An application of stochastic forecasting to monthly averaged 700-mb heights

The statistical modeling of existing climatic records complements numerical modeling methods. Statistical models can be used to forecast climate variations without the physical understanding needed in the dynamical approach.Since most climatic records cover only several decades, these statistical methods are feasible for relatively short forecast periods of weeks or months. The statistical forecast method of multivariate autoregression is applied to a 27-year record of monthly averaged 700 mb heights. According to the theorem of predictive decomposition, a stationary series can be represented as the sum of deterministic and stochastic parts that are uncorrelated. Hence the deterministic annual cycle is removed and the statistical model applied to standardized anomalies of the 700 mb heights. The field of height anomalies is represented by 504 gridded values, which is far too large a dimension for multivariate autoregressive modeling. Therefore, 95 of these values representing equal areas are selected for modeling. Principal component analysis is then used to further reduce the 95 values to 20. These 20 principal components are modeled by a third-order multivariate autoregressive (MVAR) model that is selected objectively. Dependent and independent forecasts of the standardized height anomalies for 1 month are made and compared with similar forecasts made climatologically. THE MVAR model is found to perform better on the average than climatology.Forecasts for several months are also made and examined for retrogression of anomalies. For the single case studied, the model is found to move features both eastward and westward. 1982 NESDIS (National Environmental Satellite, Data, and Information Service) Library Public domain 2286