasn’t available. This means that 86 percent (635 / 736) of the customers who wanted Yummy Cakes bought Tiny Tina instead. You thus can conclude that the likelihood of a customer substituting from Yummy Cakes to Tiny Tina in family size is 86 percent.

Here’s another way to think about estimating the rate of substitution. You know from the share estimations that family size represents 11 percent of demand. If everyone substituted from Yummy Cakes to Tiny Tina, then when Yummy Cakes was missing, family size would still represent 11 percent of sales. If no one substituted, then the 6 percent of sales to customers that prefer Yummy Cakes would be lost, and you’d only capture 94 percent of demand. The 5 percent of customers who buy Tiny Tina in family size would represent 5 / 94 = 5.3 percent of the total demand captured, and hence the share of family size would fall to 5.3 percent of sales. So family-size share when Yummy Cakes is missing will be between 5.3 percent and 11 percent of sales; closer to 5.3 percent indicates low substitution, while closer to 11 percent indicates high substitution. The share of family size is 1,248 units sold / 12,264 total units = 10.2 percent, very close to 11 percent and indicating a high level of substitution. In fact, 10.2 percent actual share / 11 percent maximum share = 93 percent, almost the same as the 86 percent substitution rate we calculated before, so the two viewpoints are similar.

TABLE 2-3

Calculation of the percentage of customers who will substitute from Yummy Cakes to Tiny Tina in family size

TABLE 2-4

Estimating substitution percentage from Yummy Cakes to Tiny Tina in single serving

Applying the same logic to the fudge flavor, you can estimate that the likelihood of substitution from Tiny Tina to Yummy Cakes in family size is 20 percent. Likewise, for chocolate chip, the likelihood of substituting from Tiny Tina to Yummy Cakes in single serving is 26 percent.

Estimating the substitution percentage from Yummy Cakes to Tiny Tina in single serving is trickier because there is no flavor in which Yummy Cakes single serving is the only missing brand-size. But table 2-4 shows how you can do this calculation, using sales data of apple-flavored cakes.

Thanks to the brand-size and substitution percentages for family size that you’ve already estimated, you know that the 60 units sold of Tiny Tina in family size correspond to 10.16 percent of total demand for apple (5 percent + 0.86 × 6 percent). You therefore can estimate total demand for apple as 591 (60 / 0.1016). You then compute the demand estimates shown in table 2-4 as brand-size shares multiplied by this total demand estimate. Once you know the estimated demand for Yummy Cakes and Tiny Tina in the single-serving size, you can compute the substitution demand for Tiny Tina and the substitution percentage from Yummy Cakes to Tiny Tina, as shown in table 2-4.

Table 2-5 summarizes the substitution frequencies we have calculated.

In our experience, most buyers guess that some customers will substitute if they can’t find what they want, but many of these buyers lack a way to estimate these values and factor those estimates into their assortment decisions. Table 2-5 shows that substitution rates can vary a lot; customers who buy Yummy Cakes in family size are not very loyal: 86 percent of them switched to Tiny Tina. This made sense to the convenience chain’s buyers when we presented our findings, as Yummy Cakes was strong in single serving but had only recently begun to offer family size. These sorts of differences in willingness to substitute have a big impact on assortment planning.

You can use the estimates of brand-size shares in table 2-2 and crossbrand substitution frequencies in table 2-5 to estimate the demand for all brand-size-flavors. Table 2-6 displays these estimates. For each flavor, you know total sales, given the assortment of brand-sizes offered in that flavor. You can also compute the fraction of potential demand captured, both from customers for whom a brand-size offered is their first choice and from customers for whom it is their second choice but who are willing to substitute. Witness the 7 percent share captured by buttercream. It results from the 6 percent of customers who made Tiny Tina family-size buttercream their first choice plus the 1 percent of customers who wanted Tiny Tina but substituted Yummy Cakes (20 percent of 5 percent equals 1 percent). You can then estimate total demand for a flavor, assuming the chain offered all four brand-sizes, as total sales divided by share captured. For buttercream, for example, the estimated total demand is 1,429 (100 / 0.07). You can then multiple total demand by brand-size shares to obtain estimates of SKU demand. The 62 percent share of Yummy Cakes in single serve, for example, implies a demand estimate for Yummy Cakes single-serving buttercream of 886 (.62 × 1,429).

