The purpose of this study is to investigate the efficacy of three supervised learning models, Logistic Regression, Random Forest and XGBoost, on two datasets of financial fraud detection that were constructed differently with differing class distributions. The Credit Card Fraud Detection Dataset (Kaggle, 2023) is a synthetic dataset that has been artificially balanced to produce a 50:50 relative proportion of fraudulent and non-fraudulent observations to allow for performance of the models to be evaluated under ideal conditions. On the other hand, the Bank Account Fraud Dataset (NeurIPS, 2022) reflects real-world monetary behavior and features extreme class imbalance characterized by only approximately 1% of the observations containing fraudulent behavior. (Jesus et al., 2022) A single pipeline was constructed using stratified 60 / 20 / 20 splits and SMOTE applied only to the training set, Evaluation metrics included F1-score and AUC-ROC. The results reflect close to perfect outcomes on the balanced synthetic dataset but large degradation in performance on the real-world imbalanced dataset. The model that consistently performed best on the imbalanced dataset was XGBoost as represented by the F1 (23.4%) and AUC (89.3%) values. These results are consistent with published benchmarks indicating that F1-scores in the 15 to 25% range represent excellent outcomes in practice in detection of fraudulent behavior. The results of the present study underscore the critical impact of data imbalance and real-world practicality of the dataset used in the performance of supervised models and indicate future study to apply techniques such as cost-sensitive learning, explainability and temporal modeling of financial data in operational settings in order to achieve generalization with the models tested.