To model, or not to model: Forecasting for customer prioritization

Abstract

Abstract Simple heuristics are usually deemed to be inferior to more complicated models. Although recent studies have demonstrated the usefulness of some forecasting heuristics, the questions of why and when a heuristic would work remain unaddressed. This study aims to answer such “why” and “when” questions by looking empirically at the specific context of forecasting for customer prioritization. Based on widely-applied probabilistic models, a series of simulations reveal that: (1) we are not usually able to identify the future top- X % of customers in a customer base accurately, even if we know the exact data generation process; (2) a simple heuristic can perform as well as a probabilistic model even if the model maps the data generation process exactly; (3) the relative performances of the model and the heuristics can be explained by several easily-obtainable descriptive statistics. The heuristic works because the minimal information it relies upon is relatively robust and relevant in a random world.