Machine Learning of Jump Dynamics in US Dollar-Ghana Cedi Exchange Returns

Abstract

An algorithm to filter the jumps in the US Dollar-Ghana Cedi log-returns has been implemented with the help of machine learning tools in this paper. The algorithm is like a classifier. For a frontier market there is no known standard algorithm to filter and classify jumps. This new algorithm was developed to extract the jump dynamics to feed the new equation for modelling the US Dollar-Ghana Cedi log returns, the Correlated Multifactor Stochastic Variance Jump Diffusion (CMSVJD). The algorithm is based on the three-sigma rule of leverage statistics but here the recursion is truncated using a tolerance convergence level defined in a classifier or with the help of a convergence graph. The justification of the algorithm has been established and was implemented with the US Dollar-Ghana Cedi log-returns as a Lebesgue unimodal data. The jump dynamics show that the jump size and jump time are random and independent; jump size is time-invariant and jump time intervals obey the Markov chain property of time inhomogeneity. There were more positive jumps (depreciation of the cedi) than negative jumps (appreciation of the cedi) but the magnitude of the negative jumps is higher than that of the positive jumps.