Seasonal forecast of rainfall in central Chile using the "logistic" method

A "logistic" method is used to forecast rainfall conditions in central Chile during winter (May-August). These method forecasts for a longer period than those performed with canonical correlation analysis, which found significant rainfall predictability for July-August period.

The model is based on categorical classification using several predictors. It forecasts the probability of occurrence of each rainfall category. Parameters are estimated using maximum likelihood method. In comparison to the linear discrimination analysis the "logistic" method uses two parameters for each category.

In the present study 33 rainfall stations grouped in four homogeneous regions (30°-32°S, 32°-34°S, 34°-35°S and 35°-36°S) are used. Two consecutive bimonthly mean anomalies of SST in the region Niño 3 are used as predictors, from 3 to 0 months in advance. The forecast model parameters were adjusted over a 30 years period (1951-80). Rainfall category forecast during this period was assessed using the cross-validation method. Independent forecasts were obtained from 1981 to 1998.

The Heidke skill score was used in order to evaluate the accuracy of the forecast. For the first three regions and between lag 0 and lag 1 this score was found to be always greater than 15%. In region 2 this value was also found even for forecast of two months lead. Finally, the model performs better in rejecting an extreme category than forecasting its occurrence.