Estimating structure of multivariate systems with genetic algorithms for nonlinear prediction.

Although we can often observe time-series data of many elements, these elements do not always interact with each other. This paper proposes a scheme to estimate the interdependency among observed elements only by time-series data, which is useful for selecting essential elements to optimize multivariate prediction model. Because this estimation is a sort of combinatorial optimization problems, we applied the genetic algorithm as a method to moderate this problem. Through some simulations, we confirmed performance of our method, which can identify interaction of multivariate system and can improve its prediction accuracy. Especially, our method can be applied to predict real foreign-exchange markets even if system has nonstational property and its structure changes dynamically.