UPI Transaction Fraud Detection Using Machine Learning: A Behavioural Risk-Scoring Prototype

Digital payment fraud detection requires identifying suspicious transactions while limiting unnecessary interventions for legitimate users. This paper presents a behavioural machine-learning prototype for UPI-style transaction risk assessment. Experiments use a project-generated synthetic dataset containing 353,118 transaction records from 2,500 simulated users between January and June 2026; the data do not represent actual UPI, NPCI, or bank transactions. A chronological partition assigns January–April to training, May to validation, and June to testing. A history-replay feature engine constructs 41 predictors from transaction attributes and information available before each transaction, including user spending patterns, payee familiarity, device history, transaction frequency, and location changes. Logistic Regression, Random Forest, and XGBoost are compared as supervised classifiers, while Isolation Forest is evaluated as an unsupervised baseline. XGBoost is selected using validation Average Precision (AP), with a validation AP of 0.9417. At the validation-selected BLOCK threshold of 0.4353, the model achieves 90.89% precision and 83.79% recall on the held-out June test set, detecting 579 of 691 fraud-labelled records with 58 false positives. A separate STEP-UP threshold of 0.2374 supports an intermediate risk decision. The prototype exposes scoring and analyst functions through a FastAPI backend and includes a Streamlit demonstration interface. Results are specific to the synthetic data generator and should be interpreted as proof-of-concept findings rather than evidence of performance on live UPI transactions.