Abstract
This study investigates an alternative modelling approach for empirical seasonal temperature forecasts over South America. Seasonal average temperatures are found to be non-stationary at most parts of South America over the 1949–2012 period. Simple persistence and lagged regression methods have considerable correlation skill in forecasting next season temperature using previous season temperature as predictor. However, the presence of trends in both predictor and predictand temperature variables can affect correlation skill. Models that can account for non-stationarity in these variables may do better in modelling and forecasting seasonal temperatures known to have trends. A novel method (cointegration), introduced here for empirical seasonal climate forecasting, is found to perform better than the traditional persistence and regression forecasts for places where the predictor and predictand temperatures have stochastic trends. Potential skill pairwise comparisons between temperature forecasts produced with cointegration and those produced using persistence and lagged regression have shown that the alternative cointegration method performs significantly better than the other two. One of the main reasons for the better performance of cointegration method is that the modelling procedure accounts for the existing non-stationarity in the process, and thus enables the estimated model to predict out of the range as efficiently as possible. Overall, this method appears to be ideal for modelling and predicting climate under the current global warming scenario. This is because most of the climatic variables including temperature in particular cannot be assumed to be stationary through time under such warming scenario.