A system that investigates, identifies and cancels fraudulent transactions comprises a fraud detection server that receives a first dataset indicating a quantity of fraud-transactions. The first dataset is generated at least in part by a fraud-score model. The system receives a second dataset including a quantity of false positive fraud-transactions from the fraud transactions. The system calculates a fraud error rate using the quantity of fraud-transactions and the quantity of false positive fraud-transactions. The system generates a new fraud-score model when the fraud error rate exceeds a predefined error rate.