Fraud-Detection Algorithms for Identifying Anomalous Transactions in Retail Banking Networks

Abstract

This study addresses the persistent problem of detecting anomalous and fraudulent transactions in retail banking networks where high transaction volume, class imbalance, and evolving attack patterns create unacceptable false-positive burden and missed-fraud risk. The purpose was to quantify which enterprise fraud-detection system conditions most strongly predict perceived fraud-detection effectiveness within a real organizational setting, using a quantitative, cross-sectional, case-based design grounded in a cloud-hosted enterprise fraud-monitoring deployment. Data were collected from a purposive sample of N = 156 stakeholders drawn from enterprise operations and governance roles (fraud analysts 46.2%, IT or ML support 31.4%, compliance or risk 22.4%), representing cloud and enterprise operational cases within the case environment. Key variables included Perceived Fraud-Detection Effectiveness (PFDE) as the dependent outcome, and Algorithm Capability, Feature or Information Quality, Real-Time Readiness, Explainability Clarity, and Governance or Compliance Readiness as primary predictors; operational impact variables included a False-Positive Burden Index (FPBI) and an Alert-Triage Efficiency Index (TEI). The analysis plan applied descriptive statistics, reliability testing (Cronbach’s alpha), Pearson correlations, and multiple regression modeling. Findings indicated moderately high PFDE (M = 3.84, SD = 0.61) alongside strong measurement reliability (α range .82–.90; PFDE α = .88). PFDE correlated most strongly with TEI (r = .66, p < .001) and Feature or Information Quality (r = .62, p < .001), and negatively with FPBI (r = −.44, p < .001). The regression model explained substantial variance in PFDE (R² = .58; F(5,150) = 41.30, p < .001), with Feature or Information Quality emerging as the strongest predictor (β = .34, p < .001), followed by Governance or Compliance Readiness (β = .24, p = .002), Explainability (β = .18, p = .011), and Algorithm Capability (β = .15, p = .026), while Real-Time Readiness was not significant at p < .05 (β = .09, p = .087). Scenario stress testing showed weakest perceived capability for mule-account funneling (M = 3.21) and account takeover with device rotation (M = 3.33), implying that multi-step fraud patterns remain the hardest to manage. Practically, the results imply that banks should prioritize improving risk-information quality, governance controls, and explanation quality to raise detection effectiveness while reducing alert overload, rather than focusing only on speed improvements.