In this paper, an original approach to frequency identification is explained and demonstrated through an application in the domain of microwave filters. This approach splits into two stages: a stable and causal model of high degree is first computed from the data (completion stage); then, model reduction is performed to get a rational low order model. In the first stage the most is made of the data taking into account the expected behavior of the filter. A reduced order model is then computed by rational H-2 approximation. A new and efficient method has been developed, improved over the years and implemented to solve this problem. It heavily relies on the underlying Hilbert space structure and on a nice parameterization of the optimization set. This approach guarantees the stability of the MIMO approximant of prescribed McMillan degree. (C) 2012 Elsevier Ltd. All rights reserved.