Abstract Objective To compare the predictive performance of CHA2DS2-VASc and HAS-BLED scores in Atrial Fibrillation (AF) patients with and without cancer. Methods and Analysis Using data from the Clinical Practice Research Datalink (CPRD) in England, we performed a retrospective cohort study of patients with new diagnosis of AF from 2009-2019. Cancer was defined as history of breast, prostate, colorectal, lung, or haematological cancer. We calculated the CHA2DS2-VASc and HAS-BLED scores for the 1-year risk of stroke and major bleeding events. Scores performance was estimated by discrimination [area under the receiver operating characteristic curve (AUC)] and calibration plots. Results Of 141,796 patients with AF, 10.3% had cancer. The CHA2DS2-VASc score had good to modest discrimination in prostate cancer AUC=0.74 (95%CI: 0.71, 0.77), haematological cancer AUC=0.71 (0.66, 0.76), colorectal cancer AUC=0.70 (0.66, 0.75), breast cancer AUC=0.70 (0.66, 0.74), and lung cancer AUC=0.69 (0.60, 0.79), compared to no-cancer AUC=0.73 (0.72, 0.74). HAS-BLED discrimination was poor in prostate cancer AUC=0.58 (0.55, 0.61), haematological cancer AUC=0.59 (0.55, 0.64), colorectal cancer AUC=0.57 (0.53, 0.61), breast cancer AUC=0.56 (0.52, 0.61), and lung cancer AUC=0.59 (0.51, 0.67), compared to no-cancer AUC=0.61 (0.60, 0.62). Both CHA2DS2-VASc score and HAS-BLED score were well calibrated across all study cohorts. Conclusion Amongst certain cancer cohorts in the AF population, CHA2DS2-VASc performs similarly in predicting stroke to AF patients without cancer. Our findings highlight the importance of cancer diagnosis during the development of risk scores and opportunities to optimize the HAS-BLED risk score to better serve cancer patients with AF.