• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

Abstract. Customer retention is a key problem in the insurance industry. As new customers are generally not profitable for the first few years, minimizing defection is critical. The objective of this study is to mine the company's operational databases to predict the insured households that will most likely defect within the next 12 months. The operational databases available for mining consisted of all active policies as of January 1994 and new policies written thereafter in a particular business region. Building the analysis dataset presented several challenges. Policy level data had to be aggregated into household level information and matched with demographics from other databases. Customers who moved to a different address had to be tracked. We constructed snapshot files of active customers for each of the years 1994-1998. Each snapshot file contained information on about 600,000 households and was used to build models using logistic regression (mainly) and decision trees. Each year's model was used to predict defection for the following year. Results showed that the yearly models had little variations in terms of model fit, gains, relative importance of predictors, and other measures. Further, they were uniformly accurate in identifying short-term defection. This means that, barring major changes in the marketplace, a model based on 1998 data should reliably predict the likely defectors in 1999.

Citations