This summer school on `Deep Learning for Actuarial Modeling’ aims at given an overview and an introduction to the latest developments in this field.
We start by giving a solid technical basis on statistical modeling by introducing the familiar framework of generalized linear models (GLMs) which is based on the exponential dispersion family (EDF) of distributions. This family contains the most important distributions for regression modeling in actuarial science, such as the Poisson model, the gamma model, or Tweedie’s model.
Based on the GLM we dive into deep learning. The first extension considered is a classical feed-forward neural network (FNN), which can be obtained by a straightforward modification of a GLM. The main extension concerns that the original covariates are replaced by a so-called feature extractor, which is a deep learning tool that performs representation learning on the original covariates, aiming at extracting the most relevant information for prediction. Furthermore, we discuss fitting these enhanced regression models, as well as mitigation of statistical biases. This gives us a solid basis for the subsequent chapters.
A crucial technique in deep learning is entity embedding, which is especially useful when dealing with many categorical covariates being of high-cardinality. We discuss these methods, which form the main tool to bring covariate information into a suitable tensor structure for more advanced deep learning tools. These are then presented, like the recently developed attention layers and transformers, which are the core deep learning modules in large language models (LLMs) such as ChatGPT. We also discuss the credibility transformer, which integrates Bühlmann credibility into the transformer architecture. This credibility mechanism improves model fitting and it integrates explainable features into the transformer architecture.
Furthermore, we discuss recurrent neural networks (RNNs) as another class of architectures designed to model sequential data, which may arise in actuarial science. By leveraging information from previous observations, RNNs are particularly effective for tasks involving time-series data, such as mortality modeling and forecasting.
This summer school is completed by discussing several special deep learning architectures such as the LocalGLMnet, that aims at locally mimicking the behavior of a GLM, a summary of the Kolmogorov-Arnold network (KAN) as well as further tools that are useful to explain the predictive models and their results.
