An Intelligent Predictive Analytics Framework for Financial Fraud Detection, Risk Assessment, and Automated Accounting Decision-Making

For financial fraud detection, these types of analytical approaches must be able to detect very rare fraudulent transactions and also inform transparent and actionable financial decision-making. This study created an intelligent predictive analytics framework that combines machine learning fraud detection, transaction-level risk assessment, explainable modelling and automated accounting decision support. Transaction characteristics and engineered balance-consistency indicators were used to train and compare Logistic Regression, Decision Tree, Random Forest, and XGBoost models. The performance of the models was measured by the accuracy, precision, recall, specificity, F1-score, ROC-AUC, PR-AUC and confusion-matrix metrics and special emphasis was placed on imbalance-sensitive metrics. Tree-based and ensemble models demonstrated best predictive accuracy, whereas the features related to balance have had a strong impact on discrimination of fraud. Predicted fraud probabilities were then converted to continuous risk scores, then assigned to operational risk levels. These risk categories were associated with specific actions in the accounts, such as normal processing, further verification, review of transactions, and investigation/referral for further review. The framework therefore goes beyond binary classification of fraud and combines the elements of fraud prediction, interpretability, risk assessment and accounting decision support. The framework needs to be tested in the future using independent real-world transaction streams and institution-specific decision thresholds.