Various bleeding risk prediction schemes, such as the Hypertension, Abnormal renal/liver function, Stroke, Bleeding history or predisposition, Labile international normalized Ratio, Elderly, Drugs/alcohol (HAS-BLED), Anticoagulation and Risk Factors in Atrial Fibrillation (ATRIA), and Outcomes Registry for Better Informed Treatment (ORBIT) scores, have been proposed in patients with atrial fibrillation. We compared the relative predictive values of these bleeding risk scores for clinically relevant bleeding and the relationship of ATRIA and ORBIT scores to the quality of anticoagulation control on warfarin, as reflected by time in therapeutic range. We conducted a post hoc ancillary analysis of clinically relevant bleeding and major bleeding events among 2293 patients receiving warfarin therapy in the AMADEUS trial. Only HAS-BLED was significantly predictive for clinically relevant bleeding, and all 3 risk scores were predictive for major bleeding. The predictive performance of HAS-BLED was modest, as reflected by c-indexes of 0.59 (P < .001) and 0.65 (P < .002) for clinically relevant bleeding and major bleeding, respectively. The HAS-BLED score performed better than the ATRIA (P = .002) or ORBIT (P = .001) score in predicting any clinically relevant bleeding. Only the HAS-BLED score was significantly associated with the risk for both bleeding outcomes on Cox regression analysis (any clinically relevant bleeding: hazard ratio, 1.85; 95% confidence interval, 1.43-2.40, P < .001; major bleeding: hazard ratio, 2.40; 95% confidence interval, 1.28-4.52; P = .007). There were strong inverse correlations of ATRIA and ORBIT scores to time in therapeutic range as a continuous variable (low risk ATRIA, r = −0.96; P = .003; ORBIT, r = −0.96; P = .003). Improvement in the predictive performance for both ATRIA and ORBIT scores for any clinically relevant bleeding was achieved by adding time in therapeutic range to both scores, with significant differences in c-indexes (P = .001 and P = .002, respectively), net reclassification improvement, and integrated discriminant improvement (both P < .001). All 3 bleeding risk prediction scores demonstrated modest predictive ability for bleeding outcomes, although the HAS-BLED score performed better than the ATRIA or ORBIT score. Significant improvements in both ATRIA and ORBIT score prediction performances were achieved by adding time in therapeutic range to both scores.