TABLE 2-5

Estimated substitution percentages

TABLE 2-6

Estimated demand for all brand-size-flavors

Table 2-6 shows demand estimates for each brand-size. If the convenience store offered both brands in a flavor and size, we used sales as our demand estimate.⁷ In this case, customers didn’t need to substitute, so sales corresponded to actual demand. We computed other demand estimates as the shares times the estimated total demand for a particular flavor.

A key question in improving this assortment is which flavors customers want most. A traditional way of answering this is to look at sales rank. But the total demand estimate column in table 2-6 gives a different—and we’d argue, better—answer. Consider the flavor vanilla/chocolate. With total sales of 2,162, this flavor ranks twelfth, near the bottom, making it a candidate for deletion. But if you take our estimates into account, its total demand of 5,759 puts it in third place, near the top of the list.

Why such a big difference? Vanilla/chocolate is offered only in single serving, and is the least popular brand for single serve to boot, and this depresses its sales relative to other flavors. Despite this handicap, if you look closely, customers love this flavor. Its sales of 2,162 ranks second only to chocolate in Yummy Cakes single serving.

This is one more example that the assortment a retailer offers distorts sales away from true demand. If you want to discern true demand, you must tease out customer insights latent in sales data.⁸

Finding an Optimal Assortment—Let’s Get Greedy

To use the demand estimates in table 2-6 to find an optimized assortment, you employ a technique called a greedy rule. For sake of concreteness, suppose you’re seeking a revenue-maximizing assortment for this store, subject to the constraint that the assortment contain no more than thirty-nine SKUs. (You could just as easily apply this approach to the chain or a cluster of stores, and you can use it to maximize unit sales or dollar gross margin.) The greedy rule chooses thirty-nine SKUs sequentially, choosing as the first SKU the one that would maximize revenue if it were the only one in the assortment. Note that the revenue of a SKU is its primary demand revenue plus any revenue from substitution demand from other SKUs. Thus the first SKU that you choose might not have the highest demand; it might be many customers’ second choice. You choose the second SKU to maximize the increase in revenue given the first SKU that you picked. You continue to pick SKUs that maximize the increase in revenue from their selection until you’ve selected thirty-nine of them.

Table 2-7 shows the resulting new assortment. We estimated that this assortment would capture $99,757 in revenue, a 21 percent increase over the current assortment revenue of $82,665. Note the other ways in which the assortment has changed. First of all, our assortment drops Yummy Cakes in family size in all flavors. Recall that 86 percent of the customers who preferred Yummy Cakes in family size were willing to switch to Tiny Tina. Thus the revenue lost by dropping Yummy Cakes is only 0.11 times its 6 percent share, or .66 percent. Our assortment also drops cheese and apple, the least popular flavors. The revenue lost is again small because few customers preferred these flavors. This frees room in the assortment to add additional brand-sizes in the more popular flavors, which generates the 20 percent sales lift.

Figure 2-3 shows the results of applying the greedy rule to the demand estimates for all of the stores of this chain. Figure 2-3 gives revenue as a function of SKU count for the two extremes of assortment localization: no localization, which corresponds to one assortment for the chain; and maximum localization, which corresponds to a unique assortment for each store. To get a single chain assortment, we applied the greedy rule to sequentially choose SKUs guided by the total revenue across all stores for a SKU choice.

TABLE 2-7

Revised assortment–shaded cells are SKUs carried in the new assortment

These curves are a by-product of the greedy rule, which adds SKUs one by one, and every point on each curve is backed up by a specific assortment that is estimated to produce the revenue shown. The figure also shows a point that gives the revenue of the thirty-nine-SKU assortment currently used by the retailer. The revenue increase from reassorting at the chain level is 28 percent; maximum localization can give another 13 percent.

FIGURE 2-3

Reassorting is estimated to add 41 % to revenue

This retailer, like many, liked the revenue increase from localization but was reluctant to have a unique assortment for each store because of the high administrative costs. Its buyers asked about having five different assortments for the chain. To do that, you’d begin with the store-specific assortments and cluster these using the following approach, until you’d reduced the number of assortments to five. First, you’d choose two stores to combine. Your criterion for clustering them could be either least loss in revenue or greatest overlap between the two store-specific assortments. You’d then apply the greedy rule to this two-store cluster. You’d continue to pair up stores into clusters until you had exactly five assortments and every store had been assigned to one of these five assortments. The good news is that the localization gain from cutting back to just five assortments only dropped from 13 percent to 12.5 percent, so most of the gain was achieved with just five well-chosen assortments.

