The variable annuity product has many desirable features for retirement saving purposes, such as stock-linked growth potential, protection against losses in the investment, and guarantees of minimum payout amount at annuitization. Therefore, it is of great interest to study this product for designing next generation retirement solutions. Policyholder behavior is one of the most important profit or loss factors for the variable annuity product, and insurance companies generally do not have sophisticated models at the current time. This paper will discuss a new approach using modern statistical learning techniques to model policyholder withdrawal behavior with promising results. Copyright © 2014 John Wiley & Sons, Ltd.