Abstract In this paper a 4SID algorithm is proposed to identify a class of linear stochastic systems from the noisy input-output data sequence. First, the standard linear stochastic models are replaced equivalently by the innovations representation of Kalman filter equation. Then, by introducing a quasi-stationarity assumption for the error covariance matrix associated with the Kalman filter, our 4SID algorithm is derived. Finally, by simulation studies the effectiveness of our algorithm is shown.