Improved Customer Lifetime Value Prediction With Sequence-To-Sequence Learning and Feature-Based Models

Abstract

The prediction of the Customer Lifetime Value (CLV) is an important asset for tool-supported marketing by customer relationship managers. Since standard methods based on purchase recency, frequency, and past profit and revenue statistics often have limited predictive power, advanced machine learning (ML) techniques were applied to this task in recent years. However, existing approaches are often not fully capable of modeling certain temporal patterns that can be commonly found in practice, such as periodic purchasing behavior of customers. To address these shortcomings, we propose a novel method for CLV prediction based on a combination of several ML techniques. At its core, our method consists of a tailored deep learning approach based on encoder–decoder sequence-to-sequence recurrent neural networks with augmented temporal convolutions. This model is then combined with gradient boosting machines (GBMs) and a set of novel features in a hybrid framework. Empirical evaluations based on real-world data from a larger e-commerce company and a public dataset from the domain of online retail show that already the sequence-based model leads to competitive performance results. Stacking it with the GBM model is synergistic and further improves accuracy, indicating that the two models capture different patterns in the data.