This paper presents a computational approach for the frequency-domain identification of multivariable, discrete-time transfer function models based on a cost function minimization. The algorithm is optimized for the parametric characterization of complex high-order multivariable systems requiring a large number of model parameters, including sparse matrix methods and QR-projections for the reduction of computation time and memory requirements. The algorithm supports a multivariable frequency-dependent weighting, which generally improves the quality of the transfer function model estimate. The overall approach is successfully demonstrated for a typical case encountered in experimental structural dynamics modelling (using modal analysis) and compared with related algorithms in order to assess the gain in computational efficiency. (c) 2005 Elsevier Ltd. All rights reserved.