Effective and robust supply chain management is necessary to meet the consumer demands in the 21st century. The blockchain architecture has been proposed for this purpose, possibly in combination with QR codes to access the data. Further, implementing detection of risks in the supply chain, such as late deliveries, can further strengthen the management system. However, with the complexity and large size of the data in the supply chain, a large amount of data may be required to properly train machine learning algorithms to forecast risk. Although such data can be available, accessing it is costly because of the nature of blockchain (distributed databases, verifications from transactions, etc.). Thus, in this research, we evaluate different ML algorithms to determine which is the best for predicting late delivery risk with a small amount of data. Results show that logistic regression is highly accurate and efficient to train, even with small data size. Random forest also performs well, but at a higher training cost. However, the use of logistic regression is limited to linearly separable problems. In conclusion, the dynamics and patterns of the data should be first understood to determine which algorithm is best for the task.