We consider the problem of multi-step ahead prediction in time series analysis using nonparametric smoothing techniques. Forecasting is always one of the main objectives in time series analysis. Recent research has shown that nonlinear time series models have certain advantages in multi-step ahead forecasting. Traditionally, nonparametric k-step ahead least squares prediction for nonlinear AR(d) models is done by estimating the k-step conditional mean function via nonparametric smoothing of X(t+k) on X(t),...,X(t-d+1) directly. In this paper we propose a multi-stage nonparametric predictor. We show that the new predictor has smaller asymptotic mean squared error than the direct smoother, though the convergence rate is the same. Hence, the proposed predictor is more efficient. Some simulation results, advice for practical bandwidth selection and a real data example are provided.