Introduction
In non-contractual retail environments, such as online grocery stores or general e-commerce, identifying when a customer has "churned" is a significant challenge. Unlike subscription services, customers do not formally cancel; they simply stop purchasing. To address this, the industry relies on Buy-Till-You-Defect (BTYD) models to estimate latent customer activity and future transaction potential.
However, standard BTYD models often suffer from a "face validity" crisis, producing "zombie" predictions—estimates that certain customers will remain active for thousands of years. This paper introduces a Mixture Hierarchical Bayes (MHB) model that resolves these unrealistic estimates by accounting for distinct customer segments.
The Variance Inflation Trap
The primary cause of unrealistic lifetime predictions in classic models, like the Pareto/NBD, is the assumption of a single, unimodal distribution for customer behavior. In reality, most customer bases are multi-modal, consisting of "one-hit wonders" and "die-hard loyals."
When a single-curve model attempts to fit these diverse groups, it must inflate its variance to cover the outliers. This creates a "fat tail" in the distribution, leading the model to erroneously conclude that some customers have a defection rate of nearly zero. The result is a population of predicted "zombies" with astronomical expected lifetimes.
The Mixture Hierarchical Bayes (MHB) Approach
The authors propose the MHB model to capture this inherent heterogeneity. Unlike standard Hierarchical Bayes (HB) models that use a single joint log-normal distribution, the MHB utilizes a mixture of $K$ multivariate normal distributions.
Key features of the model include:
- Latent Segmentation: The model automatically clusters customers into segments with similar purchase and defection behaviors.
- Concomitant Variables: By integrating a Multinomial Probit (MNP) sub-model, the MHB can use initial customer data (e.g., the size of a first basket) to predict which segment a new customer likely belongs to from day one.
- Bayesian Inference: Estimation is performed via a Metropolis-within-Gibbs sampler, treating the unobserved active/inactive status of customers as latent variables.
Empirical Results and Model Performance
The model was tested using transaction data from a Western European online grocery retailer. The MHB identified two primary segments: a "Frequent Buyer" group (41%) with high loyalty and low price sensitivity, and an "Incidental Buyer" group (59%) who purchased rarely and defected quickly.
As shown in the figure above, the MHB model demonstrates strong calibration between predicted and true lifetime values. Most importantly, it successfully eliminated the "zombie" prediction problem. While the standard HB model produced mean lifetime predictions in the range of $10^{45}$ days, the MHB model brought these estimates down to a sensible mean of approximately 20 years and a median of 7 years.
Managerial Implications
The ability to segment customers based on their first purchase offers significant advantages for targeted marketing. For instance, the grocery data revealed that frequent buyers often value convenience over price, paying higher delivery fees for smaller, more frequent baskets. By identifying these high-value segments early, managers can optimize acquisition costs and tailor retention strategies to the specific needs of each group, rather than relying on a one-size-fits-all model that produces mathematically valid but practically impossible predictions.
Summary updated