14 
Development results 
Table 7: Contributions of different groups of predictors to the reduction of error variance, broken down into phase 1 (T+1 hour) 
and phase 2 (T+3, 9 27, 33 hours) 
(1) 
(2) 
(3) 
(4) 
(5) 
(6) 
(7) 
(8) 
(9) 
(10) 
(11) 
(12) 
Fp 
RMSE 
RV 
RV 
RV 
RV 
Prs2D 
RV S ] 
P 1 i 
t 
RV 
/h 
DMO 
MOS 
MOS 
MSI 
Bk 
GFS + S 
korr 
HW/NW 
+ /- 
Tot 
(MOS,DMO) 
01 
127 
57 
80 
2 
55 
46 
17 
20 
10 
26 
85 
03 
128 
89 
52 
7 
1 
3 
45 
15 
15 
23 
63 
09 
134 
96 
49 
9 
- 
3 
40 
16 
16 
31 
65 
15 
136 
107 
40 
12 
- 
3 
30 
18 
18 
31 
59 
21 
145 
114 
38 
14 
- 
4 
24 
14 
18 
28 
56 
27 
151 
127 
31 
13 
- 
6 
15 
14 
26 
30 
52 
33 
160 
132 
35 
16 
- 
5 
15 
9 
20 
27 
53 
Ph2 
142 
111 
39 
12 
0 
4 
27 
14 
17 
28 
56 
Tot 
140 
103 
46 
11 
8 
10 
26 
15 
16 
28 
61 
The last two lines show the mean values of the RMSE and (RMSE weighted) RV values of phase 2 and 
corresponding mean values of both phases, respectively, i.e. all Fp considered. The individual columns 
include the following: 
Column 
2: 
Column 
3: 
Column 
4: 
Columns 5-8: 
Column 
5: 
Column 
6: 
Column 
7: 
Column 
8: 
Column 
9: 
Column 10 
Column 11 
Column 12: 
RMSE of DMO, mm. 
RMSE of MOS without classification, mm. 
Reduction of error variance of MOS without classification resulting from 
columns 2 and 3, compared to DMO; all RV data in percent. 
Percentage of the most important predictor groups in the total RV in column 4. 
Reduction of error varance of 1-predictor equation using only DMO as predictor, 
compared to pure DMO. 
Percentage of surge at Borkum. 
Percentage of all 200 potential predictors from GFS and MSWR-MOS base 
technology, which additionally include astronomical, harmonic, and binary 
predictors. 
Percentage of the correction of the last known initialisation error of DMO, 
already referred to above. 
Further reduction of variance versus column 4 by splitting into HW/LW. 
Further reduction of variance versus column 4 by splitting into surge+/surge-. 
Total reduction of variance by splitting into the 4 classes HW+, HW-, LW+ and 
LW-. It is only approximately equal to the RV total of columns 9 and 10 because 
the 8 MOS equations involved are all computed independently from each other. 
Total reduction of error variance RV (MOS, DMO) following classification as 
RV total of columns 4 and 11. 
It is obvious that most of the reduction of DMO error variance (on average 50 of 61 %) is due primarily 
to manipulation of DMO - not to additional external information such as observations or GFS prediction 
elements. In the long phase from T+3 to T+33 hours, this accounts for as much as 54 of 56% of the 
reduction of the error variance.