Problem
When a customer sets up AutoPay at checkout, they often have several saved payment methods. Affirm picked the default with a few fixed rules. Payments can fail even when funds are available, because of expired cards, outdated account details or temporary processing issues. The article measures the effect of the default on first-payment delinquency.
Example from the article: at AutoPay setup the customer has several saved methods and the system suggests the one more likely to support successful repayments, while any eligible method can still be chosen.
Solution in brief
PredictPay is a model that predicts payment failure risk for each saved method and ranks the eligible methods by it. The ranking only sets the default; the customer stays in control. Version 1 used XGBoost with batch scoring on existing ML platform infrastructure. Version 2 added richer payment-history features and raised ROC-AUC from 0.81 to 0.88.
Questions this case answers
- How to rank a customer's available options by predicted failure risk so that the default shown at a decision point is the one most likely to succeed
- How to move an ML prototype to a first production version on existing platform infrastructure and validate it with an online experiment
- How to choose the default payment method for recurring payments when a customer has several saved methods?
- How to predict whether a payment on a given saved method will fail?
Tools and methods
- XGBoost — v1 model that predicts payment failure risk per saved method
- InterpretML — explains individual predictions in the hackathon demo
- Batch scoring — scores are computed ahead of time and stored, so checkout reads them without waiting
- Offline evaluation and impact simulation — justify a production experiment before launch
- Checkout A/B experiment — measures delinquency, payment failure rate and take-up
Google or a one-time email code. No password.