Application fraud after refusal and which cases need review
Your monthly review of approvals after refusal has brought you to a familiar kind of decision: whether an application after a refusal belongs in a rule or a review. At the top of the list is one store. An application was refused in the morning, then approved before lunch at the same counter, for a lower amount. The store's name sits beside the changed outcome, and your message to the sales director is nearly ready. The monthly file is still open on your screen. What needs to be clear before you send it?
Approvals after refusal had lower default rates
Every client in this comparison holds a current loan, and default is counted on that loan. Clients refused earlier and approved again before this loan defaulted at 9.6%. Clients refused earlier whose current loan was the first approval since those refusals defaulted at 11.7%. The approved clients also had lower default rates in every band of refusal counts.
Default here is the lender’s own flag of payment difficulty on the current loan. A fraud loss and a credit loss look the same in that flag.
| History | Clients | Default | Lift |
|---|---|---|---|
| No earlier refusal | about 190,000 | 7.1% [7.0, 7.2] | 0.86 |
| Refused before, this loan the first approval since | about 35,000 | 11.7% [11.3, 12.0] | 1.42 |
| Refused before, approved again before this loan | about 65,000 | 9.6% [9.4, 9.8] | 1.17 |
| Any refusal | about 100,000 | 10.3% [10.1, 10.5] | 1.26 |
The same pattern holds within each refusal-count band: among clients with one refusal, default was 7.6% for clients approved again before this loan and 10.5% for clients whose current loan was their first approval since the refusal. With two refusals, the rates were 9.2% and 12.1%. With three or four, they were 11.0% and 13.3%, and with five or more, 13.8% and 15.1%.
The refusal carries more information here than the approval that followed it. Treating every approval after refusal as the dangerous outcome would put more weight on the sequence than the current-loan results support.
Build the history in the right order before you compare clients: order each client’s earlier applications by decision day, then by the hour the application started, and pair each approval with the latest refusal since the previous approval.
Changed amount retries need review
Flag a same-day retry after refusal when the amount changes in either direction, and give the same-store retry particular attention.
Among same-day retries after refusal, defaults were similar whether the amount went down or up. They were higher at the same store, at about 13%, than elsewhere or on same-amount retries.
The lower amount can look like the obvious sign because it follows a refusal. The book gives the same warning for an increase, and the approved loans with lower amounts were late less often than those with the same amount. Use the changed amount as a retry flag, not as a reason to single out reductions.
The refusal-to-approval gap barely changed the current-loan default rate across the tested time windows.
Carry the same-store detail forward with the changed-amount flag.
Refusal history identifies higher risk cases
Use the refusal history to distinguish which approvals after refusal deserve closer attention.
Among clients whose latest approval followed a refusal, default was 12.6% after three or more refusals, compared with 8.7% after one. It was 11.3% when the current application came within 365 days after an approval that followed a refusal, compared with 8.6% when that approval was more than 365 days before the current application. Clients with any refusal recorded under the fraud-score code had a 19.5% default rate. When their latest approval followed such a refusal, the current-loan rate was 17.4%.
The paired loans after fraud-score refusals were late at 1.0%, against 2.5% for the book. The code’s meaning is not documented, so use its recorded value as a flag rather than treating its wording as a confirmed explanation. The current loan’s payment outcome is the relevant result for this decision.
Build separate flags for a fraud-score refusal, three or more refusals before approval, and an approval after a refusal in the last year.
Four client flags focus case review
Use the four client flags together after checking that they add information beyond refusal alone and hold across separate parts of the book.
| Flag | Clients | Defaults | Flagged | Others |
|---|---|---|---|---|
| a. Fraud-coded refusal | 2.3% | 5.6% | 19.5% | 7.9% |
| b. 3 or more refusals | 4.1% | 6.3% | 12.5% | 8.0% |
| c. Recent approval after refusal | 8.3% | 11% | 11.3% | 7.9% |
| d. Same-day retry, changed amount | 7.6% | 11% | 11.5% | 7.9% |
| Any of a to d | 15% | 21% | 11.5% | 7.6% |
| Any refusal at all | 34% | 43% | 10.3% | 7.1% |
The combined flags covered 15% of clients and 21% of defaults, with an 11.5% default rate for flagged clients and 7.6% for others. Among refused clients, the flagged group defaulted at 11.5%, against 9.4% for refused clients without those flags.
The result held on two random halves of the book. For the combined flags, flagged clients defaulted at 11.1% and 11.9% in the two halves, while the unflagged rates were 7.6% in each. Each individual flag also showed higher default rates among flagged clients in both halves.
