We introduce a forecasting technique based on multivariate ideas previ
ously applied in remote sensing. The approach has the trivial but none
theless fundamental purpose of dividing the information inherent in th
e time series into important and unimportant. Important information is
used for forecasting purposes while the unimportant is discarded. Alt
hough related to vector autoregression, giving asymptotically the same
estimates, there are reasons to believe that the approach gives bette
r precision of parameter estimates for finite samples as well as more
precise predictions.