Hsi Chang
Title: Improving Lee-Carter Mortality Forecasts Using Multivariate Tree-Based Age Segmentation
Date: Wednesday, October 28th 2026
Time: 3pm
Location: Zoom
Supervised by: Cary Tsai
Abstract: Mortality dynamics can vary substantially across ages, whereas the Lee-Carter model uses a common time index throughout the entire age range. This project develops a multivariate regression-tree procedure that partitions ages into contiguous groups using standardized features constructed from annual changes in the logarithm of the central death rates. The Lee-Carter model is then fitted to each of the partitioned age groups. The procedure is evaluated using female and male mortality data over a range of study ages for the U.S., the U.K., and Japan from the Human Mortality Database. Forecast accuracy is assessed against the baseline Lee-Carter model using the mean absolute percentage error (MAPE). In general, tree-based segmentation improves average forecast performance in feature specifications, although gains vary between individual population鈥搒ex鈥揾orizon cases. Observed log mortality yields the largest overall MAPE reduction at each forecast horizon, while the best improvement-based specification varies across horizons.