Factors Affecting Policyholders’ Risk Aversion in the Adoption of Additional Endowment Life Insurance Using a Machine-Learning Algorithm

Abstract

This study employed machine-learning analysis on a large-scale real-world dataset comprising first and second endowment life insurance policies issued by Sarmad Insurance Company, which is recognized as a leading provider of second life insurance policies in Iran, as well as data from Fanavaran Company, an insurance technology platform operating within the insurance industry. The study population consisted of 95,962 individuals who collectively purchased their first life insurance policies through a well-known insurance company in Iran, with the policyholder and the insured being the same person. Following customer relationship management activities and follow-up contacts, 18,604 of these individuals purchased a second life insurance policy. The data were collected between 2021 and 2025 and, after data cleaning and normalization, were used as inputs for the predictive model. To predict which customers were inclined to purchase additional insurance coverage, a predictive model was developed using one of the most effective machine-learning techniques, namely Extreme Gradient Boosting (XGBoost). The predictive accuracy of the selected model was subsequently evaluated using k-fold cross-validation. Given that previous studies have consistently confirmed the risk-averse tendencies of individuals who purchase life insurance, those who purchased additional life insurance under highly unstable economic conditions—despite having the option to allocate their funds to alternative assets—may be regarded as a particularly salient group of risk-averse individuals. The model incorporated 13 demographic, behavioral, and socioeconomic factors as independent variables, while the purchase of additional endowment life insurance was considered the dependent variable. Consequently, the distinctive characteristics of customers predicted to purchase a second endowment life insurance policy were identified. The findings demonstrated significant associations between the independent variables and the probability of purchasing additional insurance. Notably, 6 of the 13 independent variables exerted a substantial influence on the purchase of additional insurance. These variables included age, gender, city and province of residence, marital status, possession of other insurance policies, and possession of automobile insurance, including third-party liability and comprehensive motor insurance. By contrast, variables such as occupation, income, and educational attainment were identified as having limited predictive influence because of the high degree of similarity among the data obtained from employees of the large organizations comprising a substantial proportion of the insurance company’s customer base.