Forecasting Repeat Sales at CDNOW: A Case Study

We describe a modeling exercise, conducted in conjunction with the online music retailer CDNOW, where the goal was to develop a simple stochastic model of buyer behavior capable of generating a medium-term forecast of aggregate CD purchasing by a cohort of new customers. Weekly sales are modeled using a finite mixture of beta-geometric distributions with a separate time-varying component to capture nonstationarity in repeat buying. The resulting model can easily be implemented within a standard spreadsheet environment (e.g., Microsoft Excel). We demonstrate that the model does a good job at describing the underlying sales patterns, as well as generating an excellent medium-term forecast.