We consider a class of nonparametric time series regression models in which the regressor takes values in a sequence space. Technical challenges that hampered theoretical advances in these models include the lack of associated Lebesgue density and difficulties with regard to the choice of dependence structure in the autoregressive framework. We propose an infinite-dimensional Nadaraya-Watson type estimator, and investigate its asymptotic properties in detail under both static regressive and autoregressive contexts, aiming to answer the open questions left by Linton and Sancetta (2009). First we show pointwise consistency of the estimator under a set of mild regularity conditions. Furthermore, the asymptotic normality of the estimator is established, and then its uniform strong consistency is shown over a compact set of logarithmically increasing dimension with respect to $\alpha$-mixing and near epoch dependent (NED) samples. We specify the explicit rates of convergence in terms of the Lambert W function, and show that the optimal rate is of logarithmic order, confirming the existence of the curse of infinite dimensionality.