Sample engagement · computed in your browser
Fraud & abuse — card transaction scoring
A transaction-fraud scorer measured on 20,000 scored card transactions, against a declared review capacity, a production cutoff and the PSD2 transaction-risk-analysis reference rate.
Mistakes we hid in this data on purpose
- — Fraud in this book sits in LOW-value transactions — the card-testing pattern — so a value-weighted PSD2 rate and a count-weighted one differ by roughly four times. An implementation that weighted by count would look correct until a supervisor asked.
- — The scoring model does not observe geographic mismatch at all, and under-weights every other driver relative to the process that generated the outcomes.
- — Labels mature over 120 days, so the most recent transactions in the book cannot yet be scored against an outcome and are excluded by name rather than counted as clean.
This is made-up data, and nothing on this page is precomputed — it runs through the real checking engine right now.
The evidence, and what was declared about it
- Book
- Card transaction book — synthetic, 20,000 scored transactions
- Source
- AlphaIQ synthetic transaction book (fraud-transaction-book-dgp.json)
- Fixture
- fraud-transaction-book.csv
- Model
- Card Fraud Scorer v7.1
This sample has no pass/fail target on file — on purpose. Nobody handed us the institution’s own cutoffs, so each number below is reported but not graded, and that is exactly why the recommendation below reads the way it does.