New-Product Demand Forecasting for Long-Lived Products

Abstract

We consider the problem of demand forecasting for long-lived products when a new product is introduced. This problem is challenging in part because sales data is limited or non-existant for the new product. In addition, a new product can change the demand for the existing products through cannibalization, so forecasts for existing products also need to be adjusted to take this into account. We propose a joint forecasting model for the demand of the new and existing products that combines a linear model, and a choice model over clusters of similar products. Clustering reduces the dimension of the choice model, which makes it easier to calibrate, and allows us to estimate cannibalization by the new product by quantifying the impact on demand when similar products were introduced in the past. Market shares are further adjusted by the linear model to account for potentially time-varying factors and product-level features that are lost in the clustering. The resulting model is easy to calibrate using an extension of the method by Theil (1969). In
out-of-sample tests using data from The Body Shop, methods like linear regression and random forest, which do not explicitly account for cannibalization, overestimate demand for existing products when a new product is introduced, while our proposed model is able to eliminate this bias.

Keywords: demand forecasting, new products, limited data, cannibalization