Pattern informatics (PI) model is one of the recently developed predictive models of earthquake phys- ics based on the statistical mechanics of complex systems. In this paper, retrospective forecast test of the PI model was conducted for the earthquakes in Sichuan-Yunnan region since 1988, exploring the possibility to apply this model to the estimation of time-dependent seismic hazard in continental China. Regional earthquake catalogue down to ML3.0 from 1970 to 2007 was used. The ‘target magnitude’ for the forecast test was MS5.5. Fifteen-year long ‘sliding time window’ was used in the PI calculation, with ‘anomaly training time window’ being 5 years and ‘forecast time window’ being 5 years, respectively. Receiver operating characteristic (ROC) test was conducted for the evaluation of the forecast result, showing that the PI forecast outperforms not only random guess but also the simple number counting approach based on the clustering hypothesis of earthquakes (the RI forecast). If the ‘forecast time window’ was shortened to 3 years and 1 year, respectively, the forecast capability of the PI model de- creased significantly, albeit outperformed random forecast. For the one year ‘forecast time window’, the PI result was almost comparable to the RI result, indicating that clustering properties play a more important role at this time scale. Pattern information (PI) model is one of the recently developed predictive models of earthquake phys- ics based on the statistical mechanics of complex systems. In this paper, retrospective forecast test of the PI model was conducted for the earthquakes in Sichuan-Yunnan region since 1988, exploring the possibility to apply this model to the estimation of time-dependent seismic hazard in continental China. Regional earthquake catalog down to ML 3.0 from 1970 to 2007 was used. The ’target magnitude’ for the forecast test was MS5.5 . Fifteen-year long ’sliding time window’ was used in the PI calculation, with ’anomaly training time window’ being 5 years and ’forecast time window’ being 5 years, respectively. Receiver operating characteristic (ROC) test was conducted for the evaluation of the forecast result, showing that the PI forecast outperforms not only random guess but also the simple number counting approach based on the clustering hypothesis of earthquakes (the RI If the ’forecast time window’ was shortened to 3 years and 1 year, respectively, the forecast capability of the PI model de- creased significantly, albeit outperformed random forecast. For the one year ’forecast time window’, the PI result was almost comparable to the RI result, indicating that clustering properties play a more important role at this time scale.