Predicting customer churn based on changes in their behavior patterns

Customer retention is one of the most important tasks of a business, and it is extremely important to allocate retention resources according to the potential profitability of the customer. Most often the problem of predicting customer churn is solved based on the RFM (Recency, Frequency, Monetary) model. This paper proposes a way to extend the RFM model with estimates of the probability of changes in customer behavior. Based on an analysis of data relating to 33 918 clients of a large Russian retailer for 2019–2020, it is shown that there are recurring patterns of change in their behavior over a single year. Information about these patterns is used to calculate the necessary probability estimates. Incorporating these data into a predictive model based on logistic regression increases prediction accuracy by more than 10% on the metrics AUC and geometric mean. It is also shown that this approach has limitations related to the disruption of behavioral patterns by external shocks, such as the lockdown due to the COVID-19 pandemic in April 2020. The paper also proposes a way to identify these shocks, making it possible to forecast degradation in the predictive ability of the model.