Neural networks have been tremendously successful in a variety of domains. Of course, they have also been extensively studied in actuarial research and applied in insurance companies. For example, an application which is highly relevant for life insurance is the modeling and forecasting of mortality rates.

While simple feed-forward neural networks are already very effective for a number of tasks, there are several techniques (such as embedding layers or ensembles) and specialized architectures (such as recurrent or convolutional neural networks) which are necessary to exploit the full potential of these powerful machine learning models.

Interpretability and explainability are often key requirements in actuarial applications. Successful approaches in this direction have been proposed and are, for example, based on combining neural networks with the more traditional generalized linear models in the so-called combined actuarial neural network or LocalGLMnet architectures.