Predictability study of the observed and simulated European climate using linear regression

Abstract

AbstractMonthly mean temperature anomalies in the regions England, Germany and Scandinavia are predicted by linear regression. Two predictors are selected from monthly mean teleconnection indices, North Atlantic sea surface temperatures (SSTs) projected on the first three empirical orthogonal functions (EOFs), and European climate variables (temperature, sea level pressure, and precipitation) averaged in the three predictand regions. The predictors are chosen separately for each month according to their correlation with the predictand. Observations from 1870–1999 and data from a 600‐year integration with the coupled atmosphere– ocean general‐circulation model ECHAM/HOPE are used to assess and compare the forecast skill. The skill is measured by the anomaly correlation coefficient (ACC) and the explained variance (EV).For a one‐month lead time the ACC for observations is up to 0:6 (EV≈35%) for February–March and August–September in the three regions. The skill for the simulated data is lower (maximum values at ACC≈0:5,EV≈25%) and its seasonal dependence differs from that of the observations. Main predictors are the preceding temperatures in the predictand region. Using segments of the simulated data the spread of skill is estimated as 0.1 in ACC (10% in EV).For lead times up to one year there is a small ACC (0.3–0.4) in the observations for England (spring and late summer), and Scandinavia (August–September), but none in Germany. The observed two‐month mean England temperature in spring and late summer can be predicted with six months' lead time for 1971–96 with 1870–1969 as a training set, selecting the first two North Atlantic SST EOF coefficients as predictors. A leave‐two‐out cross‐validation in 1870–1999 shows a distinct reduction of skill. In simulated data, the skill beyond one month is negligible compared with the observations. Copyright © 2003 Royal Meteorological Society.