<p>Credit card fraud continues to present substantial economic challenges, driven by the dynamic nature of fraudulent behavior and the inherent class imbalance in transactional data. In this study, a comprehensive comparison of fifteen machine learning algorithms, including classical classifiers (Logistic Regression, Support Vector Machine, and Ridge Classifier), discriminant analysis techniques (Linear and Quadratic Discriminant Analysis), distance-based models (K-Nearest Neighbours), probabilistic approaches (Naive Bayes), a single decision tree, ensemble learning methods (Random Forests, Extremely Randomised Trees, Extreme Gradient Boosting, Categorical Boosting, and Light Gradient Boosting Machine), and a deep learning model designed for tabular data (TabNet) is conducted. Experiments were conducted using three datasets: the 2013 European credit card transaction dataset, a synthetic dataset from 2019 and a real-world dataset collected in 2023. To address the severe class imbalance, the Synthetic Minority Oversampling Technique and Adaptive Synthetic Sampling is employed. The results indicate that ensemble methods, particularly Extreme Gradient Boosting and Categorical Boosting, achieve high accuracy and efficiency. Although TabNet delivers the best performance on the balanced 2023 dataset, its higher computational cost limits its practical deployment. Interpretability analysis using SHapley Additive exPlanations (SHAP) identifies transaction amount and features derived through Principal Component Analysis as key indicators of fraudulent activity. These insights support the development of interpretable, data-driven fraud detection models with real-world economic relevance. The study code is available at: <a href="https://github.com/Deep8s/CCFD.git">https://github.com/Deep8s/CCFD.git</a></p>