A new statistical scheme for prediction of New Zealand seasonal mean temperatures has been developed. It is significantly more skilful than earlier schemes, typically exhibiting explained variance of around 30%, compared to less than 20% for existing methods.

Improvements in skill come largely from careful selection of predictors, based on physical reasoning. Notably, summer rainfall is identified as a significant predictor of autumn temperatures, as a result of related effects on soil moisture and temperature. The emphasis is on spatial coherence: spatially dependent predictors act to decrease forecast skill, while seasonally dependent predictors act to increase forecast skill.

Based on estimates of potential predictability, the statistical relationships presented here capture about 80% of the potential predictive skill in all seasons except summer.