Cost-Aware Deep Neural Network for Credit Card Fraud Detection under Chronological Evaluation

Credit card fraud detection remains challenging due to extreme class imbalance, evolving fraud patterns, and asymmetric misclassification costs. This paper presents a deployment-oriented evaluation and decision-calibration framework for transaction-level fraud detection on the public creditcard.csv dataset, assessed under a strictly chronological train–validation–test protocol that mirrors real-world operation. Our contribution lies in evaluation design and cost-aware decision calibration rather than architectural novelty, focusing on how probabilistic model outputs are translated into operational decisions under explicit cost constraints. A standard feed-forward deep neural network (DNN) is trained on the numerical features using a class-weighted binary cross-entropy loss, with early stopping guided by validation AUC–PR. At deployment time, the decision threshold is selected on the validation window by minimizing an empirical cost function that penalizes false negatives more than false positives. On the held-out test set, the proposed pipeline achieves a ROC-AUC of 0.9489 and a PR-AUC of 0.7813. We show that decision policy choice strongly affects operational outcomes: naive thresholding yields excessive false alarms, whereas validation-based cost-sensitive calibration substantially reduces expected loss. Under C_FN = 10 and C_FP = 1, the cost-optimal threshold yields an expected test cost of 190. Comparisons with logistic regression, random forest, and XGBoost under identical preprocessing, temporal splitting, and decision calibration show that tree-based ensembles remain highly competitive, while the evaluated DNN achieves comparable cost and precision–recall performance. Overall, the results highlight the importance of combining chronological evaluation with explicit cost-sensitive thresholding for practical fraud detection under severe class imbalance.