Asset returns in deep learning methods: An empirical analysis on SSE 50 and CSI 300

In the field of finance, the application of deep learning methods to asset returns can yield more intuitive results than those derived using standard methods. We use the daily returns from the Shanghai Stock Exchange (SSE) 50 and China Security Index (CSI) 300 Index in the Chinese stock market to compare both standard linear methods (ordinary least squares, stepwise regression, principal components regression (PCR), elastic net (ENet)) and deep learning neural network (NN) methods with one, two, and three hidden layers (NN1, NN2, and NN3, respectively) from January 2012 to December 2017. We show that for both SSE 50 and CSI 300 indices, not all NN models outperform linear models in terms of out-out-of-sample Roos2 values in the 2017 test dataset. However, the NN models with one and three hidden layers do not outperform the classical PCR model in terms of out-of-sample Ros2 values for the SSE 50 Index, whereas all the deep learning methods outperform the traditional models for the CSI 300 Index. The NN model with two hidden layers performs the best in terms of the out-out-of-sample R2 values for the CSI 300 Index. In the out-out-of-sample dataset, NN3 performs poorly for both SSE 50 and CSI 300 indices, while the deepest NNs are not the best performing.