Some remarks on value-at-risk optimization∗

We discuss on observations related to value-at-risk optimization. Firstly, we consider a portfolio problem under an infinite number of value-at-risk inequality constraints (modeling first order stochastic dominance). The random data are assumed to be normally distributed. Although this problem is necessarily non-convex, an explicit solution can be derived. Secondly, we provide a (negative) result on quantitative stability of the value-at-risk under variation of the random variable. Although reduced Lipschitz properties (in the sense of calmness) may hold true at continuously distributed random variables under suitable conditions, the result shows that no full Lipschitz property (more generally: Hölder property at any rate) can hold in the neighbourhood of an arbitrary continuously distributed random variable. Even worse, this observation holds true with respect to any probability metric weaker than that of total variation.