A numerical procedure for the calibration of a source spectrum model to find the optimal combination of seismological parameters and derive unbiased ground motion prediction equations is presented, using available accelerogram recordings as calibration target. The source spectrum model is used to compute the radiated spectra of SH- waves for different seismic moments and hypocentral distances, combining point (far field) and finite (near field) source models. The expected value of the peak ground acceleration is computed from the radiated spectra using random vibrations theory. A genetic algorithm is used to search for the source spectrum parameters that best fit the available strong motion recordings. The procedure can be applied to any region with statistically significant accelerogram records. The proposed methodology is followed to derive spectral ground motion prediction equations compatible with the seismotectonic environment of Colombia, using 206 strong motion recordings as the calibration target. The obtained ground motion prediction equations are compared with those used in the latest national seismic hazard assessment study of Colombia. The comparison shows that the strong ground motion prediction equations obtained using the proposed methodology have lower residual bias, showing reasonable matching between them and the real strong motion recordings.