Systemic Risk Based Portfolio Selection

Abstract

This dissertation examines portfolio selection under systemic risk using performance measures. In the first chapter, we propose a novel performance measure to construct optimal portfolios that explicitly incorporate the occurrence of systemic event. Investors maximize an ex-ante modified Sharpe ratio that is conditional on some systemic event, with the latter interpreted as a low market return environment. We solve the portfolio optimization problem analytically under the absence of short-selling constraint and numerically when short-selling constraint is imposed. The approach is made operational by embedding it in a multivariate dynamic setting via dynamic conditional correlation and copula models. In the second chapter, we further enhance the portfolio selection approach proposed in the first chapter by using machine learning techniques. Specifically, the optimal portfolio is solved through a three-step supervised learning model. First, the smooth pinball neural network is employed to predict conditional marginal return distribution. Secondly, we use copula to model dependence between portfolio assets and the market, based on which we generate return scenarios. Lastly, we maximize the ex-ante conditional Sharpe ratio based on simulated returns. Unlike the previous chapter, where we use statistical models to forecast return distributions, in this chapter we take advantage of a distributional machine learning model along with a set of predictors that includes more than 1,000 predictive signals. In the last chapter, following the similar idea of conditional Sharpe ratio, we propose another systemic risk-based performance measure namely the conditional Rachev ratio. This measure inherits the advantage of unconditional Rachev ratio in the sense that it can account for asymmetric information of portfolio return distribution. Moreover, we build a link between our new measure and the well-know CoVaR measure in the finance literature. In each chapter, we construct a comparative analysis using data on the US stock market. Overall speaking, all the backtesting results demonstrate the superiority of our proposed approaches against popular benchmark strategies in terms of profitability and systemic risk, where the outperformance is robust to the inclusion of transaction costs.