Title: Portfolio selection under systemic risk: A QRNN-based approach
Authors: Weidong Lin - Durham University (United Kingdom)
Abderrahim Taamouti - Durham University Business School (United Kingdom) [presenting]
Abstract: The aim is to improve the traditional mean-variance (MV) portfolio selection model by accounting for systemic risk and using machine learning techniques. Our objective is to formulate the portfolio selection as a three-step supervised learning problem, which allows for considering systemic risk when constructing optimal portfolios. In the first step, we use a quantile regression neural network (QRNN) to predict conditional quantiles for stock returns. Based on the obtained quantiles, we estimate the marginal distributions for individual assets and the market portfolio. In the second step, we use copula to model the dependence structure among assets and generate return scenarios. Lastly, we solve the portfolio optimization problem dynamically by maximizing a conditional Sharpe ratio (CoSR) based on the simulated return scenarios. Thereafter, we run several comparative studies using real data on big US financial institutions. The backtesting results demonstrate the superiority of our proposed portfolio over other benchmark portfolios. In particular, we compare the out-of-sample performance of our portfolio with those of: (i) a portfolio that maximizes the unconditional Sharpe ratio (SR); (ii) a Global Minimum Variance Portfolio (GMVP); and (iii) an equally weighted portfolio ($1/N$).