Home / Current Issue / Paper 1722134
Score, Review, Retrain: A Closed-Loop Architecture for Real-Time Fraud Detection with Human-Verified Model Retraining in Digital Banking
Subject area: Science,Engineering and Technology · Area of research: Machine Learning
Abstract
The expansion of digital banking has widened the surface available to transaction fraud, and fraud detection mechanisms built on static, predefined rules have struggled to keep pace with fraud patterns that change faster than the rules describing them. Most published machine learning approaches to this problem evaluate a trained classifier against a fixed, offline dataset, without describing how that classifier would be embedded into a live transaction pipeline or how it would continue to improve once deployed. This paper presents the design, implementation, and evaluation of a fraud detection system in which risk scoring, human review, and model retraining are integrated into a single closed loop within a working digital banking backend, rather than treated as separate offline concerns. Every fund transfer and withdrawal is scored in real time by a Gradient Boosting classifier trained on more than forty behavioral, temporal, and velocity-derived features; transactions exceeding a configured risk threshold are withheld pending review by a human analyst, and confirmed outcomes are incorporated into subsequent, automatically scheduled retraining cycles. The system was evaluated on synthetically generated transaction data of two different sizes, with the final deployed configuration achieving an area under the receiver operating characteristic curve (AUC) of 0.94 and, at a class-balancing ratio tuned specifically to preserve precision, a precision of 0.24 with a recall of 0.57 at an F1-optimal decision threshold. Gradient Boosting is further compared against Logistic Regression and Random Forest baselines trained and evaluated under identical conditions. The results and their trade-offs are discussed in relation to dataset size, class-imbalance correction, and decision-threshold selection, and the paper concludes with a candid account of the system's limitations and directions for future work.
Keywords
Automated Retraining, Class Imbalance, Digital Banking, Fraud Detection, Gradient Boosting, Human-In-The-Loop, Machine Learning Operations, Real-Time Systems
How to cite this paper
@article{1722134,
author = {Aaisha Farooq},
title = {Score, Review, Retrain: A Closed-Loop Architecture for Real-Time Fraud Detection with Human-Verified Model Retraining in Digital Banking},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {10},
number = {2},
pages = {449-459},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1722134.pdf},
abstract = {The expansion of digital banking has widened the surface available to transaction fraud, and fraud detection mechanisms built on static, predefined rules have struggled to keep pace with fraud patterns that change faster than the rules describing them. Most published machine learning approaches to this problem evaluate a trained classifier against a fixed, offline dataset, without describing how that classifier would be embedded into a live transaction pipeline or how it would continue to improve once deployed. This paper presents the design, implementation, and evaluation of a fraud detection system in which risk scoring, human review, and model retraining are integrated into a single closed loop within a working digital banking backend, rather than treated as separate offline concerns. Every fund transfer and withdrawal is scored in real time by a Gradient Boosting classifier trained on more than forty behavioral, temporal, and velocity-derived features; transactions exceeding a configured risk threshold are withheld pending review by a human analyst, and confirmed outcomes are incorporated into subsequent, automatically scheduled retraining cycles. The system was evaluated on synthetically generated transaction data of two different sizes, with the final deployed configuration achieving an area under the receiver operating characteristic curve (AUC) of 0.94 and, at a class-balancing ratio tuned specifically to preserve precision, a precision of 0.24 with a recall of 0.57 at an F1-optimal decision threshold. Gradient Boosting is further compared against Logistic Regression and Random Forest baselines trained and evaluated under identical conditions. The results and their trade-offs are discussed in relation to dataset size, class-imbalance correction, and decision-threshold selection, and the paper concludes with a candid account of the system's limitations and directions for future work.},
keywords = {Automated Retraining, Class Imbalance, Digital Banking, Fraud Detection, Gradient Boosting, Human-In-The-Loop, Machine Learning Operations, Real-Time Systems},
month = {August},
}