A generalised comparison of Pareto/NBD based forecasts using MCMC, maximum likelihood, and heuristics

Abstract This study is the first generalised effort to compare the forecasting efficacy of different Pareto/NBD model-based forecasts. Monte Carlo Markov Chain (MCMC)-based estimates, Maximum Likelihood Estimation (MLE), and heuristics are applied to four different types of forecasts: (1) predicting the future number of purchases a single customer makes within a given period, (2) identifying active customers who will make at least one purchase within a given period, (3) identifying the customers who will belong to the top segments of the customer base, and (4) predicting the timing of a customer's next purchase. The results show that the model-based forecasts outperform the heuristics regarding predictive power and accuracy for the first three types of forecasts. MCMC yields slightly better results than MLE and it can additionally convince with confidence intervals for the number of future purchases. Forecasting the timing of a customer’s next purchase yields deviations that are too large to be used in practice.