The most recent International Journal of Forecasting includes a paper titled “Forecasting from other’s experience: Bayesian estimation of the generalized Bass model.” by Ramirez-Hassan and Montoya-Blandon. The Bass model is a mature diffusion model to replicate and forecast sales using few parameters (innovation or early-adopter coefficient, imitation coefficient, and short-run potential market size). The generalized Bass model (GBM) adds terms to represent market effort which includes growth rate of price, advertisement, and other variables. In the paper the authors examine Bayesian estimation for the parameters of the GBM by looking at data for related products. This data is used to estimate hyperparameters for prior distributions for the GBM parameters. I find this approach to be sensible.
The paper includes three case studies. The first is to use clothes dryer sales to estimate the sale of room air conditioners. The second uses iPod sales to estimate the GBM parameters for BlackBerry sales during the period 2005 – 2011. The third examines sales of compressed natural gas in a Colombian municipality using two other municipalities. The cast studies are diverse, even if one of them is pretty dated.
There’s a remark in the study of air conditioner sales that clothes dryers are considered to be part of the same product group (home appliances). It seems to me that more accurate hyperparameter estimates could be obtained by considering multiple products in that group.
The second case study (BlackBerry using iPod) seems potentially flawed in that it includes the period in which iPhone was introduced and competed with BlackBerry. To me this would represent a model regime shift that should be accounted for.