Can General-Purpose Models Outperform Specifically Tailored Models? Evidence from Life Insurance Under-writing: A Case of Medical Examination Sampling

Accurate risk assessments are critical to the profitability of life insurers, with medical examina-tions playing a vital role in evaluating applicants’ health risks. While increasing medical examination coverage enhances assessment accuracy, it also increases operational costs and prolongs assessment time. A proven approach to address this issue is to use predictive models to classify applicants based on risk and focus medical examinations on high-risk applicants. This study compares two gradient-boosted tree models – the specifically tailored models and the general-purpose models – that predict risk defined as hospitalization days in a five-year period. The specifically tailored models were trained on over 400,000 contracts from 2015 to 2017, em-ploying the sliding window method to detect characteristics related to risks. In contrast, the general-purpose models were trained on data from 4 million policyholders as of 2015, suitable for various use-cases like applicant assessment and marketing. Three metrics – relative ratios in hospitalization days, health abnormality rate, and claim ratio – were used to evaluate models’ performance. Our findings suggest that the general-purpose models capture a broader represen-tation of the population due to their larger dataset and more balanced target variables. On aver-age, they perform 28% better than the specifically tailored models across the aforementioned metrics. Furthermore, the wider applicability of the general-purpose models reduces the need to train multiple models for different business scenarios, aligning with sustainability goals. There-fore, the general-purpose models outperform the specifically tailored models in medical exam-ination sampling.