Mortality Modeling: Machine Learning and Mortality Shocks
Human longevity has increased significantly and often more than expected over the past decades. In response, sophisticated ways of modeling and forecasting mortality have been developed. Widely used models are mostly grounded on traditional statistical approaches such as maximum likelihood estimation and time series analysis. This thesis offers a thorough review of these techniques and existing stochastic mortality models. As a new contribution to the literature, machine learning methods are applied to mortality modeling, yielding additional perspectives on the data and improved forecasting performance. Benefits of the classical models such as interpretability and prediction uncertainty quantification are preserved as much as possible or even enhanced.
Specifically, the popular common age effect model is extended to include multiple common age effects, each of which is shared only by a subset of all considered populations. Different cluster analysis algorithms are applied to identify suitable subsets. The model extension is shown to be easily interpretable as it is based on the realistic assumption that there are several groups of populations with similar mortality dynamics within each group but not between groups. Furthermore, a new model inspired by the fuzzy clustering paradigm is introduced and analyzed. It is demonstrated to be particularly useful when a population is difficult to be classified with hard (non-fuzzy) clustering techniques.
Instead of improving an existing mortality model, a different approach is to completely replace it by a supervised machine learning method. To implement this idea, a convolutional neural network model trained on the age-period mortality surface is proposed. Its forecasting performance compares favorably to other neural networks and established mortality models. As neural networks by default only yield point forecasts, previous works applying them to mortality modeling have not investigated prediction uncertainty. This gap in the literature is addressed by a bootstrapping-based technique, which leads to reliable prediction intervals for the considered convolutional network.
Recently, the COVID-19 pandemic has interrupted the widespread trend of steady mortality improvements. The last part of this thesis addresses the new modeling challenges posed by the mortality shock it has caused in many countries. The consequences of a COVID-19-type shock both for mortality data and models are investigated. It is concluded that shocks can have a large impact on death rate forecasts and, consequently, on the valuation of mortality-related insurance products. Different ways of accordingly adjusting the models are compared and practical recommendations regarding model choice and calibration are derived.