Abstract:This article provides an empirical statistical analysis and discussion of the predictive abilities of selected customer lifetime value (CLV) models that could be used in online shopping within e-commerce business settings. The comparison of CLV predictive abilities, using selected evaluation metrics, is made on selected CLV models: Extended Pareto/NBD model (EP/NBD), Markov chain model and Status Quo model. The article uses six online store datasets with annual revenues in the order of tens of millions of euros for the comparison. The EP/NBD model has outperformed other selected models in a majority of evaluation metrics and can be considered good and stable for non-contractual relations in online shopping. The implications for the deployment of selected CLV models in practice, as well as suggestions for future research, are also discussed.

The selection of a suitable customer lifetime value (CLV) model is a key issue for companies that are introducing a CLV managerial approach in their online B2C relationship stores. The online retail environment places CLV models on several specific assumptions, e.g. non-contractual relationship, continuous purchase anytime, variable-spending environment. The article focuses on empirical statistical analysis and predictive abilities of selected probabilistic CLV models that show very good results in an online retail environment compared to different model families. For comparison, eleven CLV models were selected. The comparison has been made to the online stores’ datasets from Central and Eastern Europe with annual revenues of hundreds of millions of euros and with almost 2.3 million customers. Probabilistic models have achieved overall good and consistent results on the majority of the studied transactional datasets, with BG/NBD and Pareto/NBD models that can be considered stable with significant lifts from the baseline Status quo model. Abe's variant of Pareto/NBD have underperformed multiple criterions and would not be fully useful for the studied datasets without further improvements. In the end, the authors discuss the deployment implications of selected CLV models and propose further issues for future research to address.

The article contributes to the knowledge of customer lifetime value (CLV) models, where extensive empirical analyses on large datasets from online stores are missing. Based on this knowledge, practitioners can decide about the deployment of a particular model in their business and academics can design or enhance CLV models. The article presents predictive performance of selected CLV models: the extended Pareto/NBD model, the Markov chain model, the vector autoregressive model and the status quo model. Six large datasets of medium and large-sized online stores in the Czech Republic and Slovakia are used for a comparison of the predictive performance of the models. Online stores have annual revenues in the order of tens of millions of euros and more than one million customers. The comparison of CLV models is based on selected evaluation metrics. The results of some of the models which use additional non-financial data on customer behaviour-the Markov chain model and the vector autoregressive model-do not justify the effort which is needed to collect such data. The advantages and disadvantages of the selected CLV models are discussed in the context of their deployment.