Background: Oral anticoagulation (OAC) in patients with atrial fibrillation (AF) is a double‐edged sword, because it decreases the risk of stroke at the cost of an increased risk of bleeding. We compared the performance of a new bleeding prediction scheme, HAS‐BLED, with an older bleeding prediction scheme, HEMORR2HAGES, in a cohort of ‘real‐world’ AF patients. Methods: By individual‐level‐linkage of nationwide registers, we identified all patients (n = 118 584) discharged with non‐valvular AF in Denmark during the period 1997–2006, with and without OAC. Major bleeding rates during 1 year of follow‐up were determined, and the predictive capabilities of the two schemes were compared by c‐statistics. The risk of bleeding associated with individual risk factors composing HAS‐BLED was estimated using Cox proportional‐hazard analyses. Results: Of AF patients receiving OAC (n = 44 771), 34.8% and 47.3% were categorized as ‘low bleeding risk’ by HAS‐BLED and HEMORR2HAGES, respectively, and the bleeding rates per 100 person‐years were 2.66 (95% confidence interval CI, 2.40–2.94) and 3.06 (2.83–3.32), respectively. C‐statistics for the two schemes were 0.795 (0.759–0.829) and 0.771 (0.733–0.806), respectively. The risk factors composing HAS‐BLED were associated with varying risks, with a history of bleeding (hazard ratio HR 2.98; 95% CI 2.68–3.31) and being elderly (HR 1.93; 95% CI 1.71–2.18) being associated with the highest risks. Comparable results were found in AF patients not receiving OAC (n = 77 813). Conclusions: In an unselected nationwide cohort of hospitalized patients with atrial fibrillation, the HAS‐BLED score performs similarly to HEMORR2HAGES in predicting bleeding risk but HAS‐BLED is much simpler and easier to use in everyday clinical practise.