Granger causality in machine learning model : application to KOSPI and Asian stock market indices

Granger causality test is introduced into machine learning model. Following Granger (1969), prediction errors are compared for the test. Since no formal statistical test is established in machine learning, non-parametric Wilcoxon signed rank test is used to derive statistical significance. Causality relationships between KOSPI and major Asian stock market indices are analyzed. From the comparisons of vector autoregression (VAR) and neural network models, two major findings are derived. First, neural network model identifies more causality relationships. Second, the Granger causalities between KOSPI and one of the Asian stock indices are more transparent in neural network model than in time series model. These results may be due to the fact that neural network model incorporates almost all possible shapes of functional relationships between target and feature variables. The neural network model finds more predictability because it extracts all available predictive sources from feature variables.