SUMMARY Predictive relationships for the ground motion in the Marmara region (northwestern Turkey) are parametrized after regressing three-component waveforms from regional earthquakes, in the frequency range: 0.4‐15.0 Hz, and in the distance range: 10‐200 km. The data set consists of 2400 three-component recordings from 462 earthquakes, recorded at 53 stations. Moment magnitudes, M w, range between 2.5 and 7.2. The largest event for which we have waveforms available (M w 7.2) occurred in Duzce on 1999 November 12. The aftershocks of that earthquake, together with the aftershocks of the 1999 August 17 Izmit event (M w = 7.4), are included in the dataset. Regressions are performed, independently, on Fourier velocity spectra and on peak ground velocities, for a large number of sampling frequencies. A simple model is used to relate the logarithm of the measured ground motion to excitation, site, and propagation terms. Results obtained for peak velocities are used to define a piecewise continuous geometrical spreading function, g(r), a frequency-dependent Q(f ), and a distance-dependent duration function. The latter is used, through random vibration theory (RVT), in order to predict time-domain characteristics (i.e. peak values) of the ground motion. The complete model obtained for the peak ground motion was used to match the results of the regressions on the Fourier amplitudes. Fourier velocity spectra for the combined horizontal motion are best fit by a hinged quadrilinear geometrical spreading function for observations in the 10‐200 km hypocentral distance ranges as a function of frequency: f 100, f ≥ 1.0 Hz, r −1.0 for r ≤ 30 km; r −0.6 for 30 100 km. The frequency-dependent crustal shearwave quality factor Q (f ) coefficient Q( f ) = 180 f 0.45 . The T (5‐75 per cent) duration window provides good agreement between observed and predicted peak values. By modelling the behaviour of the small earthquakes at high frequency, we also quantified a regional parameter κ = 0.055 s. Spectral models with one single-corner frequency (Brune), and with two-corner frequencies (Atkinson and Silva) fit the observed high-frequency excitation levels equally well, whereas the model by Atkinson and Silva fits the low-frequency observations slightly better than Brune’s. RVT is used to predict the absolute levels of ground shaking, following Boore’s implementation of the stochastic ground motion model (Boore’s SMSIM codes). Our regional empirical predictive relationships are compared to the ones adopted in several regions of the world, from California to Western United States.