Hospital readmission rate has been broadly accepted as a quality measure and cost driver. However, success in reducing readmissions has been elusive. In the US, almost 20 % of Medicare inpatients are rehospitalized within 30 days, which amounts to a cost of $17 billion. Given the skyrocketing healthcare cost, policymakers, researchers and payers are focusing more than ever on readmission reduction. Both hospital comparison of readmissions as a quality measure and identification of high-risk patients for post-discharge interventions require accurate predictive modeling. However, most predictive models for readmissions perform poorly. In this study, we endeavored to explore the full potentials of predictive models for readmissions and to assess the predictive power of different independent variables. Our model reached the highest predicting ability (c-statistic =0.80) among all published studies that used administrative data. Our analyses reveal that demographics, socioeconomic variables, prior utilization and Diagnosis-related Group (DRG) all have limited predictive power; more sophisticated patient stratification algorithm or risk adjuster is desired for more accurate readmission predictions.