Non-contractual churn prediction using Hierarchical Temporal Memory

Abstract

As markets become more saturated and industry leaders compete over the existing customer base, competitors look for ways to improve customer retention with their customers. It is considered much more expensive to gain a new customer than retaining one, so the industry leaders look for ways in which churn in a customer can be predicted and potentially be avoided by incentivizing the customer to stay. Several of the previously proposed approaches struggle with combining the linearity with the non-linearity that exist within churn analysis and prediction. This emphasizes the need for research into state-of-the-art algorithms that furthers the knowledge regarding churn analysis that utilize the temporal structure of the data in a prediction based model. As contributions to this end, this thesis examines a Hierarchical Temporal Memory (HTM) approach to predict the future purchase events of customers in a non-contractual setting. The thesis compare the results of the HTM to the potential of existing state-of-the-art in the same context. The research shows HTMs potential through documenting performance with increasing data availability. The robustness of the implementation remains in question as complexity issues arise in conceptualizing a good definition for churn. HTM proves to be a viable churn detection algorithm, but has weaknesses in terms of churn prediction. The robustness of HTM increases with available data and the multimodality of that data.