Beyond isotropic uncertainty in tropical cyclone tracks with adaptive conformal prediction

Abstract Tropical cyclones are among the most destructive natural hazards, necessitating reliable uncertainty quantification in track forecasting. However, traditional probability prediction has problems such as isotropic regions and difficulty in reflecting geometric distortion. This paper proposes an adaptive uncertainty quantification framework based on distribution-free conformal prediction. This framework utilizes deep learning models to construct anisotropic probability ellipses and integrates an online calibration mechanism to achieve joint dynamic adjustment of the ellipse scale and principal axis direction. Results demonstrate our proposed framework moves beyond conventional methods, offering a more expressive and operationally relevant characterization of forecast uncertainty.