Experimental data used to carry out structural identification (either artificial or natural), usually consist of time-histories both of the excitation and of the response; however, in the civil engineering field it can be sometimes necessary to refer only to the measured response, due to difficulties to measure the forcing action, as it happens for wind-induced loads or for traffic (road or rail-road) induced loads. If the unmeasured loading process can be considered as stationary and with a wide-band and smooth frequency spectrum, as for wind or road traffic induced loads, reliable identification techniques are available both in time and frequency domain. The situation can be quite different for the case of train-induced loads, that are characterised by typical dominant frequencies; moreover, the loading process cannot be considered as stationary if the length of the train is much smaller than the bridge length. The paper discusses theoretical and computational aspects of the identification problem of bridges excited by moving trains; the results of an experimental investigation on a masonry arch railway bridge are presented. Different techniques are used to identify the modal parameters of the structure and the experimentally evaluated quantities are compared to the finite element model results. In this ambit a peculiar aspect of the excitation, that can be suitably used in the identification process, is discussed: the dependence of the positions of peaks in the travelling load frequency spectrum on the train speed.