The fraud loss for selecting the model complexity in fraud detection

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Fraudulent activities necessitate efficient detection systems due to high case volumes and rarity of fraud.
• Resource limitations require investigators to focus on a select number (k) of high-probability fraudulent cases.
• Prediction models in fraud detection require regularization to prevent overfitting and ensure performance.

Purpose of the Study:

• To propose a novel 'fraud loss' function for selecting optimal model complexity in statistical fraud detection.
• To evaluate the effectiveness of the proposed fraud loss function against existing methods, particularly Area Under the Curve (AUC).
• To determine optimal validation settings through simulation studies.

Main Methods:

• Development of a regularized prediction model for identifying potentially fraudulent cases.
• Introduction of a 'fraud loss' function to tune model complexity.
• Comparative analysis using simulation studies and a real-world credit card default dataset.
• Performance evaluation based on fraud loss and Area Under the ROC Curve (AUC).

Main Results:

• The proposed fraud loss function effectively selects model complexity.
• In simulations and on a credit card dataset, the fraud loss approach yielded results comparable or superior to AUC-based complexity selection.
• The fraud loss metric proved a reliable indicator for optimizing fraud detection models.

Conclusions:

• The proposed fraud loss function offers an effective alternative for tuning model complexity in fraud detection.
• This method enhances the efficiency of fraud detection systems by better allocating investigative resources.
• The fraud loss approach demonstrates robust performance across simulated and real-world datasets, improving fraud identification accuracy.