For customer-centric firms, churn prediction plays a central role in churn management programs. Methodological advances have emphasized the use of customer panel data to model the dynamic evolution of a customer base to improve churn predictions. However, pressure from policy makers and the public geared to reducing the storage of customer data has led to firms' self-policing' by limiting data storage, rendering panel data methods infeasible. We remedy these problems by developing a method that captures the dynamic evolution of a customer base without relying on the availability past data. Instead, using a recursively updated model our approach requires only knowledge of past model parameters. This generalized mixture of Kalman filters model maintains the accuracy of churn predictions compared to existing panel data methods when data from the past is available. In the absence of past data, applications in the insurance and telecommunications industry establish superior predictive performance compared to simpler benchmarks. These improvements arise because the proposed method captures the same dynamics and unobserved heterogeneity present in customer databases as advanced methods, while achieving privacy preserving data minimization and data anonymization. We therefore conclude that privacy preservation does not have to come at the cost of analytical operations. (C) 2016 Elsevier B.V. All rights reserved.