AI-DRIVEN PREDICTIVE ANALYTICS MODELS FOR ENHANCING GROUP INSURANCE PORTFOLIO PERFORMANCE AND RISK FORECASTING

Abstract

This study evaluated whether AI-driven predictive analytics models enhanced group insurance portfolio performance and improved risk forecasting under renewal-cycle volatility and heavy-tailed claims. The empirical dataset covered four renewal years and included 412 sponsors and 186,540 members in Year 1, expanding to 463 sponsors and 209,940 members by Year 4, with total exposure rising from 2,143,210 to 2,409,760 member-months. Claim incidence remained stable at 0.27–0.29, yet utilization intensity increased as mean claim frequency per claimant rose from 2.6 to 2.9 and mean severity increased from 2,960 to 3,360 USD, while the 95th percentile severity exceeded 21,000 USD in Year 4. High-cost members comprised only 3.4–3.9% of lives but generated 41.8–44.2% of total costs, confirming tail dominance. Sponsor performance showed baseline instability, with mean loss ratios increasing from 0.83 to 0.90 and loss-ratio standard deviation widening from 0.19 to 0.23; sponsors exceeding stop-loss attachment increased from 7.1% to 8.9%. Benchmark actuarial models achieved moderate discrimination (0.66–0.74 across tasks) and higher errors for tail-sensitive outcomes. AI models—particularly gradient boosting and hybrid actuarial–AI specifications—produced consistent uplift: discrimination improved for high-cost events from 0.74 to 0.87 and for sponsor loss-ratio forecasting from 0.66 to 0.81. Sponsor loss-ratio error declined from 0.084 to 0.057, tail-exceedance misclassification fell from 0.078 to 0.049, stop-loss attachment error decreased from 0.071 to 0.045, reserve bias reduced from 2.9% to 1.7%, and sponsor loss-ratio volatility compressed from 0.23 to 0.18. These gains persisted under sponsor-stratified out-of-sample testing, renewal-forward validation, industry-shift checks, and feature perturbation tests. Overall, AI-driven predictive analytics materially strengthened group portfolio forecasting and stability by improving tail risk identification and sponsor-level loss ratio prediction.