Machine Learning–Based Fraud Detection in Financial Records Using Anomaly Identification and Predictive Risk Scoring

Abstract

This study examined the problem of reactive, rule-bound, and opaque fraud detection in financial records, where fraudulent transactions and manipulated accounting entries are frequently identified only after loss has occurred, where fixed rule engines generate high false-positive rates, and where machine-learning models, when deployed, often function as black boxes that analysts cannot interpret or trust. The purpose of the study was to assess how a Machine Learning–Based Fraud Detection framework, integrating unsupervised anomaly identification with supervised predictive risk scoring, influences fraud detection performance in financial records. A quantitative, cross-sectional, case-based design was adopted, and data were collected through a structured five-point Likert-scale questionnaire from 142 valid respondents out of 160 distributed questionnaires, representing an 88.8% valid response rate. The sample included fraud analysts and investigators, data scientists and machine-learning engineers, auditors and forensic accountants, risk and compliance officers, and financial-systems managers, with 66.9% directly involved in fraud-detection or risk-scoring activities. The key variables were anomaly detection modeling, predictive risk scoring, feature engineering and data quality, model validation and performance, analyst and institutional engagement, framework design quality, and fraud detection performance. The analysis plan included descriptive statistics, reliability testing using Cronbach's alpha, Pearson correlation, regression modeling, a framework maturity index, and a fraud risk-control priority matrix. The headline findings showed that all major constructs were rated high, with fraud detection performance recording the highest mean score of 4.21, followed by predictive risk scoring at 4.15 and framework design quality at 4.11. Reliability was acceptable to excellent, with Cronbach's alpha values ranging from 0.81 to 0.93. Correlation results showed significant positive relationships, including r = 0.78 between framework design quality and fraud detection performance. Regression results confirmed that the model explained 71.4% of the variance in fraud detection performance, R² = 0.714, adjusted R² = 0.702, F(6,135) = 56.12, p < 0.001. Framework design quality was the strongest predictor, β = 0.34, followed by predictive risk scoring, β = 0.26, and anomaly detection modeling, β = 0.24. The findings imply that financial institutions should strengthen anomaly detection modeling, calibrated risk scoring, feature engineering, model validation, and analyst engagement to improve the accuracy, transparency, and trustworthiness of machine-learning fraud detection.