Abstract
Goldberger [1] derived a Best Linear Unbiased Predictor in a case of correlated error. For the case when the model is not necessary true, we first evaluate the goodness of the predictor of Goldberger in the expectation of Prediction Mean Square Error with respect to future sample point, and compare it with the ordinary predictor. We modify the predictors by changing the Generalized Least Squares estimator of regression parameters. We also propose a criterion for selection of predictors and variables when these two modified predictors are used.
