Machine Learning Fundamentals for Actuaries

Abstract

A mortality table built from five years of experience data — features selected, crude rates graduated, thin cells credibility-weighted, results validated against emerging experience — is, in every structural sense, supervised machine learning. The actuary who built it never called it that, but the discipline she practised predates the term by more than a century. This article makes the connection explicit: machine learning is not a foreign discipline arriving from computer science departments; it is a formalisation of the reasoning actuaries already use, extended to more data, more features, and more complex relationships than a traditional parametric model can accommodate. We map the three learning paradigms — supervised, unsupervised, and reinforcement learning — to workflows every practitioner will recognise; show that the ML pipeline is the actuarial control cycle stage for stage; demonstrate that the bias-variance tradeoff is credibility theory by another name, with regularisation playing the role of the credibility factor Z; and set out an evaluation framework in which calibration, not discrimination, is the test that matters for actuarial applications. A lapse-prediction case study — gradient boosting versus logistic regression on an 85,000-policy term portfolio — grounds the framework in the trade-off practitioners actually face: predictive accuracy against the interpretability that statutory sign-off demands.