A dynamic system is driven by random forcing, so there are only output measurements. AR model works well with a small number of lags, to predict the future values of the response. Also, n4sid was used to estimate a prediction model. There are multiple output values. In this case, AR model predicts the future values better than n4sid. To increase the accuracy of n4sid, I was wondering if lagged output variables could be used as inputs (to n4sid). However, when I try this, using lagged variables, n4sid responds with an error. Is there approach for using lagged values in n4sid?