A Parsimonious Model of SKU Choice: Familiarity-based Reinforcement and Response Sensitivity

We develop a highly parsimonious model to describe and predict stock-keeping unit (SKU) choice in any frequently-bought consumer packaged good category. The model has 20 parameters for a category that has 3 salient attributes, and in general it has 8 + 4 ¢ k parameters for a k-attribute product category. This parsimony is accomplished by neither discarding part of the data nor aggregating the level of analysis. Unlike previous choice models, the number of parameters is neither a function of the number of SKUs nor a function of the number of attribute levels in the category. Consequently, it is particularly useful for modeling product choice in categories that have many SKUs and attribute levels. The model assumes that a product consists of multiple attributes and the consumer derives utility from both consuming and shopping for each attribute. While consumption occurs only for the chosen attribute level, shopping experience applies to all attribute levels. The consumer is also allowed to respond di®erently to a product's marketing mix activities over time. The model incorporates three key features: 1. The consumer's marginal utility from consuming an attribute level depends on her level of familiarity with it. This allows us to capture potential satiation with a familiar attribute level that has been consumed previously. 2. The consumer accumulates a shopping experience, which also depends on attribute-level familiarity. If shopping experience increases with familiarity, the consumer retrieves more familiar attribute levels more readily during shopping and chooses products based on memory cues. Both the attribute-level product satiation and shopping experience provide a natural way to model variety-seeking behavior commonly found in these product categories. Our model reveals which product attribute becomes more readily satiated and predicts that the consumer is more likely to switch to an attribute level that has a stronger shopping experience. 3. The consumer's response sensitivity to a product's marketing mix activities depends on her familiarity with the product. This allows the consumer to respond di®erently to di®erent SKUs's marketing mix activities during her purchase history. Using an extensive panel-level data set of 10 food and 6 non-food categories, we benchmark our model with that of Fader and Hardie (1996) (FH), which outperforms Guadagni and Little (1983)'s model. The empirical results are summarized as follows: 1. Our model on average ¯ts 5.16% better in-sample and predicts 6.50% better out-of-sample in log-likelihood than the FH model. Moreover, this superior performance uses only one-¯fth …