Team Coursework · Advanced Data Analytics / C
Consumer Credit Strategy & Customer Segmentation
Policy Analysis / Customer Segmentation / Business Analytics
A team coursework analysis of 50,000 historical lending records, connecting policy comparison and customer segmentation with practical credit-strategy decisions.
The more lenient Policy B approved a substantially larger share of borrowers who later defaulted in the historical dataset.
Welch's t-test: p < 0.05 · Cohen's d = 1.02
01 · Context & Decision Question
Consumer lending creates a direct trade-off: broader approval can expand lending opportunities, but it can also admit more borrowers who later default.
Using 50,000 historical lending records, the project asked how approval rules should balance acceptance and credit risk, and whether one policy should be applied to every borrower segment.
This was completed as team coursework for Advanced Data Analytics.
02 · Policy A/B Comparison
Policy A was the stricter baseline, using a debt-to-income threshold of 17 and accepting grades A, B and C. Policy B widened eligibility to a debt-to-income threshold of 25 and grades A through E.
The false negative rate measured the share of borrowers who defaulted but would still have been approved. Policy A produced an FNR of 0.27, compared with 0.73 under Policy B.
A Welch two-sample t-test found the difference statistically significant (p < 0.05), with a large effect size (Cohen's d = 1.02). This was a historical analytical comparison, not a live product experiment.
| Policy | Approval rule |
|---|---|
| Policy A | DTI ≤ 17 · Grades A, B, C |
| Policy B | DTI ≤ 25 · Grades A, B, C, D, E |
03 · Customer Segmentation
Segmentation used seven financial and credit-behaviour variables: annual income, debt-to-income ratio, revolving utilisation, recent inquiries, prior delinquencies, open accounts and total accounts.
Ward.D2 hierarchical clustering on a 3,000-record sample was used to initialise K-means across the full dataset. Elbow and silhouette analysis supported three interpretable customer groups.
The result separated borrowers into High-Leverage, Prior Default and Conservative segments, each with a different risk profile and approval trade-off.
Ward.D2 clustering on a 3,000-record sample initialised K-means across all 50,000 records.
Selected solution: k = 3 · Average silhouette = 0.165
| Segment | Core characteristic | Default rate |
|---|---|---|
| High-Leverage | Higher income with higher debt and borrowing intensity | 16.4% |
| Prior Default | Stronger history of delinquency | 15.9% |
| Conservative | Lower leverage with almost no prior delinquency | 15.1% |
04 · Business Implications
A single approval policy did not create the same trade-off for every segment. Stricter controls remained more defensible for higher-risk borrowers, while the Conservative segment combined the lowest observed default rate with lower leverage and almost no prior delinquency.
For Conservative borrowers, Policy A reduced the pass rate by 37.3% relative to Policy B, indicating that uniform strictness could reject comparatively strong applicants and narrow potential lending opportunities.
The strategic conclusion was to adjust approval thresholds by segment rather than treat strictness as a universal rule: tighter criteria for higher-risk profiles, and measured flexibility where observed risk is lower.
The dataset contained approved loans only, so rejected applicants were absent and the comparison is subject to selection bias. The records also date from 2012–2013, so thresholds should be recalibrated before any present-day operational use.