PORTFOLIO OPTIMIZATION WITH MACHINE LEARNING: RELEVANT PERSPECTIVES ON IMPROVING THE BLACK-LITTERMAN MODEL

Abstract

The primary focus of this thesis is experimenting with different machine learning algorithms for generating investor views for the Black-Litterman model in the form of financial forecasts and comparing the performance of the created portfolios with other established models from the literature. The thesis also explores domains of investor sentiment, clean energy assets, the low-risk anomaly, and diverse estimation techniques in constructing portfolios. In a nutshell, we present three novel perspectives on utilizing the Black-Litterman model. First, we use daily price and technical indicators' data for the ten MSCI Asia Pacific sector indices for the past 20 years and find that our hybrid multivariate Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM) deep learning model gives reasonably better predictions when predicting index closing prices out-of-sample than using either CNN or BiLSTM alone. After utilizing these predictions as investor views inside the Black-Litterman model with time variation in the conditional distribution of returns, we find that the portfolios generated outperform all benchmark model portfolios by a considerable margin regarding financial efficiency and diversification. Second, we introduce a new dimension in constructing relative investor views for the Black- Litterman model by incorporating fear/greed technical indicator predictions as a proxy for investor sentiment in the portfolio construction process. We apply a hybrid complete ensemble empirical mode decomposition with adaptive noise - gated recurrent unit (CEEMDAN-GRU) deep learning model to forecast this indicator and the extreme gradient boosting (XGBoost) ensemble learning algorithm to forecast returns for ten country ETFs and create relative views for the Black-Litterman model. These models beat several benchmark forecasting models. Our empirical results show that the proposed approach outperforms the Markowitz, minimumvariance, equally-weighted, and risk-parity strategies along with four other Black-Litterman approaches from the literature for six investment periods. Finally, we empirically test the presence of the low-risk anomaly in clean energy markets in the pre-Covid, during Covid, and post-Covid periods and also compare the performance of various clean energy portfolios against leading conventional energy market indices to uncover the benefits and costs of switching towards clean energy investments from a risk-return standpoint. We follow a novel approach in which parameters of three traditional econometric models (GARCH, EGARCH, and FIGARCH) are taken as features inside a compute unified device architecture enabled bidirectional long short-term memory (CuDNNBiLSTM) neural network to forecast the realized volatility of fifteen clean energy indices and classify them into various categories of risk as a means of asset preselection before portfolio optimization. We employ the Black-Litterman portfolio optimization model and use the random forest machine learning algorithm to forecast index returns and translate them into absolute Black-Litterman views. We find that the models used for forecasting beat several other established models from the literature. The Black-Litterman clean energy portfolios created by the proposed procedure surpass three other leading portfolio strategies and six benchmark market indices. We find strong evidence of the low-risk anomaly in all three time periods. This dissertation offers proof that financial forecasting based on machine learning has several advantages. It also confers that investor views from financial forecasting based on machine learning can greatly improve the Black-Litterman model.