Font Size: 

In this contribution we aim at combining graphical models and neural network for predicting bank stock returns. We propose to leverage the information extracted from graphical models fitted on financial data timeseries as input features to a neural network to improve the predictive performance. We focus on the 74 largest listed U.S. banks over a period that spans from 2003 to 2017. We apply a recent and fast algorithm (LoGo) for the network estimation that allows us to calculate the banks' partial correlations network and relative indexes like mutual information and transfer entropy. The information obtained from the network are then leveraged as features in a Recurrent Neural Network and in other predictive models for comparison.