Bayesian Neural Networks with Dependent Dirichlet Process Priors. Application to Pairs Trading
Gheorghe Ruxanda, OPINCARIU SORIN · ECONOMIC COMPUTATION AND ECONOMIC CYBERNETICS STUDIES AND RESEARCH · 2018
Bayesian neural networks combine the universality of the neural networks with the principled uncertainty quantification of the Bayesian approach.The black-box character of neural networks makes it difficult establishing appropriate priors for the weights of the neural network.In this paper we propose a hierarchical model where the prior distribution of the network weights is drawn from a Dirichlet process mixture model.We further extend the model to dependent Dirichlet process mixtures to allow the model to account for non-stationarity in the data.The neural network with dependent Dirichlet priors is used to model a pairs trading experiment.