Predictive multiview embedding

TL;DR

This paper introduces predictive multiview embedding, a novel approach that combines multiple data sources to improve long-term climate variable predictions, demonstrating its effectiveness with hybrid models outperforming purely empirical ones.

Contribution

It presents a new predictive multiview embedding framework that integrates natural measurements and model outputs for enhanced climate prediction accuracy.

Findings

Hybrid models outperform purely empirical models in climate prediction.

Predictive multiview embedding reveals the local manifold structure of climate attractors.

The approach enables building prediction bounds and exploring predictability of climate variables.

Abstract

Multiview embedding is a way to model strange attractors that takes advantage of the way measurements are often made in real chaotic systems, using multidimensional measurements to make up for a lack of long timeseries. Predictive multiview embedding adapts this approach to the problem of predicting new values, and provides a natural framework for combining multiple sources of information such as natural measurements and computer model runs for potentially improved prediction. Here, using 18 month ahead prediction of monthly averages, we show how predictive multiview embedding can be combined with simple statistical approaches to explore predictability of four climate variables by a GCM, build prediction bounds, explore the local manifold structure of the attractor, and show that even though the GCM does not predict a particular variable well, a hybrid model combining information from the GCM and empirical data predicts that variable significantly better than the purely empirical model.