Robust investment portfolio management for dynamic financial markets using Bayesian neural networks

Portfolio management plays a vital role in financial investment decisions. However, uncertainties in asset returns and estimation models pose challenges to efficient portfolio selection in dynamic financial markets. In response, we propose a robust reward-risk-adjusted portfolio management model based on a Bayesian machine learning inference framework to flexibly achieve optimal portfolio strategy. Specifically, we employ the Stochastic Gradient Variational Bayesian (SGVB) method to address the uncertainties in the asset return probability distribution parameters and the portfolio model structure. Moreover, SGVB is integrated with neural networks to form a probabilistic deep learning framework, where model latent variables are modeled as probability distributions and efficient portfolio model training is achieved through stochastic gradient optimization. We utilize the financial market data from US DJIA stocks to analyze the superiority and robustness of our proposed portfolio approach. Empirical analysis indicates that the proposed reward-risk-adjusted portfolio model achieves higher portfolio returns, Sharpe ratios, and Sortino ratios performance than other models under different market factors, including market turbulence periods like COVID-19. These findings assist investors and financial institutions in performing and analyzing optimal portfolio decision-making under volatile market environments.