A Regression-Based Scheme for Seasonal Forecasting of New Zealand Temperature

Abstract

A statistical scheme for predicting New Zealand seasonal-mean temperatures at the beginning of each season has been developed. It is significantly more skillful (in terms of the percentage explained variance) than earlier schemes, raising overall predictive skill to more than 30%, compared to less than 20% skill for prior schemes. Careful selection of predictors, based largely on physical reasoning, is the key reason behind the relatively high skill levels achieved. In particular, summer rainfall is identified as a useful predictor for autumn temperatures. This appears to be related to deep-level soil moisture and temperature, vegetation cover, and their subsequent effects on air temperature. This study emphasizes the importance of spatial coherence between regions in the selection of predictors. Seasonally dependent predictors are shown to significantly improve predictive skill, while spatially dependent predictors generally decrease skill. Estimates of the potential predictability of seasonal-mean New Zealand temperatures suggest that, in all seasons except summer, the statistical relationships documented here capture around 80% of the potential predictive skill.