In this paper, we propose a semiparametric procedure called the “Model Averaging MArginal Regression” (MAMAR) that is flexible for forecasting of time series. This procedure considers approximating a multivariate regression function by an affine combination of one-dimensional marginal regression functions. The weight parameters involved in the approximation are estimated by least squares on the basis of the first-stage nonparametric kernel estimates of the marginal regressions. Under some mild conditions, we have established asymptotic normality for the estimated weights and the regression function in two cases: Case I considers that the number of the covariates is fixed while Case II allows the number of the covariates depending on the sample size and diverging. As the observations are assumed to be stationary and near epoch dependent, the approach developed is applicable to both the estimation and forecasting issues in time series analysis. Furthermore, the method and result are augmented by a simulation study and illustrated by an application in forecasting the high frequency volatility of the FTSE100 index.