This paper reports on a study regarding modal property identification for railway vehicles using a linear prediction model. The relationship between input (excitation force or axlebox acceleration) and output (carbody acceleration) of an actual railway vehicle obtained by stationary or running tests is expressed by means of an ARX (Auto-Regressive eXogenious) model, and the procedure for the extraction of modal properties is described in detail. Determination of an appropriate model order (i.e., the order of the prediction coefficients in the ARX model) is specifically discussed from the viewpoint of practical use. The implementation of average estimation errors for two different parts of the analyzed data is proposed, and their effectiveness in determining the model order is evaluated. Suitability for the MIMO (multiple-input multipleoutput) problem using the ARX model is also described. It is shown that detailed modal characteristics can be successfully identified using the proposed method from measured data for both stationary and running tests.