What Banks Know About You Before Approving a Credit Card (Data Science Explained)

Utilization, inquiries, late payments, income: every credit model runs on data. But where does it come from, what does it cost, and what is a bank actually allowed to use? In this episode we measure six data sources on 100,000 simulated card applications and keep score as we go. ✅ The application: free but weak (Gini 23%), plus the CARD Act ability-to-pay rule and the questions a lender can't ask ✅ Credit bureaus: why a credit report is every other lender's records pooled together (Gini 23% → 48%) ✅ The rules: FCRA permissible purpose, hard vs soft inquiries, prescreened offers and adverse action ✅ What a bureau pull is worth: about $41 saved per application, so it pays for itself many times over ✅ The bank's own records: a lift for existing customers (46% → 52%) ✅ Thin files and cash-flow data: 3.5 times as many thin-file applicants approved at the same risk ✅ Why raw data isn't a feature: a raw balance barely helps (22%) while engineered cash-flow features reach 36% ✅ Device and digital signals for fraud ✅ Social and digital data around the world: the digital-footprints study, app lenders in Kenya, the Philippines and India, India's 2022 limits, China's Sesame Credit and court blacklists, and the EU's ban on social scoring ✅ Where should the line be? From J. P. Morgan's "the first thing is character" to the risk of a dystopia ✅ Macro data moves levels, not rankings ✅ Which source feeds which model, and how banks test a new data source before buying it Every number on screen comes from real code on simulated data. Prices marked "assumed" are assumptions, since card issuers' bureau prices aren't public. ⏱️ Chapters 0:00 Intro 0:34 Six sources 1:02 The application 1:40 Credit bureaus 2:30 The rules 3:07 What a pull is worth 3:48 The bank's own records 4:24 Thin files & cash-flow data 5:33 Raw data isn't a feature 6:27 Device & digital signals 7:03 Social & digital data around the world 8:31 Where should the line be? 9:21 The economy 9:56 Which source, which model 10:28 Testing a new source 10:56 Recap & what's next 📚 Sources mentioned CFPB, Technical correction and update to the credit invisibles estimate (June 2025) Berg, Burg, Gombović and Puri, "On the Rise of FinTechs: Credit Scoring Using Digital Footprints", Review of Financial Studies (2020) Reserve Bank of India, Guidelines on Digital Lending (September 2022) EU Artificial Intelligence Act, Article 5 (prohibited practices, applicable from February 2025) J. P. Morgan, testimony to the Pujo Committee (December 1912) 📺 Earlier episodes: logistic regression, reading the model, evaluating a credit model, defining "bad", and time windows and leakage are all on the Gradient Room channel. 🔔 Next on Gradient Room: inside a credit report: tradelines, inquiries, payment history, and how raw records become model inputs. Like, share and subscribe so you don't miss it. 💬 Which data source does your team rely on most? And where would you draw the line on social and digital data? Tell me in the comments. Educational content on simulated data; not legal or financial advice. #datascience #creditrisk #machinelearning #alternativedata #fintech