Predicting Under- and Overperforming SKUs within the Distribution–Market Share Relationship

Abstract This research presents a retail analytics application which uses machine learning (ML) to identify and predict under- and overperforming consumer packaged goods (CPGs) using retail scanner data. Essential to measuring market performance at the SKU level is the relationship between distribution and market share (the velocity curve). We validate that ML can reproduce the velocity curve, and ML is further used to predict underperforming, in-line performing, and overperforming SKUs relative to the velocity curve, based on a range of variables (SKU features) at a point in time. Our ML approach can correctly predict 83% of SKUs as under-, in-line-, or overperforming based on their characteristics. The research analyzes 9,321 SKUs of 2,565 brands across seven product categories of CPGs which were sold in 8,117 stores from 49 different retail chains of five different retail channels located in the US states of California, New York, Texas, and Wisconsin. The retail stores comprise convenience stores, drug stores, food stores, liquor stores, and mass merchandise retail stores. The data is Nielsen retail store scanner data for the calendar year 2014. The relationship between distribution and market share is a market-wide proxy for the ratio of relative sales in a category to, for example, aggregate shelf space, a key retail productivity metric. We further find indications that the distribution of SKUs across different store sizes, the stores’ category specialization, the line length of the brands, the overall performance of the parent brand, and sales consistency are the most important characteristics for the prediction of market share performance beyond the velocity curve. The methods and results presented will help CPG marketers (suppliers and retailers) understand which SKUs are under-, in-line-, or overperforming and the potential factors contributing to that performance. Optimizing assortments and portfolios is essential to decrease failure rates of individual SKUs. ML approaches can evolve to complementary support tools for such management problems.