Quantum project 3, credit risk
Nine answers, nine boxes of points, one sum. Built on 1,000 real German loans, then graded on 300 it never saw.
Narrated, with sound: the bins, the weights, the regression, the points and the tests that try to break it.
Take the first box, the overdrawn checking account. Four steps turn it into 36 points, and the same four steps make every other box.
Checking account has four bins. Among the 700 training loans, the share who defaulted climbs from 12% with no account to 51% when overdrawn.
Numbers like loan length are binned so risk moves one way only. A lender won't sign a card where risk rises, falls and rises again.
Overdrawn holds 19.8% of the repayers but 47.6% of the defaulters: ln(0.198 ÷ 0.476) = −0.88. Above zero means safer than average, below means riskier.
Logistic regression on the WOE values, α = −0.866. Every β must be negative, so safer evidence always pushes the score up; checking account gets β = −0.833 (p = 3×10⁻¹¹).
Telephone entered at p = 0.02 but survived only 36% of 25 rebuilds, so it is out. Nine characteristics stay.
factor = 20 ÷ ln 2 = 28.85 and offset = 487.12, so 600 points mean 50:1 odds of repaying and every 20 points double them. With n = 9:
−(0.731 − 0.096) × 28.85 + 54.12 = 35.8 → 36Our applicant: 36 for the overdrawn account, then 65, 53, 70, 66, 60, 72, 63 and 66. Total 551, above the cutoff of 539.
The worst possible answers still add up to 348 and the best to 655, so every score lives between those two.
The holdout: 210 people who repaid (above the line) and 90 who defaulted (below). No step of the build touched them.
The two crowds overlap a lot. A scorecard ranks people, it can't sort them perfectly.
The cheapest line on the training book is 539. On the holdout it approves 99 people (33%) and 8 of them default: 0.530 per applicant, against 1.50 approving everyone and 0.70 approving no one.
Tuning the line on the holdout itself would show 0.503, a 5% flattery. The line is chosen first and paid for after.
Pick one repayer and one defaulter at random: 78.0% of the time the repayer scores higher. That is the AUC.
Gini = 2 × AUC − 1 = 2 × 0.780 − 1 = 0.560On its own 700 training rows it read 0.659: the gap is overfit. Rebuilt 25 times, 90% of results fall between 0.41 and 0.67, so 0.560 means "somewhere in there".
Another holdout applicant scored 552 and defaulted. At 551 the card gives odds of about 9 to 1, so roughly one in ten people with this score still won't pay.
A score is odds, not a verdict.
AUC on the same 300 unseen applications.
Bars start at AUC 0.5, a coin flip. On 700 rows binning regularises hard; on a million-row book this order would likely flip.
Sex and nationality never enter the model. Other answers can still stand in for them, so the finished card is checked.
Approval rate of women over men (30.4% vs 34.1%), above the 0.8 four-fifths line.
All nine characteristics and every bin, with the points the model gives. Tap one answer per row; it starts on our applicant.
The points table loads from scorecard_results.json.