3588 
JOURNAL OF CLIMATE 
Volume 27 
confirms the skill of the model. While the performance is 
particularly high for moderate values. Fig. 3a also sug 
gests an underestimation of the most extreme events. 
This underestimation may be caused by different fac 
tors. First, the ensemble mean may potentially average 
a few single extreme events out. Second, because of the 
low temporal and spatial resolution of the model forcing 
(the reanalysis has a temporal and spatial resolution of 
6 h and 2°, respectively) some extreme events which are 
present in the observations may be lost. Finally, model 
inaccuracies may be responsible for the underestimation 
of the most extreme events. To test whether the first 
reason mentioned above is responsible for the differ 
ences, we also tested the predictive skill for each en 
semble member separately. While for some individual 
extreme events the model performance can be slightly 
improved (Fig. 3a), the general performance cannot be 
increased significantly. Overall, the ensemble mean 
shows the highest skill (Fig. 3b). It should be noted that 
the underestimation of some extreme events is not a 
unique shortcoming of the statistical model but also 
present in numerical models (Weisse and Pliip 2006; 
Arns and Jensen 2010). This is why we suggest that the 
second reason discussed above (i.e., the temporal and 
spatial resolution of the model forcing) is most likely 
responsible for the deviations in the highest percentiles. 
Because of their stochastic occurrence, it is unlikely that 
these differences affect the long-term behavior, which is 
further analyzed in the next sections. 
As introduced our main aim is to determine whether 
the statistically significant relationship between surges 
and 20CRv2 wind and pressure fields remains stationary 
in time. Hence, we apply the regression coefficients from 
Eq. (1) to the entire reanalysis period from 1871 to 2010. 
To test the stationarity of the relationship, we compute 
different efficiency criteria—namely, the coefficient of 
determination ( R 2 ) and the RMSE—between observed 
and predicted surges for each year back to 1871. 
1840 1860 1880 1900 1920 1940 1960 1980 2000 2020 
Time [yr] 
■ 
■ 
. I A A 
LI. A 
tl T 
. T T 
. 7.1 
.11. 1 
1 
T 4 
: 
Summer 
1 1 
■ Winter 
■ 
. Annual 
1 
95 
1 
98 
1 
99 
99.9 
Percentiles 
Fig. 4. Normalized (i.e., the long-term average has been removed) 
annual (black) and seasonal [red is October-March (ONDJFM); 
blue is April-September (AMJJAS)] time series of the (a) 95th and 
(b) 99.9th storm surge percentiles. Low-pass-filtered time series 
[10-yr locally weighted scatterplot smoothing (LOWESS) filter] are 
shown as thick lines. The gray and blue shaded areas represent pe 
riods of increased and decreased storminess, respectively. For 
presentation purposes, the time series are shown with an arbitrary 
offset [0,20, and 40 cm in (a) and 0,60, and 120 cm in (b)]. (c) Linear 
trends ±2cr SE of four upper percentile time series for the period 
1843-2012. 
3. Results 
a. Storm surge trends and variability 
We present the storm surge record as a time series of 
normalized seasonal and annual 95th and 99.9th per 
centiles together with low-pass-filtered versions of the 
time series in Fig. 4a,b, respectively. Figure 4c shows 
linear trends for four upper seasonal and annual per 
centile (95th, 98th, 99th, and 99.9th) time series. The time 
series are characterized by a considerable interannual-to- 
multidecadal variability, while there is no evidence for any 
significant long-term trend: neither for the seasonal nor for 
the annual percentile time series. Generally, surge levels 
are considerably higher and more variable during winter 
(both on intra-annual and interannual time scales) com 
pared to the summer season. Periods of particular high 
surges are found at the end of the nineteenth and twentieth 
centuries (note that the periods at the end of the nine 
teenth century are slightly different in timing between the 
95th and 99th percentiles but at the moment we are not 
able to explain these differences). In both cases the high 
rates are dominated by the winter season. Between both 
peaks, the highest percentiles of surges especially are 
marked by a gradual decline until the mid-1960s, as noted 
earlier for central European storminess (Matulla et al. 
2008). After the mid-1990s maximum, surges returned to 
more moderate values.