Bayesian Neural Networks for Actuarial Mortality Modelling

The use of Bayesian neural networks (BNNs) for mortality modelling is an understudied, yet potentially promising area of research. They inherently offer robust uncertainty quantification, and are known for their application to sparse or small datasets. This research investigates the efficacy of BNN architectures in the presence of extreme data sparsity and explores methodologies for incorporating domain-specific prior information through hybrid structures. By embedding classical mortality laws directly into the neural network framework, we develop a suite of hybrid models capable of leveraging both the flexibility of deep learning, and the interpretability of parametric actuarial laws. We demonstrate that these Bayesian hybrid models provide promising forecasting capabilities and produce reliable uncertainty intervals, even under conditions of severe data sparsity. More specifically, our results indicate that these hybrid models can accurately capture underlying mortality patterns and interpolate missing values, even when 95% of the training data is absent. The performance of this mortality prediction approach is evaluated using female population data from a diverse set of countries, namely Iceland, Italy, Japan, Russia, Sweden, and the United States. Furthermore, we extend the above framework also to multi-population modelling. This is done by proposing an architecture that utilises embedding layers on population indices. These embeddings function as population-specific loading factors for the global terms within the hybrid models, allowing joint modelling across heterogeneous cohorts. While these models perform well on sparse multi-population data, we discuss the computational and runtime constraints encountered when scaling BNNs to comprehensive multi-population datasets.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Moehrke, Patrick Gary

Advisor dc:contributor.advisor

• Bierman, S.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136197

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136197