Abstract/Summary

Statistics is fundamental for climate science. It helps making sense of its complex and multi-scale features by characterizing its aggregated behaviour. However, statistical methods used to extract causal information about this behaviour are often based on correlation and pattern detection, which lack a causal interpretation. Causal networks are emerging as a framework that can bridge statistics and causal meaning. This thesis seeks to contribute in advancing the use of causal network-based methods for the understanding of sub-seasonal to seasonal variability and predictability. The system of interest is the Southern Hemisphere mid-to-high latitude large-scale circulation variability in spring-to-summer, which is characterised by a strong downward influence of the stratospheric polar vortex on the tropospheric eddy-driven jet and its seasonal latitudinal shifts. The coupling extends the predictability of the troposphere due to the usually more predictable stratospheric dynamics. In this thesis, firstly the strength and timescale of the coupling are estimated from reanalysis with a time-series causal network, revealing the biasing effect of the vortex internal dynamics on all cross-correlations with the jet. The detected coupling can explain the enhanced jet autocorrelations and the effect of ozone depletion on its poleward trend in the late 20th century. Secondly, the predictability of the coupled variability is studied with a Bayesian causal network with a large ensemble hindcast. Marginal predictability is found given long-lead drivers, such as El Nino Southern Os- ˜ cillation and Polar Night Jet oscillation. The jet is highly predictable given the vortex state, also for its poleward shift, confirming a hypothesis present in the literature. Motivated by the presence of non-stationarity in this system, a causal discovery algorithm for regime-dependent non-stationarity is proposed. Its skill is shown for a suite of synthetic systems and one real-world example.