Hidden Markov models for churn prediction

Most companies favour the creation and nurturing of long-term relationships with customers because retaining customers is more profitable than acquiring new ones. Churn prediction is a predictive analytics technique to identify churning customers ahead of their departure and enable customer relationship managers to take action to keep them. This work evaluates the development of an expert system for churn prediction and prevention using a Hidden Markov model (HMM). A HMM is implemented on unique data from a mobile application and its predictive performance is compared to other algorithms that are commonly used for churn prediction: Logistic Regression, Neural Network and Support Vector Machine. Predictive performance of the HMM is not outperformed by the other algorithms. HMM has substantial advantages for use in expert systems though due to low storage and computational requirements and output of highly relevant customer motivational states. Generic session data of the mobile app is used to train and test the models which makes the system very easy to deploy and the findings applicable to the whole ecosystem of mobile apps distributed in Apple´s App and Google´s Play Store.