In　the　seismic　reliability　analysis　of　engineering　structures,　it　is　often　necessary　to　adopt　stochastic　ground　motions　as　inputs.　However,　the　current　methods　used　to　generate　stochastic　samples,　such　as　the　spectral　representation　method,　the　spectral　transfer　method,　the　dynamic　source　method,　the　physical　source　method,　and　the　comprehensive　method,　etc.,　are　not　very　convincing　since　they　do　not　contain　the　complete　information　of　seismic　occurrence　and　propagation.　In　this　study,　an　improved　finite-fault　model　for　stochastic　simulation　of　ground　motions　is　proposed.　Compared　to　existing　stochastic　finite-fault　model,　the　improved　model　considers　the　propagation　law　of　phase　spectrum,　local　site　effect,　and　the　impact　of　wave　velocity　on　the　arrival　time　of　seismic　waves.　Besides,　identification　and　statistics　of　model　parameters　are　carried　out　based　on　ground　motion　records.　Reasonable　samples　of　stochastic　ground　motions　are　then　obtained.　In　addition,　parameter　sensitivity　analysis　is　conducted　to　determine　critical　random　variables,　which　reduces　the　complexity　of　stochastic　simulation.　Simulation　results　are　compared　with　recorded　accelerograms　from　the　1999　Chi-Chi　earthquake.　It　is　shown　that　the　improved　stochastic　finite-fault　model　is　reliable,　and　the　acceleration　response　spectra　of　simulated　samples　are　consistent　with　those　of　station　records　collected　at　various　site　classes　and　the　design　spectra.