Combining Parametric and Non-parametric Methods for Predicting Consumer Choice Using Supermarket Scanner Data

It is important for national chain stores to be able to customize their prices in individual stores to adapt to the neighborhood demand. A lot of stores collect scanner data which can be used to determine the price distribution over products which optimizes profits, and yet often this data resource is under-utilized. In previous research both parametric (such as linear regression) and non-parametric (such as artificial neural networks) models have been successfully used for prediction. We propose to approach this problem from a different angle and using both parametric and non-parametric methods to investigate ways of combining the lower-level models into a high-level model so as to achieve a better predictor than either of the simpler methods by itself. We will work with the data provided by a supermarket chain which would like to be able to price more strategically the various products in its local stores. The scanner data is for the category Chilled Juices and consists of store-level weekly reports of prices and quantities sold for the 14 products in the category. The data was assembled for two years from 100 individual stores of the supermarket chain. Our research will produce a method able to predict for a given category of products at a store level the consumer demand of the products from their prices. Previous research indicates that micro-marketing pricing strategies can increase gross profit margins by 4% to 10%, which after administrative and operating costs are subtracted would translate into an increase of operating profit margins by 33% to 83%. This research will provide a valuable tool for marketers concerned with predicting consumer choice. The new models will also generalize to other Machine Learning applications where both methods are currently used.