Faculty of Actuarial Science and Insurance with Mathias Millberg Lindholm of Stockholm University

Title: Transformed gradient boosting and exponential dispersion models

Abstract:
Gradient boosting based models, including gradient boosting machines (GBMs), have proven to be highly competitive in predictive modelling. Since the introduction of GBMs in the early 2000’s, a large number of versions of gradient boosting methods have been developed. These include extensions such as XGBoost, NGBoost and lightGBM. Apart from going through the basic ideas underpinning GBMs, I will introduce a general transformed gradient boosting (TGB) method that can be used to provide theoretical guarantees of loss improvement for different versions of GBMs. In particular, this approach is used to motivate adjustments to NGBoost that improve the theoretical properties of the algorithm, which also makes it more closely related to XGBoost.

Further, in many actuarial applications the underlying stochastic model is a member of the exponential dispersion family (EDFs). When considering loss functions motivated by EDFs, it turns out that the TGB method can be significantly simplified, providing guidance on how to fit multi-parametric EDFs and how to include duration weights in a consistent manner.

This presentation is based on joint work with Ł. Delong, T. Nazar, and H. Zakrisson

Faculty Events | Website | View More Events

Contact the organizers