Multivariate time series prediction by blind signal deconvolution

This paper proposes a new method in a prediction problem of multivariate time series based upon blind deconvolution. The method firstly predicts a one-step-ahead signal and its volatility (covariance) by means of a scheme called VARMA-ICA, which has been recently developed for system identification with unknown inputs. The predicted signal and volatility are then used to minimize the risk of prediction. Furthermore, the paper applies the proposed method to a strategy for stock trade which deals with multiple brands. Numerical simulation illustrates the effectiveness of the method.