This paper proposes a frequency-dependent weighting function that, when used together with the classical starting value algorithms (linear least squares, total least squares,... ), improves starting values for frequency-domain identification of parametric rational transfer function models. The idea behind the method is to approximate the weighting of the maximum likelihood estimator without prior knowledge of the model parameters. The proposed weighting scheme is applied to simulation and measurement data. Enhanced parameter estimates are obtained even for wide band, high-order systems with large transfer function dynamics.