Customer Credit Risk Dataset: Comprehensive Study Notes

Dataset Overview

  • 500 unique customer records (ID format CUST0001 – CUST0500)

  • Each row = one consumer’s six-month credit-behaviour time-slice

  • Tabular structure: 26 core input fields + 6 monthly payment-status labels

  • Primary analytical theme: consumer credit-risk profiling & default prediction

Core Numerical Features

• Age (yrs) – integer 18 – 75
• Annual Income ($) – ≈ 15 00015\,000 – 200 000200\,000
• Credit Score – FICO-like 300 – 850 scale
• Credit Utilization Ratio
– Already normalised (0 – 1); theoretical definition
Utilization=Current BalanceTotal Credit Limit\text{Utilization}=\dfrac{\text{Current Balance}}{\text{Total Credit Limit}}
• Missed Payments – count of historical misses (0–6)
• Delinquent Accounts – binary or small integer flag (0/1)
• Loan Balance ($) – outstanding instalment/term-loan amount
• Debt-to-Income (DTI) Ratio
– Pre-calculated (0 – 1)
DTI=Total Monthly Debt ServicingGross Monthly Income\text{DTI}=\dfrac{\text{Total Monthly Debt Servicing}}{\text{Gross Monthly Income}}
• Account Tenure (yrs) – customer’s relationship length with lender (0–19)

Core Categorical Features

• Employment Status – Employed, Self-employed, Unemployed, EMP (typo for employed), retired
• Credit-Card Tier – Student, Standard, Gold, Platinum, Business
• Location – 6 metropolitan areas: Chicago, Houston, Los Angeles, New York, Phoenix, Houston
• Monthly Payment Status (Month1 – Month6)
– Values: On-time, Late, Missed
– Captures recent behavioural trend → high predictive power

Typical Value Ranges & Observations

• Age clusters: early-20s (student cohort) & 55-70 (retired cohort)
• Income bimodal: <$60k (students / part-time) vs $120k+ (professionals) • Credit Scores span full spectrum; noticeable mass around 820–847 (prime) and 300–400 (high-risk) • Utilization extremes (≈0.050.05 & >0.900.90) indicate both highly conservative and maxed-out users
• Missed Payments: many rows at 4–6 → deliberately stress-testing modelling algorithms
• DTI often >0.40.4 (industry cutoff) → flagged for potential credit-counselling

Payment-Behaviour Encoding (6-month window)

  • Transition pattern examples
    • On-time → Late → Missed = deteriorating credit health
    • Missed → On-time → On-time = recovery scenario

  • Feature engineering idea: convert to weighted delinquency score, e.g.
    \text{Behaviour Score}=2\times(#\text{Missed})+1\times(#\text{Late})

Data-Quality Flags

• Inconsistent employment spellings: EMP, employed, Employed
• Missing numeric cells expressed as scientific notation placeholders (e.g. 3.6701E-01)
• Some rows show truncated / shifted columns (e.g. CUST0041 missing Income)
• Verify duplicate city names & case sensitivity
• Out-of-range values: negative or >1 utilisation in a few records

Potential Modelling Targets

  1. Binary Default (any "Missed" in last 6 months)

  2. Multi-class Payment Behaviour (On-time vs Late vs Missed)

  3. Continuous Risk Score combining credit metrics

Key Formulas & Derived Metrics

• Weighted Credit Index (illustrative)
Index=0.4 (850−Credit Score)+30 Util+50 DTI+5 Missed\text{Index}=0.4\,(850-\text{Credit Score})+30\,\text{Util}+50\,\text{DTI}+5\,\text{Missed}
• Projected Loss Given Default (LGD)
LGD=Loan Balance×Recovery Rate\text{LGD}=\text{Loan Balance}\times\text{Recovery Rate} (Recovery rate estimated from historical location-level data)

Real-World Relevance

• Banks decide limit increases / reductions based on utilization & behaviour
• Regulatory bodies (e.g. Basel III) monitor DTI and delinquency signals
• Marketing teams cross-sell Gold → Platinum upgrades to low-risk users
• Cities offer financial-literacy programs where chronic lateness is concentrated (Chicago & Phoenix in dataset)

Ethical & Philosophical Considerations

• Bias Risk – location & employment status may act as proxies for protected classes
• Privacy – granular payment history requires robust anonymisation (GDPR / CCPA)
• Fairness Metrics – evaluate Equal Opportunity\text{Equal Opportunity} & Disparate Impact\text{Disparate Impact} before deployment
• Transparency – explainable models (e.g. SHAP) preferred over black-box scoring

Connections to Previous Lectures

• Week 2: Data Cleaning → handle typos, scientific-notation strings, missing cell shifts
• Week 4: Feature Engineering → one-hot encode categorical tiers; compute rolling delinquency counts eighted Credit Index
✓ Recognise data-quality pitfalls (shifted columns, text-case issues)
✓ Explain fairness considerations when using location / employment in models
✓ Practice building a logistic-regression & tree model on this tableyments
✓ Be able to compute derived Behaviour Score & Weighted Credit Index
✓ Recognise data-quality pitfalls (shifted columns, text-case issues)
✓ Explain fairness considerations when using location / employment in models
✓ Practice building a logistic-regression & tree model on this table