A Statistical Multimodel Ensemble Approach to Improving Long-Range Forecasting in Pakistan

We have designed, developed, and tested a method for generating long-range forecasting systems for predicting environmental conditions at intraseasonal to seasonal lead times (lead times of several weeks to several seasons). The resulting systems use statistical, multimodel, and lagged average ensemble approaches. The ensemble members are generated by multiple regression models that relate globally distributed oceanic and atmospheric predictors to local predictands. The predictands are three tercile categorical forecast targets. The predictors are selected based on their long-lead correlations to the predictands. The models are selected based on their lagged average ensemble skill at multiple leads determined from cross-validated, multidecadal hindcasts. The main system outputs are probabilistic long-lead forecasts, and corresponding quantitative assessments of forecast uncertainty and confidence. Our forecast system development process shows a high potential for meeting a wide range of military and national intelligence requirements for operational long-lead forecast support. The main test bed for our system development was long-range forecasting of environmental conditions in Pakistan. This problem was selected based on DoD and national intelligence priorities for long-range support. For this test case, the system uses 81 ensemble forecast members that predict the probability of summer precipitation rates in north-central Pakistan up to 6 months in advance. The cross-validated hindcast results from the test case system are substantially more skillful than reference climatological forecasts at all leads. The test results also show that the combination of multiple forecast member predictions in a multimodel, lagged average ensemble approach yields more accurate forecasts than any one forecast member individually. The original document contains color images.