Gradient boosting

Gradient boosting is a supervised learning method that builds a predictive model by successive additions. It starts with a simple prediction, then adds a small decision tree trained to correct the current model's residual errors, then another, and so on, each step reducing a loss function in the direction of its gradient. Formalized by Jerome Friedman in 2001, the method was popularized by fast implementations, including XGBoost in 2016, which dominated data science competitions on tabular data, precisely the format of insurance data. In pricing, it captures interactions and nonlinear effects that generalized linear models only see if they are specified, and it often improves discriminatory power. But the resulting model is a sum of hundreds of trees that no one can read directly, which raises problems of justification with regulators, policyholders and management. Practitioners therefore often combine it with the linear model: they use it to discover relevant interactions, which they then build into an explainable tariff, or they accompany it with interpretation tools such as Shapley values.