The issue is that the Mean-Variance (MV) formulation for portfolio optimization leads to a quadratic programming problem. GLPK is for linear programs, thus the quadratic objective cannot be directly introduced to GLPK. For the first part of this assignment, however, you can work around the issue by using the first order necessary conditions that define a solution for the MV portfolio.

Look at page 30 of the notes for Lecture 7 entitled "General Case". There you will find linear equations which constrain solutions to the MV problem. Two of these are simple equality constraints that can be directly translated into GMPL using portfolio weights as the decision variables. The remaining

*n*linear equations involves the weights and two additional decision variables for the Lagrange multipliers. Adding those two decision variables to the model allows you to translate these equations into a system of GMPL constraints. You won't need an objective function because the linear constraints are sufficient to determine a unique solution for the MV portfolio.

For your convenience, I've put a GMPL data file called PortfolioMVO.dat on the downloads list with the necessary data for Assignment 7.

The trouble you'll face is that the second part of Assignment introduces an additional inequality constraint. This change requires you to go back and establish KKT necessary conditions that include this new constraint. In the end you will have have a somewhat more complex GMPL model with additional decision variables and some discrete logic. If you're interested in doing this, let me know and I'd be glad to work you through the mechanics.

## No comments:

## Post a Comment