Banking risk indicators, machine learning and one-sided concentration inequalities.

Summary This PhD thesis includes three essays on the implementation, and where appropriate the improvement, of financial risk measures and bank risk assessment based on machine learning methods. The first chapter develops an elementary formula, called E2C, for estimating credit risk premiums inspired by CreditGrades, and improves its accuracy with a decision tree forest algorithm. Our results highlight the prominent role played by this estimator and the additional contribution of the financial rating and the size of the firm considered. The second chapter infers a one-sided version of the inequality bounding the probability of a unimodally distributed random variable. Our results show that the unimodality assumption for stock returns is generally admissible, allowing us to refine bounds on individual risk measures, to comment on the implications for extreme risk multipliers, and to derive simplified versions of bounds on systemic risk measures. The third chapter provides a decision support tool grouping rated banks by risk level based on an adjusted version of the k-means algorithm. This fully automated process is based on a very large universe of individual and systemic risk indicators synthesized into a subset of representative factors. The results obtained are aggregated by country and region, offering the possibility to study areas of fragility. They highlight the importance of paying particular attention to the ambiguous impact of bank size on systemic risk measures.