Portfolio Selection with Convolutional Neural Network

In this work, we applied Convolutional Neural network (CNN) models to select the portfolio weights that lead to the highest realized return in one time-step ahead. In fact, given four possible portfolio optimization methods, the CNN is used to forecast the one-step-ahead returns of the assets in the portfolio implicitly and selects the portfolio optimization approach leading to the highest portfolio return. We construct four different datasets based on the daily return time series of the 36 stocks in our portfolio. In each dataset, one instance is composed of the mean vector and the covariance matrix. The four datasets are obtained as a result of using four different methods to calculate mean vectors and covariance matrices. Six different CNN model architectures are constructed, and the models' performances are compared. The obtained results demonstrate the effectiveness of CNNs in portfolio selection.