Testing slow varying predictors in a Nearest Neighbor model for statistical prediction of South East Asian Monsoon

The Indian summer monsoon rainfall contributes about 75% of the total annual rainfall and exhibits considerable subseasonal variations. The timing and magnitude of these 10-40 days rainfall variations are largely responsible for the success or failure of regional agriculture. Unfortunately, tropical interseasonal variability has proven difficult to simulate and even more difficult to predict from the first principles. An alternative approach is to use a statistical model. To date, a number of different statistical models have been proposed for monsoon forecast. But the main difficulty with the statistical models is in the selection of the predictors. There are two general approaches for the predictor selection. One is a pure statistical approach in which several predictors are selected from a large number of relevant variables based on the maximum correlation coefficients between the predictors and predictor. Another approach is to select the predictors from the physical principles governing the relations between the predictand and predictors. However both approaches have some drawbacks. A new Nearest Neighbor model of monsoon prediction has been developed to take advantage of the latter approach. The model was used to test different types of the predictors including slow varying predictors and predictors with nonlinear functional dependence between the predictand and predictor for short-term forecast of monsoon rainfall during 2001-2004. We demonstrate that incorporation of slow varying predictors results in more accurate forecast compared to a standard linear model. Advantages and disadvantages of the Nearest Neighbor model will be discussed.