André Philipp Mundt Dynamic risk management with Markov decision processes Dynamic risk management with Markov decision processes von André Philipp Mundt Dissertation, Universität Karlsruhe (TH) Fakultät für Mathematik, 2007 Referenten: Prof. Dr. Nicole Bäuerle, Prof. Dr. Ulrich Rieder Impressum Universitätsverlag Karlsruhe c/o Universitätsbibliothek Straße am Forum 2 D-76131 Karlsruhe www.uvka.de Dieses Werk ist unter folgender Creative Commons-Lizenz lizenziert: http://creativecommons.org/licenses/by-nc-nd/2.0/de/ Universitätsverlag Karlsruhe 2008 Print on Demand ISBN: 978-3-86644-200-9 Preface During almost four years of work at the Kompetenzzentrum Versicherungswissen- schaften GmbH, Leibniz Universita¨t Hannover, and the Institut fu¨r Stochastik, Universita¨t Karlsruhe (TH), many people supported me. First of all, I would like to thank my supervisor Nicole B¨auerle. She draw me to the subject of dynamic risk management and made many fruitful suggestions for the research treated in this work. Furthermore, she always took time for me and encouraged me when I was stuck and came to her with questions and problems. Secondly, let me thank Ulrich Rieder for being the second advisor and for useful discussions. From all the other people, I firstly and most importantly have to mention Anja Blatter. She read almost the whole manuscript and made many valuable comments that led to improvements in formulations and the layout. Furthermore, I have to thank Gunther Amt and Bruno Ebner for helping me improve parts of the thesis and Mirko K¨otter and Lars Michael Hoffmann for useful discussions. Altogether, my colleagues and former colleagues from the Kompetenzzentrum Versicherungswissenschaften GmbH, the Institut fu¨r Mathematische Stochastik in Hanover and the Institut fu¨r Stochastik in Karlsruhe always provided a friendly and enjoyable atmosphere at work and during working breaks. When I moved to Hanover and Karlsruhe they made it very easy for me to settle in there. Finally, I have to mention my family and in particular my parents for supporting my studies and their confidence in my ability to complete my diploma and this thesis. Karlsruhe, October 2007 Andr´e Mundt v Contents Introduction ix 1 Static and conditional risk measures 1 1.1 Model and definition . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.3 Further properties and representations . . . . . . . . . . . . . . . . 8 1.4 Remarks on conditional risk measures and measurability . . . . . . 11 2 Portfolio optimization with constraints in a binomial model 15 2.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.2 Risk minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 2.2.1 The unconstrained case . . . . . . . . . . . . . . . . . . . . . 19 2.2.2 A constraint on the final value . . . . . . . . . . . . . . . . . 21 2.3 Utility maximization . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.3.1 Comparison with the risk minimization . . . . . . . . . . . . 34 2.3.2 The unconstrained case . . . . . . . . . . . . . . . . . . . . . 36 2.3.3 Intermediate constraints . . . . . . . . . . . . . . . . . . . . 37 3 Dynamic risk measures 51 3.1 An overview on the literature . . . . . . . . . . . . . . . . . . . . . 51 3.2 Definitions, axioms and properties . . . . . . . . . . . . . . . . . . . 55 3.3 Stable sets of probability measures . . . . . . . . . . . . . . . . . . 62 4 A risk measure by Pflug and Ruszczyn´ski 67 4.1 Definition of the dynamic risk measure . . . . . . . . . . . . . . . . 67 4.2 Properties of the dynamic risk measure . . . . . . . . . . . . . . . . 70 4.3 Solution via Markov decision processes . . . . . . . . . . . . . . . . 73 4.4 A stable representation result . . . . . . . . . . . . . . . . . . . . . 84 4.5 Martingales & co. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 vii 5 A Bayesian control approach 93 5.1 The model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5.2 A Bayesian control approach to dynamic risk measures . . . . . . . 94 5.3 Explicit solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 5.3.1 The case when the parameter ϑ is known . . . . . . . . . . . 101 5.3.2 Beta distributions as initial distribution . . . . . . . . . . . 102 5.4 Comparison of value functions . . . . . . . . . . . . . . . . . . . . . 109 5.4.1 A comparison result for general distributions . . . . . . . . . 109 5.4.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 A Absolutely continuous probability measures 125 B The Beta distribution 129 Bibliography 131