Code with multiple constraints in Proc Optmodel, running slow, trying to improve efficiency

[opti_miser]

I'm trying to run the optimization pasted below in optmodel.  The program is running, whereas earlier I was getting an out of memory message.  However, it is very slow.  I tried to do what I could to make the code more efficient.  Any other ideas?

On average, i=3,000.  t=465 and k=3.  One problem may be that I need to use 465 constraints, but I have to have these constraints in place.  I can run a similar program in Proc IML in a fraction of the time, but Proc Optmodel is so flexible that I really need to migrate as much of my coding as possible. Thank you for whatever ideas you can offer.t

proc optmodel PRINTLEVEL=0;

    *create a numerical index based on permno and date, read in below;

    set <num,num> STOCKS_DATES;

    *create another numerical index just based on dates;

    set DATES = setof {<i,t> in STOCKS_DATES} t;

    *create another numerical index just based on permnos;

    set STOCKS {t in DATES} = slice(<*,t>, STOCKS_DATES);

    *create a numerical index based on the characteristics upon which portfolio weights are based;

    set CHARACTERISTICS = {'alpha', 'beta', 'stdfirm'};

    *name variables that will be read in from the data set;

    num char {STOCKS_DATES, CHARACTERISTICS};

    num ret {STOCKS_DATES};

    *read in the sas data set, specifying that each column in the char matrix is a different characteristic, k;

    read data foriml into STOCKS_DATES=[permno obs]

        {k in CHARACTERISTICS} <char[permno,obs,k]=col(k)>

        ret;

    var Theta {CHARACTERISTICS} init 0;

    impvar portmean = (1/card(DATES)) * (sum {t in DATES} (sum {i in STOCKS} ((1/card(STOCKS)) + (1/card(STOCKS)) * sum {k in CHARACTERISTICS} Theta * char[i,t,k]) * ret[i,t]));

    impvar portvar = (1/card(DATES)) * (sum {t in DATES}(((sum {i in STOCKS} ((1/card(STOCKS)) + (1/card(STOCKS)) * sum {k in CHARACTERISTICS} Theta * char[i,t,k]) * ret[i,t])-portmean)**2));

    impvar sumneg_w {t in DATES} = sum {i in STOCKS} min(((1/card(STOCKS)) + (1/card(STOCKS)) * sum {k in CHARACTERISTICS} Theta * char[i,t,k]),0);

    con short50 {t in DATES}: sumneg_w >= -0.5 ;

    max Objective1 = portmean - (&y/2) * portvar;

    Problem Prob1 Include Theta short50 Objective1;

    USE PROBLEM Prob1;

    solve with NLP OBJ Objective1 / ms;

[RobPratt]

You might try making portmean an explicit variable instead, to reduce the density of the expression for portvar.  Replace the impvar portmean statement with:

var portmean;

con portmean_con:

     portmean = (1/card(DATES)) * (sum {t in DATES} (sum {i in STOCKS} ((1/card(STOCKS)) + (1/card(STOCKS)) * sum {k in CHARACTERISTICS} Theta * char[i,t,k]) * ret[i,t]));