Neural Network Applications for Life & Health Insurance

In contrast to many other areas, neural networks remain out the model toolbox of most Life & Health actuaries. This might be due to the perceived complexity of neural networks and due to the lack of datasets with sufficient quality, consistent long-term coverage and sufficiently large sizes that would be necessary to benefit from neural networks.

In this presentation we will provide a brief overview of three applications of neural networks in a Life & Health context that demonstrate that with comparatively simple, shallow neural networks we obtain surprisingly good results. We will show that neural networks can and should be considered side by side with classical models in the actuaries' toolbox:

• Detection of anomalies in mortality rates: mortality rates might be affected by errors in exposures or by migration. Convolutional neural networks, which have found great success in image recognition tasks (using much larger and deeper networks), are also an excellent tool to detect these anomalies.
• Mortality forecasting: while classical mortality models like Lee-Carter or the age-period-cohort model usually only consider parameters along the age, time and cohort axes, neural networks provide much more flexibility and allow for example for effects that are bound to certain age and time ranges.

Neural networks as an alternative to classical survival models: more and more, large longitudinal datasets containing medical, socio-economic and lifestyle factors become available. These can be used to build models to predict relative mortality risk used in Life & Health underwriting. The more attributes are considered the more challenging the modelling of dependencies becomes. As an alternative to classical models like Cox regression, neural networks do perform very well on these use cases.