Tire Assortments—Where the Rubber Meets the Road

We also applied our approach to planning the assortment for a national tire retailer. When buying tires, people care about brand, mileage warranty, and size. This retailer offered several nationally known brands, including BFGoodrich, Goodyear, and Michelin, which we treated as one category called National. It also offered three house brands of decreasing quality, which we denote as House 1, House 2, and House 3, with House 1 being the best and costliest. Manufacturers provided a couple dozen distinct mileage warranties on their tires. Some of them varied only slightly, so the staff at the retailer believed that consumers viewed them as equivalents. For our analysis, we grouped the mileage warranties into three levels: high, medium, and low.

It doesn’t make sense to sell an expensive tire with a low mileage warranty, or vice versa. The retailer therefore offered the following six brand-warranty combinations, which it believed to be logical pairings: National-High (NH), National-Medium (NM), House 1-High (H1H), House 2-High (H2H), House 2-Medium (H2M), and House 3-Low (H3L). The retailer could choose from 64 tire sizes, resulting in 384 (64 × 6) possible SKUs. It offered 122 of the 384 possibilities. That raised an obvious question: might any of the 262 tires not offered generate enough demand to warrant selling them?

As with many products, a customer might substitute another tire if she didn’t find her first choice. The retailer didn’t have sufficient data for us to estimate all possible substitution probabilities. We relied instead on management to tell us the most likely substitution patterns, and focused on those. Substitution across sizes didn’t happen. Table 2-8 depicts the likelihood of substitution from a given brand-warranty level listed in a row to another listed over a column, as estimated by the vice president of the tire category. His analysis suggested the need to estimate three substitution frequencies: most likely, likely, and somewhat likely.

When creating a mathematical model of a problem, you face a choice between building a simple, tractable model with a few parameters and a more accurate, but complex, one with many parameters. Here, we could have used a complex model of substitution that allowed for a different substitution frequency from any brand-warranty to any other brand-warranty, which would have led to thirty substitution parameters (six brand-warranties, each of which could substitute to five other brand-warranties). A more complex model often seems more appealing at first; after all, why wouldn’t you want as much precision as possible? Unfortunately, that kind of complexity contradicts a statistical principle called the law of large numbers (also known as the law of averages), which says that you get a better estimate when you average a large number of values. Imagine, for example, that you wanted to predict the chances a coin would come up heads by flipping it a number of times. If you flipped it ten times, you might see six heads and estimate a 60 percent chance of heads. But this six is likely due to chance. If you flipped it one hundred times, and it was a fair coin, you’d likely see, say, forty-nine or fiftyone heads and estimate a likelihood of about 50 percent. Extend your experiment to a million flips, and you’re likely to get a very tired wrist—and almost exactly 50 percent heads.

TABLE 2-8

Management’s estimate of the most likely substitution patterns

Now return to our example of modeling substitution. The more complex model had thirty substitution parameters; the simpler model, three. With a fixed amount of data, that’s ten times more data per parameter in the simple model than in the more complex one, which is like flipping the coin one hundred times rather than ten. So the increased accuracy you get with more data per parameter often favors simplicity, especially if, as in this case, the retailer believes that substitution rates are similar between various options.

TABLE 2-9

Input data for a store is sales during the last six months of 2004 by size-brand-warranty level for SKUs that were offered. A subset of sizes is shown.

Table 2-9 shows sales data for a given store for a subset of sizes. We applied the analysis approach described above to this sales data for each store to estimate the six brand-warranty shares, sixty-four size shares, and the three substitution frequencies. Table 2-10 shows the average across all stores of the estimated brand-warranty demand shares and substitution probabilities. Figure 2-4 is based on store-specific estimates and shows that the House 3-Low share in a specific store is correlated with the income level in the store’s zip code, which makes sense. (You should always seek confirmation for the conclusions of your analyses of sales history, and store demographic data is a wonderful place to start.)

FIGURE 2-4

Share of the low-price tire correlates with income

Table 2-10 also gives the sales shares for comparison with the demand-share estimates. Notice that the demand-share estimate for House 3-Low greatly exceeds the sales share. For House 2-Medium, it falls short of the sales share. This retailer offered House 3-Low in only nine of sixty-four sizes, so customers had few purchase opportunities. But as shown in table 2-11, when the retailer did offer House 3-Low, customers strongly preferred it over the next-highest price selection: House 2-Medium.

