| Abstract [eng] |
Following the global financial crisis, researchers and practitioners have paid close attention to risk-based asset allocation strategies, which do not depend on the calculation of expected returns and are therefore viewed as more stable than the standard mean-variance framework. However, as the number of investable assets grows, so does the complexity of the optimum asset allocation problem. Traditional risk-based allocation techniques involve the inversion of a potentially ill-conditioned covariance matrix, which in turn results in an amplification of estimation errors. The optimization problem also becomes computationally challenging when the portfolio consists of more than a few hundred assets. In order to solve the aforementioned problems, modern portfolio optimization methods have introduced clustering-based allocation approaches. In this thesis, we examine various risk-based optimization strategies, clustering algorithms, and covariance matrix estimation methods in terms of their contribution to portfolio risk and risk-adjusted returns. The empirical study is performed on 100 randomly sampled 350-asset portfolios featuring realistic diversification across 11 sectors. Based on overall risk characteristics and risk-adjusted performance, this thesis suggests a combination of nested global minimum variance optimization, partitioning around medoids with dynamic time warping distance, and a Markov switching multifractal model with dynamic conditional correlation type structure and nonlinear shrinkage. Nonetheless, this choice heavily depends on the investor’s risk profile as well as desired portfolio turnover and weight concentration. |
