Predicting Revenues With the Multiplier Heuristic

Abstract

Using machine learning to forecast revenues per customer, product, or store has become a major industry. It contrasts with a simple, intuitive managerial heuristic: multiply the revenue observed in the first t days by a constant. We test the predictive accuracy of this multiplier heuristic in 20 data sets. Surprisingly, the heuristic performs overall on par with machine learning models. We identify three central drivers when the heuristic can even outperform these models: limited sample size, time-to-prediction, and changes across time. The results provide insights when to rely on heuristics and managerial intuition or when on machine learning.