Find the riskiest loan book segments
Drop in applications with their attributes and outcomes. The tool compares segments by one attribute or two together, then ranks those with a bad rate above the book’s by more than chance.
Applications file
Of 3,860 applications with a known outcome, 207 are bad and 3,653 are good, and 140 with no known outcome are left out.
The file is read in this browser and never uploaded. Only counts are sent if you ask for the result by email.
The columns used as attributes are store, region, product, channel, age, loan_amount, and numbers are cut into five bands of about equal size.
application_id were left out because almost every application has its own value, like an ID, or because only the first ten attributes are used.
A smaller segment can show a high bad rate from a handful of bads, so segments below this size are not tested.
More segments mean more can look bad by chance, so the test requires stronger evidence from each. Higher confidence flags fewer segments.
These are 4,000 generated demo applications from a point-of-sale lender, with a few risky segments planted in them.
Flagged segments
10 segments have a bad rate above the book’s that chance does not explain at 95% confidence, and the top 10 of them hold 69.1% of all bads in 45.4% of the applications. A segment is a place to look, not a cause.
Top segment bad rates
Red bars are flagged segments, and the dashed line is the book’s bad rate.
Segments by excess bads
| Rank | Segment | Apps | Bads | Bad rate | Lift | Excess bads | Flag |
|---|---|---|---|---|---|---|---|
| 1 | age = <24 | 745 | 71 | 9.53% | 1.8× | 31 | Yes |
| 2 | product = Phone · channel = online | 360 | 50 | 13.89% | 2.6× | 31 | Yes |
| 3 | channel = online | 1,164 | 89 | 7.65% | 1.4× | 27 | Yes |
| 4 | store = S07 | 178 | 30 | 16.85% | 3.1× | 20 | Yes |
| 5 | product = Phone | 1,171 | 80 | 6.83% | 1.3× | 17 | Within chance |
| 6 | channel = online · age = <24 | 240 | 30 | 12.50% | 2.3× | 17 | Yes |
| 7 | product = Phone · age = <24 | 239 | 29 | 12.13% | 2.3× | 16 | Yes |
| 8 | region = East · age = <24 | 180 | 25 | 13.89% | 2.6× | 15 | Yes |
| 9 | region = East | 1,005 | 69 | 6.87% | 1.3× | 15 | Within chance |
| 10 | age = <24 · loan_amount = <540 | 136 | 20 | 14.71% | 2.7× | 13 | Yes |
| 11 | store = S07 · product = Phone | 54 | 15 | 27.78% | 5.2× | 12 | Yes |
| 12 | region = East · product = Phone | 321 | 29 | 9.03% | 1.7× | 12 | Within chance |
| 13 | channel = online · loan_amount = [710,870) | 220 | 23 | 10.45% | 1.9× | 11 | Within chance |
| 14 | channel = online · loan_amount = <540 | 244 | 24 | 9.84% | 1.8× | 11 | Within chance |
| 15 | region = North · age = <24 | 208 | 21 | 10.10% | 1.9× | 10 | Within chance |
| 16 | region = East · channel = online | 303 | 26 | 8.58% | 1.6× | 10 | Within chance |
| 17 | age = <24 · loan_amount = [540,710) | 159 | 18 | 11.32% | 2.1× | 9 | Within chance |
| 18 | loan_amount = <540 | 738 | 49 | 6.64% | 1.2× | 9 | Within chance |
| 19 | loan_amount = [710,870) | 776 | 51 | 6.57% | 1.2× | 9 | Within chance |
| 20 | store = S07 · channel = online | 69 | 13 | 18.84% | 3.5× | 9 | Yes |
| 21 | store = S13 | 220 | 20 | 9.09% | 1.7× | 8 | Within chance |
| 22 | region = East · loan_amount = [710,870) | 187 | 18 | 9.63% | 1.8× | 8 | Within chance |
| 23 | product = Phone · loan_amount = <540 | 470 | 33 | 7.02% | 1.3× | 8 | Within chance |
| 24 | product = Appliance · age = <24 | 136 | 15 | 11.03% | 2.1× | 8 | Within chance |
| 25 | age = [34,46) · loan_amount = [710,870) | 158 | 16 | 10.13% | 1.9× | 8 | Within chance |
46 more segments are in the download.
Get the result by email
Use the result to test a stricter check or a referral for a segment in the credit decision against past applications. Start free
How segments are tested
Bad rate = Bads ÷ Apps Lift = Bad rate ÷ Book bad rate Excess bads = Bads − Apps × Book bad rate z = (Bad rate − Rest of book) ÷ √(p × (1 − p) × (1 ÷ n + 1 ÷ (N − n))) Flag = p ≤ (1 − Confidence) ÷ Segments tested
Each segment is tested against the rest of the book, and the confidence is shared across every segment tested, so testing many segments does not flag more by chance. A segment of two attributes is shown only when its bad rate is above both of its single attributes.
Find high-risk loan book segments
A bad rate by segment shows the share of applications with a bad outcome in a group, such as a store, channel or product. The tool compares each segment’s bad rate with the book’s rate and tests whether chance could explain the difference.
Bads can gather in particular stores, channels or customer groups, while the overall book rate hides where they concentrate. Use flagged segments as places to look, then test any stricter check or referral against past applications.
Common questions
How do I find high-risk segments in a loan book?
Compare the bad rate for each segment with the book’s rate. This tool tests whether the difference is more than chance and ranks segments by excess bads.
How do I compare bad rates by store or channel?
Use the store or channel as an attribute and compare the bad rate for each group with the book’s rate. The tool also shows applications, bads, lift and excess bads for each segment.
What does lift mean in risk segmentation?
Lift is a segment’s bad rate divided by the book’s bad rate. It shows how the segment’s rate compares with the overall book.
How can I tell a risky segment from chance?
The tool tests each segment against the rest of the book and adjusts for the number of segments tested. A flagged segment has a bad rate above the book’s that chance does not explain at the chosen confidence.
Why do bads concentrate in a few stores or channels?
The overall book rate can hide differences between groups. Comparing bad rates by attribute can show where bads are concentrated, but a segment is a place to look, not a cause.