When I started working on credit card fraud detection, I assumed the goal was simple: catch as much fraud as possible. Then I realized a model could catch more fraudulent transactions by flagging more legitimate purchases too. If you’ve ever had your card declined for no apparent reason, you know why that matters.For this project, I compared four machine-learning models and two combinations of models called soft-voting ensembles. I tested them on a large synthetic transaction dataset, where fraudulent purchases made up less than 1% of the transactions. That imbalance made accuracy a pretty misleading measure: a model could label every purchase legitimate, look highly accurate, and catch no fraud at all.XGBoost performed best among the approaches I tested. But the most interesting part for me was deciding how to evaluate the models in the first place. Catching fraud, avoiding false alarms, and choosing when to flag a transaction all involve tradeoffs. A strong score on a dataset does not mean a model is ready to make decisions about real customers.You can read my full paper, published in the American Journal of Student Research, at the link below.

https://doi.org/10.70251/HYJR2348.45829841