The retailer preferred to sell the higher-priced House 2-Medium, believing that its sales staff could persuade customers to trade up. The substitution estimate of 45 percent shows that many customers did trade up, and this explains the high sales share for House 2-Medium relative to its demand share. But the 55 percent of the 61 percent of customers preferring House 3-Low suggests that the retailer lost more than a third of demand due to the meager amount of House 3-Low in its assortment. It could have increased sales by adding House 3-Low in more sizes.

Figure 2-5 shows the results of our analysis. This retailer carried 105 SKUs per store in its current assortment. In the optimized assortment, it would replace 47 of these. The biggest contributor to the 36 percent increase in revenue would be carrying the House 3-Low tire-warranty combination in more sizes.

FIGURE 2-5

Reassorting is estimated to add 36 percent to revenue

When Does Localization Pay?

We sought to understand why localization had a significantly lower impact on tires than snack cakes (6 percent versus 13 percent revenue impact). We first thought that the store-to-store variation in demand patterns for snack cakes was greater than for tires, but found that the variation was nearly equal for the two cases, and in fact slightly higher for tires. We then realized that the percentage of maximum demand captured by the single chain optimal assortment would also affect the benefits of localization. If a retailer carried such a broad assortment that the single chain assortment captured nearly all of the potential demand, then there would be very little improvement potential left for localization. The snack cakes assortment offers 40 percent of the potential SKUs, while the tire assortment offers 27 percent of potential SKUs, so it might seem that the snack cakes assortment is broader. However, as shown by the curves in figure 2-6 (which show cumulative percentage of maximum revenue captured by a given percentage of the SKUs in the chain optimal assortment, where the SKUs have been sorted in decreasing order of contribution to revenue), the 27 percent of tire SKUs offered capture 82 percent of maximum revenue, whereas the 40 percent of snack cake SKUs offered capture just 68 percent of maximum potential revenue. Thus for tires, only 18 percent of maximum revenue is left to be captured by localization, whereas 32 percent is left to be captured by localization in the snack cakes case, nearly twice as much.

FIGURE 2-6

Cumulative revenue share versus percentage of SKUs

Thus the benefits of localization depend not only on the amount of demand variation between stores but on the percentage of demand captured by the single chain optimal assortment. This is an important principle for retailers to consider in choosing which categories to localize. The natural inclination is for a retailer to localize its most important categories, but we have seen cases where the retailer carried such a broad assortment in its most important categories that there was little to be gained from localization.

While localizing assortments by store added only 6 percent to revenue for the tire assortment, there were meaningful differences between stores. Table 2-12 shows demand estimates for two stores, in Nashville and New Brunswick, and gives median incomes for the zip codes in which these stores are located. New Brunswick customers, with 70 percent higher income than the Nashville customers, buy more of the most expensive national brand and less of the least expensive House 3 brand, and this is reflected in a higher demand share for National-High and a lower demand share for House 3-Low in the New Brunswick store.

These differences in demand estimates influence the store specific assortments for the two stores. The Nashville assortment contains twenty-one SKUs that are not carried in the New Brunswick store and, correspondingly, the New Brunswick store has twenty-one SKUs not carried in Nashville. Ten of the Nashville unique SKUs are the lowest-priced House 3-Low tires and another five are the next lowest-priced House 2-Medium. Conversely, twelve of the New Brunswick unique SKUs are the highest-priced National-High.

TABLE 2-12

Demand shares and income for two stores

These SKU differences didn’t arise because we explicitly considered customers’ income levels. Our results stemmed purely from the demand data. Higher-priced tires ended up being carried in the wealthier zip code because higher incomes led to their purchase. And that, in turn, led to higher demand estimates for the costlier tires, which caused our revenue-maximizing greedy rule to choose more of those tires. Even so, the apparent causal relationship that emerged comports with common sense. It’s a real-world check on the reasonableness of our results.

Implementation Produced a 6 Percent Revenue Increase

The retailer implemented a portion of these recommendations by adding eleven of the SKUs on our list of forty-seven and deleting eleven of the SKUs that we recommended dropping. We estimated these changes would increase revenue by 13 percent. In fact, the retailer got a revenue increase of 6 percent. The lower result stemmed from changes in demand patterns from the sales-history period to the implementation. Even so, a 6 percent improvement is large relative to what retailers typically achieve through enhancements to existing stores. In its annual report, for example, Canadian Tire reports achieving a 3 percent to 4 percent annual revenue increase in existing stores during 2005 through 2007 and is targeting the same increase through 2012.⁹