Credit risk service, live demo

Receipts

Measured on the held-out split.

Every number on this page is read from a file the pipeline wrote; nothing is typed by hand. Each tile links to that file.

Test split, n = [todo]

Measured [todo] ROC-AUC Ranking quality, higher is better reports/metrics.json
Measured [todo] PR-AUC Average precision, higher is better reports/metrics.json
Measured [todo] Brier score Calibration, lower is better reports/metrics.json
Measured [todo] Trained UTC date of the run that produced the shipped model reports/metrics.json
Measured [todo] p95 latency POST /predict, host not recorded yet reports/loadtest.json

Try it

Score an applicant.

Three real applicants from the held-out split are preloaded. Change any field and submit; the API validates every value and returns a probability.

Presets load from the API: [todo]

Account

Credit limit is the line of credit in NT dollars, including family supplementary credit. Education uses the dataset codes: 1 graduate school, 2 university, 3 high school, 4 other.

Credit limit LIMIT_BAL

Education EDUCATION

Age AGE

Repayment status

One code per statement for the last six months. -2 no consumption, -1 paid in full, 0 revolving credit, 1 to 8 months of delayed payment.

Most recent month PAY_0

1 month back PAY_2

2 months back PAY_3

3 months back PAY_4

4 months back PAY_5

5 months back PAY_6

Bill amounts

Statement balance in NT dollars for each of the last six months. A negative value is a credit balance.

Most recent month BILL_AMT1

1 month back BILL_AMT2

2 months back BILL_AMT3

3 months back BILL_AMT4

4 months back BILL_AMT5

5 months back BILL_AMT6

Payment amounts

Amount paid in NT dollars toward the previous statement, for each of the last six months. Zero means no payment was made that month.

Most recent month PAY_AMT1

1 month back PAY_AMT2

2 months back PAY_AMT3

3 months back PAY_AMT4

4 months back PAY_AMT5

5 months back PAY_AMT6

POST /predict

Drivers

Top global drivers.

How much ROC-AUC drops when one feature is shuffled and the rest are left alone. Larger bars matter more to the model overall; this is not an explanation of a single prediction.

Loads from reports/importance.json through the API. [todo]

Permutation importance on the validation split. [todo]

Run it yourself

Same service, one command.

The image on GHCR contains the model, its version and metric receipts, the presets, and this page. The curl example posts the first preset above, so the payload is a real held-out applicant.

Docker

docker run -p 8000:8000 ghcr.io/zulqarnain-10/credit-risk-service:latest

Curl, first preset

curl -X POST http://127.0.0.1:8000/predict -H "Content-Type: application/json" -d '[todo]'

Then open /docs for the OpenAPI schema, /version for the receipts as JSON, and /metrics for Prometheus.