A general framework is proposed for (auto) regression non-parametric estimation of recurrent time series in a class of Hilbert Markov processes with a Lipschitz conditional mean. This includes various non-stationarities by relaxing usual dependence assumptions as mixing or ergodicity, which are...