Can different learning paths produce better estimates in empirical asset pricing via Machine Learning?

This paper demonstrates how a portfolio composition technique can leverage the power of machine learning approaches in empirical asset pricing, using a very large set of identical Neural Networks (NN) and differentiating them only by the initial set of parameters. We implement the portfolio employing a trivial rule of ensemble, demonstrating how the variety generated by the initial conditions of the Neural Networks can produce better results than the average. This approach shed a new light on the potential application of ensemble methods to outperform a single NN involved in portfolio construction strategies, using more complex rules to extract the information discovered by the different training paths of identical NN.