You can use many methods of multivariate analysis to portfolio optimization. However, it must be remember that the paradigm of portfolio risk assessment before the crisis was radically changed in the post-crisis period.

Economic interpretation of the portfolio optimization results must be done in accordance with the requirements under Basel III – new paradigm of portfolio risk assessment.

Try decision trees (J48 usually performs great results), can be easily applied for building a multi-class model.

There are two things that I would advise is:

a)As Sylantyev mentioned the time period is important due to crisis post-crisis period. you should test the data to see if the features correlate to the time periods. Also, see if shorter time periods have an effect on the data (spot cyclical or specific time periods which might bias the data)

b) At the stage that you are mentioning right now it is better to test several algorithms and not just one. Unless you have previous literature that already has done the work for you, I would suggest that you try several methods.

For portfolio optimization there are several works with Genetic Algorithms. This model allows to maximize distincts aspects keeping constraints with the others. In general gave better solutions than other models based in energy minimization (like Simulated Annealing)

Best!

SVMs are not a bad idea. Of course, for multi-class problems there are very good heuristics for SVM although not strong theory. Another very intuitive algorithm is learning vector quantization introduced by Kohonen. Mathematical theory is knwon for a variant called generalized learning vector quantization (GLVQ). Relevance learning for this algoritm (GRLVQ) provides you the additional information, which features provides most for good classification or which feature combination (matrix learning). If you are interested in detail, do not hesitate to contact me ...

~ Thomas