Credit card fraud has become a major issue in the online financial environment, requiring the implementation of smart and automated tools for real-time detection of frauds. Machine Learning (ML) has been an important asset in this area because of its capability to discover underlying patterns, learn new fraud methods, and offer scalable solutions. This study investigates the usage of different classical machine learning and deep transfer learning based on predictive models for credit card fraud detection with a focus on their comparative performance on six important parameters: time elapsed, accuracy, precision, recall, TNR and F1 score. The investigation makes use of a PCA transformed benchmark dataset with a total of 2,84,807 credit card transactions to train models. In depth experimentation is performed using five classical ML models named Random Forest, Logistic Regression, Linear SVM; Non-Linear SVM; XGBoost and four classical Deep Learned models named MLP, Shallow ANN, ID CNN, and LSTM. To enhance experimental validity, prediction capability of four GNN based CNN models such as Boosting-GNN, Jump-Attentive GNN, GNN and PC-GNN are also tested. Deep learning based neural network models are analysed using seven different activation functions and each model is fit using 10 epochs of batch size 512. Testing results point out that overall best performance in classical ML models is shown by Non-Linear SVM with best recall score depicted by ANN on RBG kernel and GPU. In the ensemble category, Random Forest model exhibits overall best performance with best recall for XGBoost. Precision, accuracy and F1 score of random forest and XG boost are highest. Results have shown that in case of Random Forest the accuracy, precision, recall and F1score are 99.9%, 97.7%, 81.9% and 89.12% respectively whereas for XG boost the values for accuracy, precision and F1 score are 99.96%, 92.63%, 83.81% and 88% respectively. Deep Learned models showed high accuracies, however they were significantly utilized computational resources in respect to elapsed time. The study provides a roadmap to financial institutions for efficient model selection while deciding on implementing automated and trustworthy fraud detection systems and helps shape the dynamic world of intelligent financial security solutions to reduce financial losses.