Data Correction Layer · Tool 01 · Version 2.0
Synthetic Augmentation Decision Tool
A governed decision path from missing-label checks to measured gender parity, diagnosis, and fairness-controlled synthetic augmentation.
Reviewer guide · Tool 01
What it is
An interactive decision path a model development team follows when a credit model shows a gender approval-rate gap. It decides whether synthetic training data is the right fix, sends biased historical decisions for remediation instead, and checks whether any fix worked.
Why it matters
Synthetic data can fill gaps where women are under-represented. It cannot repair historical decisions that held women to a different standard; used there, it repeats the bias while the fairness numbers improve. It can also make a model look fairer while making it worse at predicting who repays.
How to read it
Enter sample values to see each route.
Prepare
Records with no decision, or with no recorded sex, are excluded rather than guessed, and the tool checks whether excluding them changes the share of women in the data. If the institution does not hold applicant sex at all, the model is marked not evaluable and referred to compliance.
Measure
Enter the demographic parity ratio (DPR): the female approval rate divided by the male approval rate on a hold-out sample. A result of 0.80 or above moves to routine monitoring under Tool 02. The tool also shows the result at a stricter 0.90.
Diagnose and route
Two checks: are there too few women in the segments the model scores (fewer than 100 approvals or denials in a segment), and did past decisions hold women to a different standard? Too few women goes to synthetic data. Biased past decisions go to remediation. If both are found, the decisions are corrected first.
Correct and re-measure
Synthetic records are added to training data only. The correction is accepted only if DPR is at least 0.80 and AUC has fallen by no more than 0.02. A team gets 3 attempts before returning to diagnosis; an unresolved gap goes to the model risk committee.
Questions this tool informs
| Item | What to look at |
|---|---|
| Q1–Q2 | Whether under-representation causes unfair outcomes at your institution, and how much structured attention it gets today |
| Q3 | Whether your institution could generate fairness-controlled synthetic data, in-house or through a vendor |
| Q4 | Whether a development team could follow the steps without material outside help |
| Q5 | The decision rules: DPR floor of 0.80, AUC tolerance of 0.02, and the minimum of 100 per segment |
| Q6 | Whether the decision record would meet your model risk framework’s data-quality documentation expectations |
| Q7 | Barriers to using the tool, and what you would change |
- 0Prepare
- 1Measure
- 2Diagnose
- 3Route
- 4Correct
- 5Re-measure
Step 0 — Prepare
Missing labels
Check exclusions before measuring parity. Records with sex not recorded or joint applications are tallied separately from the women/men counts; their sex is not inferred.
How to complete this step
- Enter the model/version and its Tool 04 registration reference. Apply the data rules fixed there before reading results.
- Confirm whether applicant sex is held. Count full-file records and exclusions separately for women and men; tally unknown-sex and joint applications separately. Never infer sex or guess declined applicants’ repayment outcomes.
- On HMDA data, codes 1–2 are approvals, 3 is denial, and 4–8 are excluded. Review the displayed female-share shift and record any selection-bias warning. If sex is unavailable, stop and refer the issue to compliance.
HMDA action-taken codes 1–2 = approval, 3 = denial, 4–8 excluded; sex is never inferred from other fields; no reject inference.
Decision record
How to complete this step
- Complete one record per run: review exclusions, bounding results, diagnosis, remediation, settings and attempt history. Enter owner/date and export your inputs explicitly to retain them in the Tool 04 audit trail.
- Model / registration
- — / —
- Records excluded
- —
- Attrition shift
- —
- Bounding result
- Not provided
- DPR / sensitivity
- —
- Representation
- —
- Label method / result
- — / —
- Route
- —
- Remediation / committee decision
- —
- Generation settings / attempts
- — / Not yet applied
- Post DPR / AUC fall
- — / —
- Decision
- Pending
Export your inputs
Key terms
- DPR
- Demographic parity ratio: the female approval rate divided by the male approval rate. The ratio and its interval describe the hold-out sample.
- AUC
- Area under the receiver operating characteristic curve, measuring predictive discrimination. AUC fall is baseline AUC minus post-augmentation AUC.
- Representational deficit
- Too few women in a segment of the training data, assessed separately for approvals and denials.
- Label bias
- Evidence that historical credit-decision labels were drawn differently for women.
- Synthetic augmentation
- Adding generated records to training data under fairness controls; validation and test data remain untouched.
- Bounding check
- A second DPR calculation treating withdrawn or incomplete files as denials to assess sensitivity to non-decisions.
- Attrition
- The change in the share of women after records without credit decisions are excluded.
- Threshold test
- A screen for whether lender decision thresholds differ for women. It can screen for label bias but cannot confirm it.
Thresholds are declared starting points fixed in the Tool 04 pre-registration. They are not regulatory requirements.