This paper proposes a novel semiparametric time-varying model for long-horizon predictive regressions in which the coefficients are allowed to change over time with unspecified functional forms. A linear projection method is employed to deal with the embedded endogeneity issue. We pursue an indirect approach to estimate the long-horizon coefficients through the implication of the short-horizon coefficients, which is less subject to overlapping observations. Monte Carlo simulation illustrates that the nonparametric estimation method works well in finite and large samples. Empirically, the predictability of return or dividend growth changes significantly over time. The dividend-price ratio predicts stock return in most of our samples, but the dividend growth is much less predictable than the stock return. Moreover, the dividend-price ratio predicts either stock return or dividend growth, or both in a local period. Our findings are robust to several forecast horizons.