Name: Gurbinder Kaur & Aashish Arora
Country: India

Year: 2026
Volume: Volume No: 13, Issue No: 1, Year: 2026
Page Number: 463-473
DOI: https://doi.org/10.5281/zenodo.20769818

The rise of digital financial transactions has created challenges for fraud detection that have become even more pressing over time. Traditional methods of detecting fraud may be ineffective at scaling up with large amounts of transactions and uncovering sophisticated fraud patterns, which has prompted a search for new Machine Learning-based solutions to detect financial fraud. Therefore, comparative analyses of Machine Learning algorithms' performance in fraud classification have gained significant popularity recently.
This paper presents the comparative analysis of three popular algorithms for fraud detection: Logistic Regression, Random Forest, and XGBoost. For this purpose, the Credit Card Fraud Detection Dataset available on Kaggle was selected for analysis, which contains records both of legitimate and of fraudulent transactions. The analysis started by applying several techniques to process and engineer features within this dataset. Afterward, a set of metrics for analyzing predictive performance of selected Machine Learning algorithms was chosen, including Accuracy, Precision, Recall, F1-Score, ROC-AUC, Fraud Detection Efficiency, and False Positive Rate. Additionally, ANOVA tests were run to assess whether the performance of each algorithm differed significantly from the others.

It was observed that there were considerable variations in the performance of the models. Logistic Regression performed well in terms of classification and acted as the best baseline model for the current research work. As an ensemble learner, Random Forest was better at prediction than others, while XGBoost was found to have the best overall performance amongst all selected models. Thus, it can be concluded from statistical testing that the selection of different models had a significant effect on their performances.
The findings obtained from the study reveal that ensemble methods are more efficient than conventional classification algorithms in detecting fraud cases. Out of the selected algorithms, the XGBoost algorithm was found to have been the most successful model when it comes to classifying the data points. Random Forest and Logistic Regression followed XGBoost respectively.

Keywords: Machine Learning, Fraud Detection, Digital Banking, Logistic Regression, Random Forest, XGBoost, Financial Fraud, Classification Models.