The main aim of the present thesis is to investigate the effect of diverging priors concerning model uncertainty on decision making. One of the main issues in the thesis is to assess the effect of different notions of distance in the space of probability measures and their use as loss functionals in the process of identifying the best suited model among a set of plausible priors. Another issue, is that of addressing the problem of ``inhomogeneous" sets of priors, i.e. sets of priors that highly divergent opinions may occur, and the need to robustly treat that case. As high degrees of inhomogeneity may lead to distrust of the decision maker to the priors it may be desirable to adopt a particular prior corresponding to the set which somehow minimizes the ``variability" among the models on the set. This leads to the notion of Frechet risk measure. Finally, an important problem is the actual calculation of robust risk measures. An account of their variational definition, the problem of calculation leads to the numerical treatment of problems of the calculus of variations for which reliable and effective algorithms are proposed. The contributions of the thesis are presented in the following three chapters. In Chapter 2, a statistical learning scheme is introduced for constructing the best model compatible with a set of priors provided by different information sources of varying reliability. As various priors may model well different aspects of the phenomenon the proposed scheme is a variational scheme based on the minimization of a weighted loss function in the space of probability measures which in certain cases is shown to be equivalent to weighted quantile averaging schemes. Therefore in contrast to approaches such as minimax decision theory in which a particular element of the prior set is chosen we construct for each prior set a probability measure which is not necessarily an element of it, a fact that as shown may lead to better description of the phenomenon in question. While treating this problem we also address the issue of the effect of the choice of distance functional in the space of measures on the problem of model selection. One of the key findings in this respect is that the class of Wasserstein distances seems to have the best performance as compared to other distances such as the KL-divergence. In Chapter 3, motivated by the results of Chapter 2, we treat the problem of specifying the risk measure for a particular loss when a set of highly divergent priors concerning the distribution of the loss is available. Starting from the principle that the ``variability" of opinions is not welcome, a fact for which a strong axiomatic framework is provided (see e.g. Klibanoff (2005) and references therein) we introduce the concept of Frechet risk measures, which corresponds to a minimal variance risk measure. Here we view a set of priors as a discrete measure on the space of probability measures and by variance we mean the variance of this discrete probability measure. This requires the use of the concept of Frechet mean. By different metrizations of the space of probability measures we define a variety of Frechet risk measures, the Wasserstein, the Hellinger and the weighted entropic risk measure, and illustrate their use and performance via an example related to the static hedging of derivatives under model uncertainty. In Chapter 4, we consider the problem of numerical calculation of convex risk measures applying techniques from the calculus of variations. Regularization schemes are proposed and the theoretical convergence of the algorithms is considered.
Formulation and properties of a divergence used to compare probability measures without absolute continuity
