The Fees Nobody Puts on the Receipt: How ML Predicts the Cost of a Card Swipe
About this session
Every card payment triggers a web of interchange fees, network fees, and processing costs that customers never see, but that directly affect the economics of moving money. Predicting those costs accurately at massive transaction volume sounds like a modeling problem. In production, it becomes an engineering problem.
In this talk, I’ll share lessons from building and deploying machine learning for fee estimation at PayPal, where a prediction ultimately translates into dollars. I’ll walk through what happens after a promising model leaves the notebook: dealing with messy transaction signals, rare fee structures, and long-tail cases; validating predictions against financial ground truth; and designing inference and monitoring that remain reliable as payment rules and transaction patterns change.
The session will also examine why strong aggregate accuracy can mask financially meaningful errors, and how evaluation must change when different mistakes carry very different costs.
Attendees will leave with practical lessons for taking ML from model to money: evaluating what matters, engineering for edge cases, validating at scale, and building production systems where statistical accuracy is only the beginning. The real test is whether the prediction is financially right when it matters.
Speaker
Key takeaways
- Gain a practical framework for linking ML model performance to measurable business value
- Understand how to build ML systems that stay reliable as data and rules change
- Learn how to prioritize ML errors by business impact, not just model metrics