Predicting term life insurance surrender using deep neural networks

Life insurance has a very low adoption rate in Iran, mainly due to policy surrender. This research aims to analyze the individual characteristics and insurance contract features that influence the surrendering of term life insurance policies.

The study utilizes a pilot database of 35,171 policy-holders and pensioners registered by an Iranian insurance company in 2021. Data mining, deep learning, and neural network algorithms are used for analysis due to their high accuracy in prediction:

The model demonstrates desirable performance based on evaluation metrics with a 74 percent accuracy in predicting both types of surrendered and non-surrendered insurance policies. The model performs better in predicting non-surrendered insurance policies more attention is given to interpreting those results. Despite imbalanced data, the model still performs well. In the dataset, surrendered policies make up only 3 percent of the total, leading to bias towards predicting the majority class. Nonetheless, the model accurately predicts and categorizes most surrendered policies, covering 59 percent of the total 244 cases.

The results indicate that certain demographic characteristics, such as age, female gender, health surcharge, and accident risk rate, as well as specific contract characteristics, including policy term, time since start date, longer premium payment methods, higher annual increase in capital and premium, fewer covered risks, and lower benefits, are negatively correlated with policy surrender. Furthermore, the results suggest that if the insured person is the policy surrender themselves, the probability of surrender is minimized. On the other hand, if the insured person is someone else, especially distant relatives, the probability of surrender increases.