Learning Causal Graph: A Genetic Programming Approach

Representing causal relation between set of variables is a challenged objective. Causal Bayesian Networks has been classified as good modeling technique for this purpose. However structure learning for causal Bayesian networks still suffering from several problems including the causal interpretation of the model and the complexity of the learning algorithm. In this research the author presents an approach for learning causal graph based on Wiener-Granger causal-theory, with minor modifications, and use Genetic Programming to determine the parameters of Granger formula. This approach enjoys necessary advantages: reasonable complexity and cover nonlinear equation. A case study of 5 global stock markets is presented to experimentally explain and support this approach. The finding show that SP500 has Granger-causal influence on NIKKE: the accuracy of forecasting NIKKE stock market can be incremented by 24% when integrating past data from SP500. Whereas Euro STOXX 50 is reported to be the least stock Granger-causally affected by the others.