Credit Risk Scorecard — Consumer Lending

An end-to-end credit scorecard for consumer lending, built the way digital lenders (蚂蚁 / 微众 / 京东白条 style) actually do it: WOE binning → Information Value selection → logistic regression → a point-based scorecard, evaluated with KS and AUC.

Credit scoring is one of the oldest and most successful machine-learning applications in FinTech — the model that decides "how much credit, at what rate" for a retail borrower. This repo implements the full pipeline from raw credit data to a deployable scorecard.

What It Does

Step	Method	Output
1. Load & clean	pandas	1000 borrowers, 20 features, 30% bad rate
2. Binning	qcut (numeric) / category (categorical)	each feature → bins, missing handled
3. Weight of Evidence	WOE = ln(P_good / P_bad)	per-bin evidence
4. Information Value	IV = Σ (pct_good − pct_bad) · WOE	feature screening (keep IV ≥ 0.02)
5. Logistic regression	scikit-learn on WOE-transformed features	default-probability model
6. Scorecard scaling	base score + points per bin (PDO)	human-readable scorecard
7. Evaluation	KS, AUC	discrimination power

Results

On the UCI German Credit dataset (in-sample):

• KS = 0.523, AUC = 0.826
• 15 of 20 features retained (IV ≥ 0.02)
• checking_status is the dominant signal (IV = 0.666), followed by credit_history (0.293), duration (0.216), savings_status (0.196) and purpose (0.169).

Scorecard excerpt (base score 600, PDO 20 → 20 points double the good:bad odds):

Feature	Bin	Points
checking_status	>=200	+9.33
checking_status	no checking	+27.06
checking_status	0-200	−9.24
checking_status	<0	−18.82
duration	(12.0, 15.0]	+14.58
duration	(30.0, 72.0]	−16.80
credit_history	critical	+14.64

Higher score = lower default risk. A negative-balance account (<0) costs ~19 points, while a longer loan duration (>30 months) costs ~17 points — both intuitively raise default risk.

A note worth knowing (interview material)

no checking scoring +27 looks counter-intuitive — "no bank account" reads like a risk signal. But in this dataset that category has the lowest default rate (11.7%), while <0 balance is the highest (49.3%). The model correctly learns from the data, not from prior assumptions — exactly what you'd verify as a risk analyst before trusting a variable's direction.

Repository Structure

Credit-Risk-Scorecard/
├── README.md
├── requirements.txt
├── scorecard.py            # the full pipeline
├── data/
│   └── german_credit.csv   # cleaned, labelled UCI German Credit
└── output/                 # generated: scorecard.csv, iv_table.csv, ...

Getting Started

pip install -r requirements.txt
#   or (CN mirror):  pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

python scorecard.py --input data/german_credit.csv --output-dir output

# optional: hold out 30% of the sample to get an out-of-sample read
python scorecard.py --input data/german_credit.csv --output-dir output \
  --holdout-frac 0.3 --seed 42

bin_series relies on pd.api.types.is_numeric_dtype, so both string columns and the pandas 3.0 str dtype are binned as categories instead of being fed to qcut.

Outputs written to output/:

• scorecard.csv — the point-based scorecard (feature, bin, WOE, coef, points)
• iv_table.csv — every feature's Information Value
• woe_table.csv — per-bin counts and WOE
• metrics.json — KS, AUC, bad rate, scaling parameters (plus a holdout block with train / holdout metrics and score PSI when --holdout-frac > 0)

Why the holdout matters

With --holdout-frac 0.3 --seed 42 (700 train / 300 holdout, stratified so the holdout keeps the 30% bad rate):

Sample	KS	AUC
Train	0.563	0.843
Holdout	0.471	0.746

Score PSI between train and holdout is 0.046 — comfortably inside the "< 0.1 = stable" band, so the drop is generalisation error rather than a shifted population. It also puts a number on the caveat in Production Next Steps: the in-sample KS of 0.523 is optimistic by roughly nine points.

--cv 5 runs the same idea as stratified 5-fold cross-validation, refitting binning, WOE and the model inside every fold (reusing WOE values estimated on the full sample would leak the target into an out-of-sample estimate):

KS 0.467 ± 0.042 | AUC 0.772 ± 0.018

Two independent out-of-sample estimates agreeing that KS sits near 0.47 and AUC near 0.75–0.78 is the real result here; 0.523 is the number the training sample wants you to believe.

Data

UCI Statlog (German Credit): 1000 applicants, 20 attributes (7 numeric, 13 categorical), target 1 = default / 0 = good. The raw german.data was mapped to readable labels and re-coded; see the column meanings below.

Column	Meaning
checking_status	status of existing checking account
duration	loan duration in months
credit_history	credit history
purpose	loan purpose
credit_amount	credit amount
savings_status	savings / bonds
employment	present employment since
installment_rate	instalment rate (% of disposable income)
personal_status	personal status and sex
other_debtors	other debtors / guarantors
residence_since	present residence since (years)
property	property
age	age in years
other_installment	other installment plans
housing	housing
existing_credits	number of existing credits at this bank
job	job
num_dependents	number of dependents
own_telephone	telephone
foreign_worker	foreign worker

Production Next Steps (the FinTech angle)

This is an in-sample baseline. A production credit model adds:

• Train / validation / out-of-time (OOT) splits — the in-sample KS here (0.52) is optimistic; a real OOT figure is typically lower (≈0.35–0.45 on this data).
• Reject inference — model only sees approved applicants, so the training sample is biased; techniques (parcelling, augmentation) correct for that.
• Model explainability — SHAP / LIME to explain individual decisions to regulators and customers (a core FinTech compliance requirement).
• PSI / drift monitoring — track score stability after deployment.
• PD calibration — map scores to absolute probability of default for pricing and capital (Basel / IFRS 9).

Tech Stack

Python · pandas · numpy · scikit-learn