This paper is concerned with consistent nearest neighbor time series estimation for data generated by a Harris recurrent Markov chain. The goal is to validate nearest neighbor estimation in this general time series context, using simple and weak conditions. The framework considered covers, in a unified manner, a wide variety of statistical quantities, e.g. autoregression function, conditional quantiles, conditional tail estimators and, more generally, extremum estimators. The focus is theoretical, but examples are given to highlight applications.