Simple El Ni\~{n}o prediction scheme using the signature of climate time series

TL;DR

This paper introduces a machine learning model that uses the path signature of climate time series to predict El Niño events, capturing the sequence and nonlinearity of climate phenomena for improved forecasting.

Contribution

It presents a novel approach applying path signature transformation to climate indices for El Niño prediction, enhancing understanding of climate event sequences.

Findings

Achieved a mean square error of 0.596 for 6-month predictions.

Identified North Pacific Index and NINO indices as key precursors.

Demonstrated El Niño's predictability based on climate event correlations.

Abstract

El Ni\~{n}o is a typical example of a coupled atmosphere--ocean phenomenon, but it is unclear whether it can be described quantitatively by a correlation between relevant climate events. To provide clarity on this issue, we developed a machine learning-based El Ni\~{n}o prediction model that uses the time series of climate indices. By transforming the multidimensional time series into the path signature, the model is able to properly evaluate the order and nonlinearity of climate events, which allowed us to achieve good forecasting skill (mean square error = 0.596 for 6-month prediction). In addition, it is possible to provide information about the sequence of climate events that tend to change the future NINO3.4 sea surface temperatures. In forecasting experiments conducted, changes in the North Pacific Index and several NINO indices were found to be important precursors. The results suggest that El Ni\~{n}o is predictable to some extent based on the correlation of climate events.