Abstract
We test the ability of Gaussian Process Regression (GPR), a non-linear Bayesian regression method, to reconstructthe Atlantic Multidecadal Variability (AMV) over the last 2000 years (Common Era). The historical observationrecord is too short to provide a long-term perspective of the AMV. Therefore, drivers of the AMV, its response toexternal forcing and its timescales are a topic of active research and debate. Reconstructions of the AMV over theCommon Era have yielded conflicting results about the stationarity of the AMV timescales. These reconstructionshave mostly relied on terrestrial proxy networks and were based on linear regression methods which are known tounderestimate the true variability. Combining marine and terrestrial proxies and using non-linear methods couldimprove the AMV reconstructions. The GPR provides a flexible non-linear framework that can take measurementuncertainties into account and fit networks with a variable number of records in time. As a starting point, wecreate pseudo-proxies from a coupled Common Era simulation and test the GPR reconstruction of the AMV in thiscontrolled environment. We investigate the effects of different spatial networks, including marine proxy sites fromthe PAGES2k network, as well as different levels of proxy complexity such as the amount of non-climatic noise ornon-linear proxy-temperature relationships.
