A major use of linear regression models is to predict the future. An improved shrinkage estimator of the regression coefficients leads to a better of the dependent (response) variable in terms of lower Prediction mean squared error. Therefore, in this paper we consider the Problem of improved estimation of the unknown coefficients of a linear Regression model under usual normality assumption. It is well known that the ordinary least squares (OLS) estimator of the regression coefficients can be dominated by the Stein-rule estimators which are again dominated by their "Positive-part" versions. In this paper we show that the estimators can be dominated by a new type of estimators which are quite different from the "positive-part" estimators.