A Three-Stage Model for Forecasting Store-Level Consumer Demand

Abstract

Brand distribution and promotional decisions require store-level forecasts for Brand sales. This may be accomplished by developing individual store-level models using store-level sales and promotional data. The disadvantage in this is that the number of models could run into hundreds. Also, accuracy may be diminished where forecasts need to be extended to stores outside the modeled sample and the sales are not homogenous across stores.

In recent years Mixed Effects Modeling has emerged as a very powerful predictive technique in situations where data heterogeneity limits the application of standard regression models. The advantages of these models are that they provide response effects at multiple levels of aggregation by pooling sub-groups within a level and also allow the estimation of a large number of parameters with relatively fewer observations by leveraging information across pooled groups using empirical and hierarchical Bayes' estimation.

We combine this technique within a three-stage model using store-level data for stores that sold the brand, to forecast sales in Stores that have not yet sold the brand. Using Mixed Effects models to account for store-level heterogeneity reduces the effect of aggregation bias, but the store-specific parameter estimates generated, cannot be effectively leveraged to forecast brand sales in un-modeled stores or channels. We first cluster the sample population of stores into homogenous groups based on similar demand and supply characteristics. A Mixed model is then identified in the second-stage with parameters estimated at the cluster and total levels.

Treating cluster-membership to have a multinomial distribution, we develop a multinomial logit model for cluster-membership prediction in the third-stage. This is used to assign new stores to existing clusters, and then develop forecasts for those stores using the cluster-level parameter estimates.

Keywords: Store-level forecast, cluster, segmentation, Multinomial Logit

JEL Classification: C22, C1