Leroux et al. propose SVM for extreme quantile regression with heavy-tailed data
New SVM framework handles extreme value predictions with finite-sample guarantees and real-world river data.
A new paper from Leroux, Dombry, and Sabourin tackles the challenge of quantile regression when covariates take unusually large values—a problem common in finance, climate science, and engineering. The authors introduce a Support Vector Machine framework that leverages reproducing kernel Hilbert spaces to perform regression on extreme observations. By focusing on the angular components of data under regular variation assumptions, the method localizes learning in the tail of the covariate distribution, enabling robust extrapolation even in high-dimensional and nonlinear settings. The model also avoids boundedness assumptions, making it suitable for unbounded response variables. Finite-sample learning guarantees are established under mild regularity conditions.
The theoretical contributions are complemented by an empirical study on river flow data from the Danube, demonstrating practical relevance. The work unifies ideas from statistical learning and multivariate extremes, offering a principled approach to out-of-distribution generalization. This is particularly valuable for applications like flood risk assessment, extreme weather forecasting, and anomaly detection where rare but catastrophic events must be predicted accurately.
- Uses SVM with reproducing kernel Hilbert spaces for nonlinear, high-dimensional extreme quantile regression.
- Provides finite-sample learning guarantees for heavy-tailed inputs without restrictive transformations.
- Validated on Danube river flow data, showing practical applicability in environmental extremes.
Why It Matters
Enables reliable predictions for rare, catastrophic events by unifying statistical learning with extreme value theory.