Empirical Fokker-Planck-based test of stationarity for time series.

Abstract

We propose a test of stationarity based on the drift coefficients of the Langevin type and the associated Fokker-Planck equations. The test relies on the estimation of the drift coefficients of the underlying probability densities and posits that a time series is nonstationary if the estimated drift term is a nonlinear function of the random variable of the observed time series and the Markov property holds. We provide ample empirical evidence that demonstrates that well- known stationary systems give rise to linear estimates of the drift coefficients, whereas nonstationary time series exhibit nonlinear estimates of the drift term. This does not, indeed, imply that a nonlinear drift term in the Fokker-Planck equation of a dynamic stochastic process causes nonstationarity.