I primarily use wavelets because Fast DWT (wavelet) algorithm is much more computationally efficient than SVD ( O(N) compared to O(M*N) ) and the improvement offered by SVD is marginal in most cases.

SVM and SVD are very powerful techniques though, to be sure. You can perform real-time training of SVM using CUDA, see cuSVM which has a cuda optimized c library and matlab wrapper.

It is pretty well documented that SVM outperforms Neural net (simple feedforward NN anyways) for many tasks but NN is generally more flexible and lends itself better to hybrid techniques such as neuro-fuzzy as well as ensemble techniques.

My first day here and I really enjoy this thread. Lets hope that NT7 comes out with future bars plotting..