Title: Bayesian non-negative constrained regularised regresssion
Authors: Mauro Bernardi - University of Padova (Italy) [presenting]
Michele Costola - Ca' Foscari University of Venice (Italy)
Abstract: The aim is to propose a novel Bayesian approach to the problem of variable selection and shrinkage in high dimensional sparse linear regression models when the regression coefficients are also constrained to be positive. The regularisation method is an extension of the Lasso which has been recently cast in a Bayesian framework. Moreover, to deal with the additional problem of variable selection, we propose a Stochastic Search Variable Selection (SSVS) method that relies on a Dirac spike and slab prior where the slab component induces the sparse non-negative regularisation. The non-negative Lasso shrinkage and the SSVS algorithm are then extended to deal with the positive Elastic-Net penalisation and regressor selection. The methodologies are applied to the problem of passive index tracking of large dimensional index in stock markets without short sales. The tracking results indicate that non-negative-Lasso and Elastic Net bases SSVS are successful in asset allocation as compare to the classical solution and alternative Bayesian methods.