OPERATIONS RESEARCH (ONLINE FORUM COMMENTARY)Vol. 57, No. 1, May–June 2009, online commentaryissn 0030-364X |eissn 1526-5463 |09

INFORMS

Determining Optimal Assortment: ForecastingDemand of New Products and Estimating

Substitution PatternComment on “Rocket Science Retailing” by Marshall Fisher

Juin-Kuan ChongAssociate Professor of Marketing

NUS Business School

National University of Singapore

Singapore bizcjk@nus.edu.sg

Tek-Hua HoWilliam Halford Jr. Family Professor of Marketing

Haas School of Business

University of California Berkeley

Berkeley, CA USA hoteck@haas.berkeley.edu

To deliver the right products to the right customers profitably requires a fundamental shift inretail decision making from art to science; and from one that is based on human intuition to one thatis driven by customer data. Marshall Fisher and his collaborators pioneered this transformationby developing many rigorous and practical statistical and operations research methods for theretailing industry. This “Rocket Science” approach brings retail decision making to a whole newlevel of efficiency and sophistication.

Retail decision making begins with assortment planning and assortment drives pricing and inven-tory sizing decisions. It is thus apt for Fisher to begin his lecture on assortment planning. Thiscommentary focuses on the excellent works of Fisher and collaborators in assortment planning andaims to achieve 2 goals. First, we will describe other research for demand forecasting (includingnew products that have new attribute levels and lost sale due to imperfect demand substitution)that may be better suited in some industries (e.g., frequently purchased consumer products). Sec-ond, we highlight the importance of using alternative sources of data to validate their method ofestimating demand substitution, which uses managerial judgment. Our suggestions and approachesmay be added to the existing toolbox for effective retail decision making.

The key to determining an optimal assortment at a store lies in the correct characterization ofdemand for each product and the substitution patterns across products. This characterization ischallenging because there are often a large number of products in a category and an even largernumber of substitution patterns. For example, if there are N products at a store, one must estimate(N*(N-1))/2 substitution patterns even with the assumption of symmetry in substitution. Theestimation complexity for SKU demand can be substantially reduced if one represents products

1

: Commentary on Marshall Fisher2 Operations Research (Online Forum Commentary) 57(1), 2009 INFORMS

as a bundle of attribute levels (e.g., brand, warranty, and size). By estimating demands for eachattribute level that SKU has, one can then derive a SKU demand from these attribute level demandsby a multiplicative model (see Table 4 in Fisher, 2009). See Fader and Hardie (1996) and Ho andChong (2003) for an alternative attribute-level modeling approach that uses panel level data.

Let us use an example to demonstrate how this reduction in complexity works. If there are 37brands, 9 sizes, and 145 flavors in an ice-cream category, there are a total of 37 x 9 x 145 =48,285 possible SKUs. Instead of estimating 48,285 product demands, one only needs to estimate37+9+145 = 191 attribute level values. We further reduced the dimensionality burden by replacingthe attribute level values with behavioral equations that captured historical demand and substitu-tion patterns (see Ho and Chong, 2003). Our approach requires only 6 parameter estimates for eachattribute and hence a total of 18 parameter values for a product category that has 3 attributes.Despite this parsimony, our approach fits the data well.

There is however a barrier in implementing the Fader and Hardie (1996)’s and Ho and Chong(2003)’s approaches. Both approaches use customer panel level data which may not be possible insome product categories either because such data cannot be captured or because the inter-purchasetime is long. However, the key advantage of both approaches is that it allows for correlation indemand at both the product and attribute levels by allowing their values to evolve as new purchasesare observed. (Fisher’s approach allows for correlation at the product but not at the attributelevel since it invokes the independence assumption in deriving SKU demand from attribute leveldemands.) The correlation at the attribute level can sometimes be important. For example, Marymay prefer chocolate flavor but only from Ben & Jerry’s (in this case, demand are correlated acrossbrand and flavor). In addition, our approach allows one to forecast demand for a new product(with attribute levels that were previously not available in the category). For example, a store mayintroduce a new flavor to the ice-cream category. Our approach automatically imputes a startingvalue for a new attribute level and uses it to forecast demand for any new product with that specificattribute level. (The starting values capture the customer’s response to past introduction of newattribute levels). As more purchase data come in, the approach will revise the starting value andproduce more accurate forecasts of the new attribute level value.

Fisher and colleagues simplify the number of substitution pairs by incorporating managerialjudgment. Store managers were asked to rule out unlikely substitution patterns and classify pos-sible patterns into distinct classes of substitution probabilities. A potential problem of managerialjudgment is that store managers may exhibit systematic biases in classifying these substitutionpatterns. Hence it may be important to validate human judgments using customer level data.Specifically, one may use the consideration set formed by customers to validate. Consideration setis the set of products that customers consider before their choice. One can determine considerationset by eliciting it directly from customers via a survey or estimating it indirectly using a panel dataset. In an indirect estimation, a consideration set is frequently defined as the set of products thatdeliver utilities above an empirically determined threshold (see Roberts and Lattin, 1991). Thus,only products that are above the threshold are included for substitution.

As long as a customer who wishes to purchase a product cannot find a perfect product substitute(i.e., the substitution probability is strictly less than 1), a sale may be lost. Under the Fisher’sapproach, lost sale is defined as 1 - substitution probability summed over all possible substitutes.An alternative approach to estimate lost sale is to explicitly incorporate a purchase incidencemodel, commonly used in marketing. Chong, Ho and Tang (2001) develop such a model and usethe assortment and household’s expected inventory level as predictors for the purchase incidenceprobability. Lost sale is then estimated to be the difference between a target assortment and theactual assortment. This alternative method may work better with a panel level data set.

References

: Commentary on Marshall FisherOperations Research (Online Forum Commentary) 57(1), 2009 INFORMS 3

Chong, Juin-Kuan, Teck-Hua Ho and Chris Tang (2001), “A Modeling Framework for CategoryAssortment Planning,” Manufacturing & Service Operations Management, 3(3), pp. 191-210.Fader, Peter S. and Bruce G. S. Hardie (1996), “Modeling Consumer Choice among SKUs,” Journalof Marketing Research, 33(4), pp. 442-452.Fisher, Marshall (2009), “Rocket Science Retailing: The 2006 Philip McCord Morse Lecture,”Operations Research, 57(3), pp. 527-540.Ho, Teck-Hua and Juin-Kuan Chong (2003), “A Parsimonious Model of Stockkeeping-UnitChoice,”‘ Journal of Marketing Research, 40(3), pp. 351-365.Roberts, J. H. and James M. Lattin (1991), “Development and Testing of A Model of ConsiderationSet Composition,” Journal of Marketing Research, 28(4), pp. 429-440.