The increasing sophistication of financial fraud necessitates advanced detection systems to protect institutions and consumers from significant financial losses. This study aims to enhance fraud detection in financial transactions identifying the most effective individual machine learning (ML) algorithms for fraud detection, developing a hybrid ML model that combines multiple algorithms, evaluating model performance using financial transaction datasets and optimising the model to minimise false positives and false negatives. The research employs Fraudulent Transactions Data, synthesised through PaySim to simulate mobile money transactions, as the experimental dataset. A range of ML algorithms were assessed individually, including DTC, RFC, LGBM, AdaBoost, SVM, and KNN. A hybrid model was constructed using a StackingClassifier with a logistic regression meta-learner to utilize these classifiers' complimentary strengths. The methodology encompassed data cleaning, exploratory analysis, preprocessing, and feature selection using the Variance Inflation Factor (VIF). Experimental results demonstrated that the hybrid model outperformed all individual models, achieving prediction accuracy of 97.12%, followed closely by LGBM at 97.08%, while SVM performed the lowest at 87.78%. The hybrid model also achieved a fraud detection recall rate of 98.81%, correctly identifying most fraudulent transactions while reducing false positives to 113 out of 2484 legitimate cases. These results highlight the effectiveness of hybrid ML models in improving fraud detection accuracy while minimising operational disruptions. This research provides a scalable and robust framework for financial institutions to combat evolving fraud threats and offers a valuable foundation for future work in ML-driven financial security.

Financial Fraud Detection Hybrid Model Machine Learning Stacking Classifier Variance Inflation Factor (VIF) Decision Tree (DT) Random Forest Support Vector Machine Gradient Boosting Machine.