Abstract

Causal networks are increasingly being adopted as a framework to guide robust statistical analysis in climate science, as well as a statistical modelling approach. Causal networks represent a system of variables as nodes and their causal relationships as directed links, which can be parametrized with functions or conditional probabilities. Their structure implies a set of conditional independence relationships, which provide testable implications given data. After a brief introduction to the subject, in this talk I will show an application of probabilistic causal networks to study long-range predictability of the Southern Hemisphere eddy-driven jet variability.

In the spring-to-summer months, the jet is influenced by the stratospheric polar vortex, whose variability is in turn affected by long-lead drivers in both the stratosphere and the troposphere. However, a quantification of the predictability arising from these drivers and their combination has been lacking. Here a probabilistic causal network model of the coupled stratospheric-tropospheric monthly variability is constructed to generate synthetic predictions of the jet and quantify the skill arising from the above-mentioned drivers. The vortex state is confirmed to be determinant for skilful predictions of jet’s variability. However, the vortex long-lead drivers provide only moderate skill. This suggests that long-range jet predictability may be ultimately constrained by how well models capture internal stratospheric variability on sub-monthly time scales.