Statistical Learning Methods
Franck Vermet · 2018
This chapter presents the statistical learning methods most commonly used in actuarial science. These are complementary methods to the more conventional statistical models, such as linear and logistic regression, which have long been applied in actuarial science. The chapter discusses a general distinction between supervised statistical learning and unsupervised statistical learning methods. The purpose of supervised learning is to learn the link between two variables. Supervised learning methods include decision trees, layered neural networks, support vector machines (SVM) and model aggregation methods (bagging, random forests, boosting, stacking). In unsupervised learning, there is no variable to be explained, which therefore rather concerns a clustering problem. The objective is to construct homogeneous classes that group together the most similar individuals, and the classes have to be as dissimilar as possible. Among the conventional methods, the chapter discusses the ascending hierarchical classification and algorithms by dynamic reallocation (k-means).