Abstract Background Surgical treatment for hepatocellular carcinoma (HCC) is advancing, but a robust prediction model for survival after resection is not available. The aim of this study was to propose a prognostic grading system for resection of HCC. Methods This was a retrospective, multicentre study of patients who underwent first resection of HCC with curative intent between 2000 and 2007. Patients were divided randomly by a cross-validation method into training and validation sets. Prognostic factors were identified using a Cox proportional hazards model. The predictive model was built by decision-tree analysis to define the resection grades, and subsequently validated. Results A total of 16 931 patients from 795 hospitals were included. In the training set (8465 patients), four surgical grades were classified based on prognosis: grade A1 (1236 patients, 14.6 per cent; single tumour 3 cm or smaller and anatomical R0 resection); grade A2 (3614, 42.7 per cent; single tumour larger than 3 cm, or non-anatomical R0 resection); grade B (2277, 26.9 per cent; multiple tumours, or vascular invasion, and R0 resection); and grade C (1338, 15.8 per cent; multiple tumours with vascular invasion and R0 resection, or R1 resection). Five-year survival rates were 73.9 per cent (hazard ratio (HR) 1.00), 64.7 per cent (HR 1.51, 95 per cent c.i. 1.29 to 1.78), 50.6 per cent (HR 2.53, 2.15 to 2.98), and 34.8 per cent (HR 4.60, 3.90 to 5.42) for grades A1, A2, B, and C respectively. In the validation set (8466 patients), the grades had equivalent reproducibility for both overall and recurrence-free survival (all P < 0.001). Conclusion This grade is used to predict prognosis of patients undergoing resection of HCC.