A Neural Network Approach for Generating Investors’ Views in the Black-Litterman Model

Abstract: This thesis investigates how neural networks can be used to produce investors' views for the Black-Litterman market model. The study uses two data sets, one with global stock market indexes and one with stock market data from the S&P 500. The task of the neural networks is to produce forecasts for the returns for the next quarter and the following year. The neural network will have to predict whether the market will move up or down and determine if the market movement is less than or equal to one standard deviation, creating four different scenarios. The forecasts are used as input to the Black-Litterman model to generate new portfolios, which are backtested from 2017 until 2022. The index data set was compared to a benchmark portfolio and a portfolio with naive risk diversification, while the S&P 500 data set was compared to market capitalization-weighted and naive portfolios. This resulted in eight different backtests where the neural networks obtained AUC values in the range of 0.56-0.73 and prediction accuracies in the range of 20.9% - 42.1%. The network used for yearly predictions on the index data set was the only network to outperform the benchmark portfolio. It obtained a Sharpe ratio of 1.782, a Sortino ratio of 2.165, and a maximum drawdown of -30.9% compared to the benchmark portfolio, where the corresponding metrics were 1.544, 1.879, and -32.8%.