A Nonanticipative Analog Method for Long-Term Forecasting of Air Temperature Extremes

TL;DR

This paper introduces a nonanticipative analog method for long-term prediction of air temperature extremes using diverse global datasets, achieving up to 18.2% specific prediction accuracy and perfect sign accuracy, surpassing traditional distribution-based methods.

Contribution

The paper presents a novel nonanticipative analog approach that predicts temperature extremes without relying on probability distribution estimation, improving long-term forecasting accuracy.

Findings

Up to 18.2% of temperature extremes are specifically predicted.

Method achieves 100% accuracy in predicting the sign of temperature extremes.

The approach outperforms traditional methods by not requiring probability distribution estimation.

Abstract

A nonanticipative analog method is used for the long-term forecast of air temperature extremes. The data to be used for prediction include average daily air temperature, mean visibility, mean wind speed, mean dew point, maximum and minimum temperatures reported during the day from 66 places around the world, as well as sea level, average monthly Darwin and Tahiti sea level pressures, SOI, equatorial SOI, sea surface temperature, and multivariate ENSO index. Every dataset is split into two samples - learning (1973-2010) and validation (2011-2013). Initially, the sum of variables in datasets for two locations, minus corresponding climatological values, is calculated over a summation interval of length from 1 to 365 days. A "quality criterion" selects datasets for two locations with appropriate lead-time and summation interval, which have maximum (or minimum) sum compared with the rest of data four times at least, when extreme events occur later within the learning sample. Up to 18.2% of all extremes are specifically predicted. The methodology has 100% accuracy with respect to the sign of predicted and actual values. It is more useful than current methods for predicting extreme values because it does not require the estimation of a probability distribution from scarce observations.