Modelling and trading commodities with a new deep belief network

Abstract

The scope of this project is to study a novel methodology in the task of forecasting and trading the crack spread modelled index. More specifically in this research we are expanding the earlier work carried out by Karathanasopoulos et al. (2016c) and Dunis et al. (2005) who model the Crack Spread with traditional neural networks. In this research paper we provide for first time a more advanced approach to non-linear modelling and trading the ‘Crack’. The selected trading period covers 4500 trading days and the proposed model is a deep belief network (DBN). To model, test and evaluate the crack spread we use an expansive universe of 500 inputs correlated with the main index. Moreover we have used for reasons of comparison a radial basis function combined with partial swarm optimizer and two linear models such as random walk theorem and buy and hold strategy.