Long-Lead Seasonal Forecast Skill in Far Eastern Asia Using Canonical Correlation Analysis

Abstract

Canonical correlation analysis (CCA), a linear statistical method designed to find correlated patterns between predictor and predictand fields, is applied to the eastern Asian region of Korea and Japan. The cross-validation technique is used to estimate the levels and sources of forecast skill for 3-month-averaged surface air temperature and total precipitation for the 37-yr time period of 1961?97. Quasi-global SST, Northern Hemisphere 700-hPa height, and prior values of the predictand field itself are used as predictor fields in an attempt to maximize the strength of the predictive relationships. The global SST field turns out generally to contribute the most to the final skill, with the exception of the winter season, in which the geopotential height field contributes most. The highest skill for temperature forecasts occurs in early spring, with relative insensitivity to the forecast lead time. This skill is statistically significant, averaging over 0.3 but including higher values locally. A secondary seasonal skill maximum appears in late summer. The forecast skill of precipitation is not high overall but is relatively highest in early winter, with area-averaged skill of nearly 0.2. Diagnostic analysis of the CCA loading patterns indicates that the strongest predictive mode for temperature is a long-term warming trend that pervades the entire seasonal cycle and contributes to skill during both the cold and the warm peak skill seasons. Interannual temperature fluctuations related to ENSO are captured in the second mode. Because of the dominant role of trends, CCA skill decreases substantially when the datasets are detrended. Forecast skill for precipitation is found to be due to a combination of interdecadal variations and ENSO-related variations.