Gastric cancer is one of the most common malignant tumours in the world. As one of the crucial hallmarks of cancer reprogramming of metabolism and the relevant researches have a promising application in the diagnosis treatment and prognostic prediction of malignant tumours. This study aims to identify a group of metabolism-related genes to construct a prediction model for the prognosis of gastric cancer. A large cohort of gastric cancer cases (1121 cases) from public database was included in our analysis and classified patients into training and testing cohorts at a ratio of 7: 3. After identifying a list of metabolism-related genes having prognostic value, we constructed a risk score based on metabolism-related genes using LASSO-COX method. According to the risk score, patients were divided into high- and low-risk groups. Our results revealed that high-risk patients had a significantly worse prognosis than low-risk patients in both the training (high-risk vs low-risk patients; five years overall survival: 37.2% vs 72.2%; p < 0.001) and testing cohorts (high-risk vs low-risk patients; five years overall survival: 42.9% vs 62.9%; p < 0.001). This observation was validated in the external validation cohort (high-risk vs. low-risk patients; five years overall survival: 30.2% vs 40.4%; p = 0.007). To reinforce the predictive ability of the model, we integrated risk score, age, adjuvant chemotherapy, and TNM stage into a nomogram. According to the result of receiver operating characteristic curves and decision curves analysis, we found that the nomogram score had a superior predictive ability than conventional factors, indicating that the risk score combined with clinicopathological features can develop a robust prediction for survival and improve the individualized clinical decision making of the patient. In conclusion, we identified a list of metabolic genes related to survival and developed a metabolism-based predictive model for gastric cancer. Through a series of bioinformatics and statistical analyses, the predictive ability of the model was confirmed. © 2020 The Authors