Recovering Causal Networks based on Windowed Granger Analysis in Multivariate Time Series

Reconstruction of causal network from multivariate time series is an important problem in data science. Regular causality analysis based on Granger method does not consider multiple delays between elements of a causal network. In contrast, the Windowed Granger method not only considers the effect of mutiple delays, but also provides a flexible framework to utilize various linear and nonlinear regression methods within Granger causality analysis. In this work, we have used four methods with Windowed Granger method including hypothesis tests of linear regression, LASSO and random forest. Then, their performance on two simulated and real-world time series are compared with ground truth networks and other causality recovering methods.