Forecasting Variable Insurance Surrender Rates in Korea: Explainable Machine Learning and Macro‐financial Transmission
TL;DR
Out‐of‐sample SHAP analysis indicates that business cycle conditions, interest rates, surrender‐history variables, and selected insurance‐industry characteristics receive substantial but product‐specific predictive attribution.
Abstract
This study examines aggregate life insurance surrender risk as a macro‐financial transmission mechanism linking household contract adjustment to insurer balance sheets. Using monthly data for Korean Variable Annuity Insurance (VAI) and Variable Universal Insurance (VUI) from 2006 to 2019, we compare econometric and machine learning forecasts over 1‐, 3‐, 6‐, and 12‐month horizons. Forecast performance is evaluated over 2017–2019 using expanding and 120‐month rolling windows. No model dominates uniformly across products, horizons, and estimation windows. Tree‐based methods frequently perform well, but SARIMAX and LASSO remain competitive in other settings, indicating that the value of model flexibility is state‐, product‐, and horizon‐dependent. At the 1‐month horizon, the best rolling models reduce RMSE relative to lag‐1 persistence by 22.7% for VAI and 30.1% for VUI; the corresponding expanding window reductions are 28.2% and 34.2%. Scaling these improvements by average 2017–2019 in‐force amounts yields approximately KRW 30.7 billion and KRW 154.7 billion of exposure‐equivalent forecast error reduction for VAI and VUI, respectively. Out‐of‐sample SHAP analysis indicates that business cycle conditions, interest rates, surrender‐history variables, and selected insurance‐industry characteristics receive substantial but product‐specific predictive attribution. The findings suggest that explainable machine learning can complement conventional econometric forecasting by capturing nonlinear and product‐specific surrender dynamics.