The webinars will take place on Zoom and a link to the virtual room will be sent out to all those who registered at the registration page.
Speaker: Péter Vékás (Corvinus University of Budapest)
Title: AI in Longevity Risk Management: Improved Long-Term Projections by Machine Learning
Abstract: While human mortality has decreased significantly since the beginning of the past century, resulting in unprecedented increases in human life expectancies, several authors have noted a historical pattern of diminishing mortality decline at relatively younger ages along with accelerating improvements among the elderly. Li, Lee and Gerland (2013) call this phenomenon the ’rotation’ of the age pattern of mortality decline. A somewhat simplistic explanation of this is that spectacular decreases in infant and childhood mortality rates (e.g., due to widespread vaccination programs and improved child nutrition) are less and less possible, while costly medical procedures to extend life at advanced ages are increasingly available.
The practical actuarial significance of the topic is that ignoring rotation in long-term mortality forecasts may lead to a severe and systematic underestimation of the old-aged population, which exacerbates longevity risk and may lead to serious adverse financial consequences for life and health insurers as well as pension schemes.
The popular model of Lee and Carter (1992) as well as many other mortality forecasting techniques do not allow for rotation at all. To correct this shortcoming, Li, Lee and Gerland (2013) introduced a variant of the Lee–Carter model including rotation. This model extension assumes that the evolution of mortality improvement rates follows a parametric equation, whose two parameters govern the speed of rotation and the level of life expectancy where the process begins.
We use age-specific mortality rates of all countries by gender from the Human Mortality Database (HMD), and split the available time periods by country into a training set spanning from the first available year up to 1990, a validation set from 1991 to 1999 and a test set containing all years after 1999. Instead of fixed values of the two parameters mentioned in the previous paragraph, as suggested by Li, Lee and Gerland (2013), we propose to treat them as hyperparameters and optimize them on the validation set, as it is customarily done in machine learning, in order to improve longterm forecasting performance. Additionally, we propose deep neural networks specifically designed to capture the rotation of mortality decline in order to produce even more data-driven rotation schedules free of any prior assumptions, and we tune the hyperparameters of the networks on the validation set. As a third candidate, we also propose a generalized additive model involving the bivariate spline approximation of the residuals of the Lee–Carter model. This approach is halfway between fully parametric models such as the variant of the Lee–Carter model including rotation and fully data-driven ones such as deep neural networks.
We use the test set to assess and compare the performance of the rotated variant of the Lee–Carter model including hyperparameter tuning, the deep neural network capturing rotation and the spline GAM approach. We will point out which approach works best in the long run in every country, which countries are more or less prone to rotation, and how actual rotation schedules differ from the parametric form hypothesized by Li, Lee and Gerland (2013).
Finally, we use our models to assess longevity risk in a pension scheme and point out the potential financial benefits of implementing our improved methods of capturing rotation in mortality data, and also elaborate on the potential impact of COVID-19 and how it is best incorporated into these models.
The talk is based on a joint work with Ronald Richman (ETH Zürich) and László Kovács (Corvinus University of Budapest)
This series of webinars addresses all interested people in probability, stochastic analysis, control, risk evaluation, statistics, with a view towards applications, in particular to renewable energy markets and production. This series brings together the major research themes of the projects STORM, SCROLLER, and SPATUS.