Modeling of time series using random forests: theoretical developments (Q2209824)

The authors apply random forests methodology to nonlinear time series modeling. The paper provides theoretical justification for applications in time series settings. A uniform concentration inequality across all regression trees built on nonlinear autoregressive processes is obtained under a mild condition on their minimum leaf size. Then this result is used to prove consistency for a large class of random forests. The approach is based on the theory of Markov processes and Bernstein type concentration inequalities. The assumptions are explicit in terms of a suitable smooth regression function \(f\) and the distribution of the noise term. For example, they are applicable if \(f\) is bounded and Lipschitz continuous, and the noise sequence is i.i.d. with a light-tailed distribution. Simulation studies for various specifications of \(f\) are presented.