The Phantom Borrower: Machine Learning Detection of Synthetic Identities and the Econometrics of Precision-Recall Trade-Offs in Digital Credit Markets

Abstract

Digital lending platforms face escalating synthetic identity fraud—fabricated identities combining real and fictitious data—causing an estimated $20 billion in U.S. losses in 2020. Traditional verification systems fail to detect these "phantom borrowers," creating severe information asymmetries in credit markets that undermine portfolio quality and financial stability.This study compares machine learning algorithms to detect synthetic identities and explore the econometrics impact of precision-recall trade-offs in fraud detection systems, and includes evidence-based recommendations for selecting optimal threshold. This is secondary research, which is a synthesis of empirical evidence from recent literature.  Asymmetric information theory and cost-sensitive learning approaches (2019-2024). Ensemble methods (XGBoost: AUPRC 0.847) and graph neural networks (AUPRC 0.891) outperform logistic regression (AUPRC 0.724). Feature importance analysis reveals identity consistency (28.4%), network/relational patterns (24.1%), and behavioral biometrics (22.7%) as primary detection signals. Optimal operating points vary by context: high-value loans require recall of 0.88-0.92, while small-dollar products prioritize precision at 0.82-0.90. Contextual factors significantly influence threshold selection, with F-beta scores ranging from 0.71 to 0.79 across scenarios. Machine learning is an effective method for detecting synthetic identities but thresholds should be optimized according to the costs of the product. To prevent fraud effectively, there needs to be a balance between the accuracy of fraud detection and the financial inclusion objectives, which means there is not a universal threshold for fraud. Regulatory framework should support a range of operating points to ensure financial stability and equitable access to credit.