Construction of uncertainty sets for portfolio selection problems

While modern portfolio theory grounds on the trade-off between portfolio return and portfolio variance to determine the optimal investment decision, postmodern portfolio theory uses downside risk measures instead of the variance. Prominent examples are given by the risk measures Value-at-Risk and its coherent extension, Conditional Value-at-Risk. When avoiding distributional assumptions on the process that generates the risky assets' returns, historical return data or expert knowledge remain the only data available to the investor. His problem is then to maximize the return of his portfolio given the risk constraint that his portfolio does not fall short of some threshold return. For the Conditional Value-at-Risk, the solution is known to be achievable by a linear program. This paper extends the solution to the investor's problem whenever his risk preferences are given by any coherent distortion risk measure. More precisely, it is shown that whenever the risk constraint is given by a coherent distortion risk measure, a linear program leads to the solution. A geometric interpretation of this solution is immediate, which is related to the non-parametric description of data by socalled weighted-mean trimmed regions. The connections of the solution to robust optimization and decision theory are illustrated.