Credit Card Default Analysis

Default-risk analysis of 30,000 Taiwanese credit card clients (April–September 2005), combining engineered risk metrics, three hypothesis tests, two classifiers, and a Power BI dashboard.

Of the 30,000 clients, 22.12% (6.64K) defaulted the following month.

Deliverable
src/features.py	Derived risk metrics from raw repayment and billing history
src/hypothesis_tests.py	Three hypotheses about who defaults
src/modeling.py	Logistic Regression baseline, then Random Forest
Credit_card.pbix	Three-page Power BI dashboard

Dashboard

Page 1 — Delay behaviour by demographics. Average payment delay and maximum delay streak, both by age and by sex. Clients aged 21–25 show the best payment-delay behaviour; those aged 23–57 average under one month of delay streak. Female clients show more delays on average, while male clients carry a longer delay streak (higher by 0.13).

Page 2 — Default drivers. Default rate by most recent repayment status (PAY_0) against an industry-acceptable benchmark, average utilisation against a benchmark, credit limit by age, and default counts broken down by education, marital status and sex. Includes the overall 77.88% / 22.12% split.

Page 3 — Risk segmentation. Utilisation against average payment delay coloured by outcome, plus default counts by delay streak and by credit limit, sliceable by age and limit.

Risk metrics

Built in src/features.py from the raw monthly columns:

Metric	Meaning
AVG_PAY_DELAY	Mean repayment status across the six months
DELAY_TREND	Slope of repayment status over time — positive = worsening
MAX_DELAY_STREAK	Longest run of consecutive months two-plus payments late
AVG_UTIL	Mean of monthly bill ÷ credit limit
AVG_PAY_RATIO	Mean of payment ÷ bill, over months with a positive bill
BILL_VOLATILITY	Standard deviation of the six monthly bills
LAST_MONTH_SHOCK	Most recent bill minus the average of the preceding five

On DELAY_TREND and column order

The source orders repayment columns newest to oldest: PAY_0 is September 2005, PAY_6 is April 2005. Fitting a slope in that column order produces a value whose sign is inverted — a positive slope would mean delays improving.

DELAY_TREND is therefore fitted oldest → newest, so a positive value means delays are getting worse, which is what the hypothesis assumes. This matters: fitted in source order, the correlation with default comes out negative and looks like a counterintuitive finding, when it is only the columns being backwards.

On AVG_PAY_RATIO

Payment ÷ bill is undefined when the bill is zero and meaningless when it is negative (a credit balance). Those months are excluded from the average rather than contributing inf, which would otherwise propagate through every downstream statistic.

Hypothesis tests

pip install -r requirements.txt
python -m src.hypothesis_tests

#	Hypothesis	Test
H1	High utilisation raises default risk	Chi-square on proportions + φ
H2	A worsening repayment trend predicts default	Point-biserial correlation
H3	Bill volatility differs by outcome	Mann-Whitney U + rank-biserial

H1 compares proportions, not means. The outcome is binary, so the question is whether the default rate differs between utilisation groups. A t-test assumes roughly normal data; a 0/1 Bernoulli variable is not, and the t-statistic it returns does not mean what it appears to.

H3 uses a rank test. Bill volatility is a standard deviation — strictly positive and heavily right-skewed — so medians and Mann-Whitney are reported rather than means and a t-test.

Effect sizes accompany every p-value. At n = 30,000 nearly any difference reaches significance, so φ, rank-biserial correlation and relative risk are reported alongside; where an effect is negligible, the output says so.

Modelling

python -m src.modeling

Logistic Regression as the baseline, then a Random Forest that has to beat it. 70/30 stratified split, random_state=42.

Features: AVG_UTIL, AVG_PAY_DELAY, DELAY_TREND, BILL_VOLATILITY, MAX_DELAY_STREAK, AGE, LIMIT_BAL.

Both models use class_weight="balanced". With 78% of clients not defaulting, an unweighted model maximises accuracy by predicting "no default" almost always — respectable accuracy, near-useless recall on the class that actually matters.

The decision threshold is tuned, not assumed at 0.5. With the positive class at 22%, the default cut-off trades away most of the recall that makes the model useful. The scripts report F1 at both 0.5 and the tuned threshold.

ROC-AUC is reported with PR-AUC and per-class precision/recall. On an imbalanced problem ROC-AUC alone flatters a model; precision-recall on the default class is the honest view.

Outputs figures/model_curves.png (ROC and PR for both models) and figures/feature_importance.png.

Data

The raw file is not committed. Download the Default of Credit Card Clients dataset from UCI and save it as Credit_card.xlsx in the repository root. Both scripts fail with that instruction if it is missing.

Source columns

Column	Meaning
LIMIT_BAL	Credit limit in NT dollars
SEX	1 = male, 2 = female
EDUCATION	1 = graduate school, 2 = university, 3 = high school, 4 = other
MARRIAGE	1 = married, 2 = single, 3 = other
AGE	Years
PAY_0…PAY_6	Repayment status, September → April. −1 = paid in full, 0 = revolving credit, n = n months late
BILL_AMT1…6	Monthly bill, September → April
PAY_AMT1…6	Monthly payment, September → April
default.payment.next.month	1 = defaulted, 0 = repaid

Structure

src/features.py           risk metrics from raw history
src/hypothesis_tests.py   H1-H3 and their figures
src/modeling.py           both classifiers, curves, importance
figures/                  generated output
Credit_card.pbix          Power BI dashboard
requirements.txt          pinned versions

Limitations

• One market, six months, 2005. Taiwanese credit behaviour two decades ago; nothing here transfers to another market or period without revalidation.
• Prediction, not explanation. The models rank who is likely to default. They do not establish why, and feature importance is not a causal claim.
• Modest separation. These features leave the classes substantially overlapping. The output is useful for prioritising review, not for automated decisions about individual clients.
• Demographic fields are present in the data. SEX, EDUCATION and MARRIAGE are deliberately excluded from the model features. Using them to decide credit outcomes would be discriminatory regardless of predictive value; they appear in the dashboard for descriptive purposes only.
• EDUCATION and MARRIAGE contain undocumented codes (0, 5, 6) not covered by the source description.