Application fraud screening: strategy review

Where the scenario lands

Counts and rates are measured on the held-back period. Figures marked are arithmetic over the three assumptions you set below, and are not observed money.

Scenario

Nothing here approves or declines an application. Changing an input changes the arithmetic above. It does not change the evidence.

Screening approach Which method ranks the queue.

Review capacity Share of applications the team can work.

Concentration rules Force a case to review regardless of score. Aggregate signals, not identities.

Assumptions

Exposure avoided per fraud caught

Cost per case worked

Cost per good customer held up

Exhibit 1

Is fraud pressure rising or falling?

Applications assessed Fraud attempt rate

Observed across all eight months of the modelling population.

Exhibit 2

What does each level of review capacity buy?

Selected approach Screening in place today Your selected capacity

Observed on the held-back period, scored by each approach.

Exhibit 3

Where do the applications go?

Observed on the held-back period under the selected scenario. Actions are simulated and non-binding.

Exhibit 4

Which customer groups see the most fraud?

Observed fraud attempt rate by group. A group with fewer than 200 fraud attempts is withheld rather than shown, because the estimate would not be stable enough to act on.
Group	Applications	Fraud attempts	Attempt rate	Relative

Exhibit 5

How has the desk performed, period by period?

Monthly fraud KPI at the operating review capacity, aggregated in the fraud database from the scores and the review queue. Periods are the eight relative months of the source data, not calendar months.
Period	Applications	Fraud attempts	Attempt rate	Reviewed	Caught	Catch rate	Change	Reviewer hit rate

Top of the review queue

Highest-ranked cases under the selected approach. The outcome column is the retrospective label a reviewer would not have had at the time; it is shown so queue quality can be judged.
#	Case	Estimated risk	Simulated action	Rule fired	What drove the score	Outcome

Analyst detail Model performance, calibration, population stability, stress tests, matching, and provenance.

How the approaches compared

Held-back period. Ranking quality is PR-AUC; the calibration intercept gate is ±0.10.
Approach	PR-AUC	AUROC	Calib. intercept

Population stability

Features whose distribution moved enough to block automatic promotion.
Feature	Blocking months

Frozen-model stress tests

The selected model scored against bias-injected variants. Never used for training. Not a production-performance claim.
Variant	PR-AUC	AUROC

Cross-application matching

Measured on a separate synthetic fixture with held-out truth.
Corruption	Worst pair F1	False merges

Recorded governance wording

The refusal exactly as the evaluation program recorded it. The decision block at the top of this page is a plain-language rendering of these lines, not a replacement for them.

Provenance

Every source file checksum-verified before use.
File	Rows	Fraud rate	Checksum