Not seeing any obvious correlelations, we throw all the data into the s.v.d. and neural network programs. 2/3 of the records are used for training, 1/3 for verification. Using a neural network, with 18 predictors, there is evidence of "over-fitting" with the training data. The over-fitting problem is not evident when the number of predictors is reduced to the best seven predictors identified from s.v.d.. But note s.v.d. predicts the verification data almost as well as the neural network does.
skill scores

6 surface obs

6 surface obs +
12 reanalysis variables

best 7 predictors


1: n.n. training

2: n.n. verification

3: s.v.d. training

4: s.v.d. verification