چكيده لاتين
One of the methods for optimizing the weights of assets within an investment portfolio is the mean-variance method, where the input data consists of the assetsʹ vector of expected returns and their variance-covariance matrix. To implement this method, the vector of expected return and the variance-covariance matrix of the assets must be estimated. However, in practice, due to estimation error of the input, this method has not been able to produce satisfactory results compared to some simple diversification strategies such as equal weighting. In this study, it is attempted to estimate the expected return of the assets in the investment portfolio by using a recurrent neural network (RNN), and then the out-of-sample performance of this method are compared with the method that use the historical average return as the expected return. Finally, it is examine whether the method used would have improve the estimation error and, consequently, the performance mean-variance optimization method.
The test assets used in this study are the long andshort legs of famous investment factors in the US stock market, which include size, value, profitability, investment, and momentum factors as well as the market factor. Using economic and technical variables, the expected returns of these factors will be predicted on a weekly basis. The economic variables that are used for predicting returns include the output gap, short-term interest rate, risk-free interest rate, gross domestic product growth, past market returns, term and credit spreads of interest rates, and the dividend-price ratio. Additionally, the technical variables used include the simple moving average (SMA), exponential moving average (EMA), relative strength index (RSI), and moving average convergence-divergence (MACD).
The results indicate that the methdos used have outperformed the historical average in terms of predictive power for expected returns, achieving an out-of-sample R-squared of up to 7.2% for predicting the factor reutns. Furthermore, after applying the predicted returns in the mean-variance method, both the average returns and the Sharpe ratio have improved compared to using the historical average. Additionally, if the predicted negative returns from the model in the mean-variance method are replaced with zero, along with a minimum and maximum weight constraint of -0.4% and 0.4%, the strategy has significantly outperformed the equal-weight strategy in terms of both the average returns and the Sharpe ratio.
