Robust Bayesian Neural Networks for the Prediction of Financial Assets Returns
Bayesian neural network, Hierarchical model, Student likelihood, Financial returns
The success of machine learning models and the vast amounts of financial that are automatically collected have led to an increased interest in applying the machine learning techniques in the prediction of the evolution of financial assets. Most machine learning are black boxes, there is no way to assess the model’s confidence in its results which is undesirable in risk sensitive domains like finance. In this paper we used the bayesian methodology as a principled way of building models with well calibrated uncertainty estimates. We discuss about the three types of uncertainty (structural, epistemic and random). We discuss on how one can construct bayesian extensions of one of the most successful black boxes: neural networks. While the literature focuses on the bayesian neural network with normal likelihood for general regression, we propose using a bayesian hierarchical model with a Student likelihood for financial regression. Doing so we account for the well-known stylized the non-gaussianity of the returns, making the regression more robust t outliers. We apply the robust bayesian neural network to the prediction of 10 time series and discussed the results obtained.