Current research on predictive regressions studies the case where for each t, Y t is regressed on a lagged predictor variable, X t-1 , and predictive series, {X t }, is of first-order autoregressive. The problem with these predictive models is that the OLS-estimated slope coefficient is biased in the finite sample case, when the errors of the autoregressive model for {X t } are correlated with the errors in the predictive regression model. However, there are predictor series having autoregressive structure of order greater than 1. Predictive regression with autoregressive predictors that are not necessarily of the AR(1) structure are quite common in finance and economics. So, a method that could reduce bias of estimators for coefficients in predictive regressions with order-p autoregressive predictors can be useful. In this thesis, we consider predictive model where for each t, Y t is regressed on X t-1 , X t-2 , …, X t-p and {X t } being AR(p) with p 1. Using generalized augmented regression method for this case, where the predictor variable is AR(p), a method has been proposed for having bias-reduced point estimation for the predictive coefficients and a corresponding hypothesis testing procedure. This method has been generalized to the case of multiple AR (p) predictors. We compare OLS and augmented regression methods in terms of the bias in estimating the predictive coefficients and in terms of the size of the statistical tests on hypothesis tests for the coefficients using simulation and empirical analysis. For New York Stock Exchange data, our method applied to a model in which quarterly stock returns are predicted by dividend yields shows that the predictor series is AR(2). For these data, dividend yield is a significant predictor of stock return not only based on OLS, but also based on the standard bias-correction method that assumes that the predictor series is AR(1). However, the predictor series is found to be AR(2). The result of applying this method was that the estimated predictor coefficients are insignificantly different from zero. Also in this thesis, Iran Stock Exchange data were evaluated based on two models. We showed that the predictability of stock return using dividend yield does not exist.