A hierarchical Deep neural network design for stock returns prediction

We present in this paper a hierarchical Deep Neural Network for stock returns prediction. This DNN is trained in a high frequency context, we use 5 minutes returns of TUNINDEX stocks in a period of 4 years. The designed network aims to predict the next 5 minutes return of a given stock. The predictive power of our network is improved by the hierarchical design and stocks classification while the training process is simplified by dimensionality reduction techniques. Experimental study shows an accuracy up to 71% and a considerable improvement comparing to recent related works.