Design-Adaptive Pointwise Nonparametric Regression Estimation For Recurrent Markov Time Series

A general framework is proposed for (auto)regression nonparametric estimation of recurrent time series in a class of Hilbert Markov processes with a Lipschitz conditional mean. This includes various nonstationarities by relaxing usual dependence assumptions as mixing or ergodicity, which are replaced with recurrence. The cornerstone of design-adaptation is a data-driven bandwidth choice based on an empirical bias variance tradeoff, giving rise to a random consistency rate for a uniform kernel estimator. The estimator converges with this random rate, which is the optimal minimax random rate over the considered class of recurrent time series. Extensions to general kernel estimators are investigated. For weak dependent time-series, the order of the random rate coincides with the deterministic minimax rate previously derived. New deterministic estimation rates are obtained for modified Box-Cox transformations of Random Walks.