TITLE: Experiments With Robust Portfolio Optimization

SPEAKER: Professor Daniel Bienstock,
IEOR Department at Columbia University, NY

DATE/TIME: Friday, September 28, 2007 / 2:30-3:45 p.m.

LOCATION: Room 453 Mohler Lab, 200 W. Packer Avenue

ABSTRACT: In this paper we study portfolio optimization problems under histogram-like models of return errors. Robust optimization is a relatively recent approach at handling uncertainty in data for optimization problems – unlike stochastic programming, it does not assume an explicit distribution for data errors and instead allows data to be realized through an adversarial process. Robust optimization has been criticized for being overly conservative; but it can also prove not conservative enough. A key ingredient is the use of an appropriate uncertainty model. A 'right' model should attempt to achieve three goals which may in principle prove mutually incompatible: it should be faithful to the observed data, it should be agnostic with respect to structural assumptions on the data deviations, and, above all, it should prove sufficiently flexible so as to allow a decision-maker the ability to tune his or her risk aversion level.

Our models are data driven, and involve explicit non-convexities. Nevertheless, we show that using modern optimization techniques the models are quite practicable. We will also describe qualitative tests using various related models, and the empirical impact of robustness on the makeup of optimal portfolios.

BIOGRAPHY: Daniel Bienstock is a professor of Operations Research at the IEOR department at Columbia University, where he has been since 1989. His research encompasses theoretical and computational aspects of discrete and continuous optimization, and applications to logistics, finance, network design and routing. He received the Presidential Young Investigator award in 1990. He gave a plenary talk at the 2005 SIAM Conference on Optimization (Stockholm) and a semi-plenary talk at the 2006 Mathematical Programming Symposium (Rio). He is the author of over fifty papers in refereed journals, and a book, "Potential Function Methods for Approximately Solving Linear Programs.".


ALL FULL-TIME ISE DEPT. GRADUATE STUDENTS ARE REQUIRED TO ATTEND
REFRESHMENTS WILL BE SERVED FOLLOWING THE SEMINAR