Stacked AI model catches banking fraud with 95.4% AUC
Logistic regression hits 0.946 AUC, but stacked generalization edges out at 0.954.
As banking moved online during the COVID-19 pandemic, fraudsters found new opportunities through digital platforms and fraudulent charity funds. In response, a research team from Lviv Polytechnic National University led by Bohdan Mytnyk developed a machine learning approach to recognize fraudulent banking operations automatically. Their paper, posted on arXiv (2608.07471), evaluates multiple classification algorithms on real transaction data, with a strong focus on preprocessing techniques that handle highly imbalanced datasets, feature transformation, and feature engineering to boost detection accuracy.
The team found that traditional logistic regression already performed well, achieving an AUC of approximately 0.946. However, a stacked generalization ensemble — combining multiple model predictions — delivered the best result, reaching an AUC of 0.954. The study also details an artificial neural network model that effectively improves fraud detection accuracy, and visualizes the performance of all algorithms for easy comparison. While the paper is an arXiv preprint and not yet peer-reviewed, it offers a practical, reproducible methodology that banks and fintech companies can adapt to build more accurate fraud detection systems, reducing false positives and catching more bad transactions.
- Logistic regression achieves 0.946 AUC on fraudulent banking transaction detection
- Stacked generalization ensemble outperforms with 0.954 AUC across tested algorithms
- Study details imbalanced-data handling, feature engineering, and an ANN-based model for practical deployment
Why It Matters
For fintech and banks, this research offers a proven ML blueprint to slash fraud losses and secure digital payments.