Improving Market Risk Management with Heuristic Algorithms

Recent changes in the regulatory framework for banking supervision increase the regulatory oversight and minimum capital requirements for financial institutions. In this thesis, we research active portfolio optimisation techniques with heuristic algorithms to manage new regulatory challenges faced in risk management. We first study if heuristic algorithms can support risk management to find global optimal solutions to reduce the regulatory capital requirements. In a benchmark comparison of variance, Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR) objective functions combined with different optimisation routines, we show that the Threshold Accepting (TA) heuristic algorithm reduces the capital requirements compared with the Trust-Region (TR) local search algorithm. Secondly, we introduce a new risk management approach based on the Unconditional Coverage test to optimally manage the regulatory capital requirements, while avoiding to over- or underestimate the portfolio risk. In an empirical analysis with TA and TR optimisation, we show that our new approach successfully optimises the portfolio risk-return profile and reduces the capital requirements. Next, we analyse the effect of different estimation techniques on the capital requirements. More specifically, empirical and analytical VaR and CVaR estimation is compared with a simulation-based approach using a multivariate GARCH process. The optimisation is performed using the Population-Based Incremental Learning (PBIL) algorithm. We find that the parametric and empirical distribution assumption generate similar results and neither of them clearly outperforms the other. However, portfolios optimised with the simulation approach reduce the capital requirements by about 11%. Finally, we introduce a global VaR and CVaR hedging approach with multivariate GARCH process and PBIL optimisation. Our hedging framework provides a self-financing hedge that reduces transaction costs by using standardised derivatives. The empirical study shows that the new approach increases the stability of the portfolio while avoiding high transaction costs. The results are compared with benchmark portfolios optimised with a Genetic Algorithm.