We present an algorithm for automated S-phase arrival time determination of local, regional and teleseismic events based on autoregressive (AR) prediction of the waveform. The waveforms of the horizontal components are predicted using a scalar AR model for multicomponent recordings. The AR coefficients are estimated in a short moving window using a least-squares approach minimizing the forward prediction error. Synthetic tests with single-component data show that the least-squares approach yields similar or even better results than the YuleWalker and Burgs algorithms. We discuss the choice of the AR model and show that the corresponding prediction error of the AR model, applied to both horizontal components, is sufficient to detect instantaneous changes in amplitude, frequency, phase and polarization. The rms prediction error of both horizontal components defines the characteristic function, to which an algorithm for the estimation of the arrival time is applied. The proposed algorithm also accounts for automatic quality assessment of the estimated S-onset times. Four quality criteria are used to define the weight of the automatically estimated S-arrival time. They are based on two different estimations of the slope of the characteristic function and on two signal-to-noise ratios (SNRs).