MSDF in time-series prediction using delay coordinate embedding

.A powerful forecasting method working with few assumptions was developed in the framework of chaos theory It uses reconstruction of process dynamics through time delay embedding and is usually applied on univariate time series. Mixing simultaneous measurements and delay coordinates of several variables in the embedding vector allows to fuse data from multiple sources. This paper presents the corresponding method for multivariate time series prediction. This novel method is tested in prediction of computer generated series and electrical load demand. Effects on performance and application details are described. The method allows to easily estimate sensitivity of the forecasting accuracy to past, current, and future values of exogenous variables. The inherent parallelism and lack of training provide certain computational advantages in comparison to other nonlinear black-box predictive models.