Skip to content

Best optimizer for large scale problems

5 messages · Bastian Offermann, Dominykas Grigonis, Alexios Ghalanos +2 more

#
Hi all,
I am working on a large scale portfolio optimization problem with up to 
500 assets. My objective function is simple

w*returns - 1/2 * 1/constant * w * Matrix * w

subject to sum(w) == 1 and w is a vector of weights with 0 <= w[i] <= 1 
for all i = 1, ..., n.

I have tried quadprog, alabama and DEoptim. What are your experiences 
with those and possibly other options?
Thanks in advance!
#
For quadratic problems you really should use quadprog.

You can also try the 'parma' package which provides a nice interface to 
frame your problem and constraints and then solve it using either an LP, 
QP, NLP (with analytic derivatives) or GNLP formulation (depending on 
the intersection of problem and constraint type).

Regards,

Alexios
On 06/06/2013 08:58, Dominykas Grigonis wrote:
#
On Thu, 06 Jun 2013, Bastian Offermann <bastian2507hk at yahoo.co.uk> writes:
You say you "have tried" these packages, so did you encounter any
specific problems?  If yes, you should post some examples that
demonstrate these problems.

quadprog sounds reasonable, but if your covariance matrix is not
full-rank, quadprog's solve.QP will not work.  (Which is actually more
an empirical than a computational problem.  For example, there might be
no unique solution.)  Other methods, such as Differential Evolution, do
not have such a constraint.

Regards,
        Enrico