In this note we suggest a new iterative least squares method for estimating scalar and vector ARMA models. A Monte Carlo study shows that the method has better small sample properties than existing least squares methods and compares favourably with maximum likelihood estimation as well.