anomaly is defined with respect to the 5-year mean. The analysed 500mb height anomalies (defined with respect to the analysed 5-year mean) for DJF 1979/80, 1982/3 and 1983/4 are shown in Fig 3.
The ensemble-mean 500hPa height anomalies for the winters DJF 1979/80, 1982/83 and 1983/4 are shown in Fig 4. In each case, the anomalies have been calculated with respect to the model's own 5-year DIE climatology. Superimposed on the contours of ensemble mean anomaly are regions of stippling where, according to a t-test, the ensemble is statistically significantly different (at the 95% confidence level) from the ensemble of all integrations from the other 4 years.
In 82/8 3 in the northern extratropics, there is a negative height anomaly over the north-east Pacific, extending down into the southwestern USA, which is statistically significant. The ECMWF model has a central anomaly of nearly 16 dam, very close to the analysed value. Further "downstream" over Europe, the model shows a statistically significant meridionally-oriented height dipole over Europe, which appears to agree well with the analysis anomalies. For 1979/80, the ensemble spread is large. Hence the ensemble mean anomalies are small, and there are hardly any regions where the response is statistically significant. By contrast, for both 1982/83 and 1983/4, the ensemble mean appears significant in a number of places, and, moreover, the field appears skilful when compared with the verifying analysis.
d)winter distributions of skill score
In this section we discuss the northern hemisphere skill score distributions based on anomaly correlation. The observed anomaly for year "N" is simply the difference between the DIE 500hPa height from reanalysis for year "N", minus the average DIE 500hPa height for the 5 winters, excluding year "N". Within the ensemble, there are many realizations of the simulated anomaly for year "N". Specifically, one member of the ensemble for year "N", and one member from each of the four ensembles which are from years other than "N", are chosen. An anomaly is created by taking the difference between the chosen year "N" member and the average of the other four chosen members. There are approximately 59,000 ways of choosing such an anomaly field. The northern hemisphere spatial correlation of the chosen anomaly and the observed anomaly is calculated, and the calculation is repeated over all 59,000 combinations. Distributions based on categories of equal correlation intervals of 0.2 are then constructed (Fig 5), taken over all winters, and just over the winter DIE 82/83.
Fig 5 shows that, taken over all years, the ensemble distribution is biased towards positive values. The mode of the distribution lies somewhere between 0.0 and 0.4. About a quarter of all scores are greater than 0.4, whilst only about 2% are less than -0.4. About 5% of scores are higher than 0.6, whilst there are no scores less than -0.6.
The distribution for 1982/83 shows a considerably higher level of skill than the 5-year distribution. There is a strong mode between 0.4 and 0.6, and nearly a quarter of skill values are greater than 0.6.
