System identification of continuous-time model based on discrete-time data can be performed using a algorithm combining linear regression and LQGbalanced model reduction. The approach is applicable also to unstable system dynamics and it provides balanced models for optimal linear prediction and control.