A rule for any refusal would reach more clients, but it would also bring a larger share of the book into the flagged group. The four flags concentrate attention on histories that separate more clearly from other clients. Check both the refused-client comparison and the two-half result in a fraud rule backtest before using the flags in the decision flow.
Keep the combined flags only when both checks hold.
Store rankings guide application handling reviews
Use store patterns to direct review, and test store rankings against product mix and a later period before using them as a risk measure.
In the out-of-time test, the top 5% of stores by same-store pair share had raw late-payment rates 1.96 times those at other stores. After adjustment for product and decision period, the ratio was 1.04. The other store-pattern measures and random store lists told the same story.
The raw comparison follows product mix. In the later period, loans at the top 5% of stores by same-store pair share were 55% cash, 27% point-of-sale and 19% revolving. At other stores, the mix was 8% cash, 88% point-of-sale and 4% revolving. Book late-payment rates were 1.7% for cash, 2.6% for point-of-sale and 5.2% for revolving loans.
For the review list, rank eligible stores by the share of their approvals that followed a refusal of the same client at the same store. Use the working floor of 300 applications. The book share was 2.8%, so a store at twice that level, 5.6% or more, goes on the list. In the older-period test, the book share was 2.3%. The corresponding threshold was 4.5%. The store risk ranking tool is for this store-level comparison. The available store value does not identify an individual salesperson. A separate sales-level view belongs with the agent default concentration tool.
Use the list to inspect application handling and the cases behind the pattern. Product and period explain the raw store comparison. The adjusted results do not establish higher loan risk at those stores.
Store 3 stays on the review list
Store 3 stays on the review list because 7.9% of its approvals followed a refusal of the same client at the same store, against 2.8% for the book. Its approved loans went late 2.0% of the time, against 2.6% for eligible stores overall.
For Store 3, inspect the paired applications and their decision order, the amount change, the refusal reason recorded, and whether the same-store retry is part of a repeated pattern. Use that review to examine how applications are being handled at the store. The store’s position directs the review, while the four client flags determine the treatment of each application.
In the decision flow, use any of the four flags to refer an application for human review before approval, rather than making the flag an automatic decline. Put the flags into a decision table and keep the store list as a separate review queue.
Apply this procedure to the book:
- Order each client’s earlier applications by decision day, then by the hour each application was started.
- Pair each approval with the most recent refusal since the client’s previous approval.
- Flag any refusal coded fraud-score, an approval after three or more refusals, an approval after a refusal in the last year, and a same-day retry after refusal with an amount change in either direction.
- Check that the flags add information beyond refusal alone among refused clients, and confirm that the results hold on two random halves of the book.
- Include the same-store detail in the retry flag, since same-store changed-amount retries had the highest default rates.
- Rank stores with at least 300 applications by same-store pair share, and put stores at twice the book value on the review list.
- Rank stores on an older period, then test their later loans against rates adjusted for product and decision period and against random store lists of the same size before treating the ranking as a risk measure.
Questions about application fraud after refusal
How should I compare an application that was approved after an earlier refusal?
Pair the approval with the applicant’s latest refusal, then compare the submitted fields side by side. Preserve both versions of income, employer, time in job, address, identity details, requested amount and terms, along with the timing and order of each attempt. Review the differences before approval, and require evidence for changed personal facts that affect the decision.
Which changes between attempts should trigger an investigation for staff editing?
Prioritise changed personal facts that could help an application pass, such as income, employer, time in job, address or identity details. A same-day retry with a changed amount in either direction is also a review trigger, especially at the same store. Check the source evidence, who handled each attempt and whether several changes move in the same direction. A lower amount, down payment or longer term can be a normal response to refusal, so assess those alongside the other changes.
What should I record for every application after a refusal?
Keep a dated record of each attempt, the applicant, store or agent, staff member, decision and recorded refusal reason. Preserve each changed field’s earlier and submitted value, plus the evidence used to verify the change. Record overrides, who approved them and the eventual payment outcome, including first-payment default. That record lets a reviewer reconstruct what changed and how the application moved from refusal to approval.
How many retries should I allow before requiring an independent review?
Use one same-day resubmission as a starting limit, then require an independent review before another attempt. Allow a further attempt when the applicant supplies new, verifiable information that addresses the refusal. A fixed cooling-off period can delay a valid correction, so reserve it for repeated applications that add no new evidence or repeat the same edits. Review exceptions for both fraud risk and lost legitimate sales.
When should a rise in approvals after refusal at a store or agent trigger an investigation?
Investigate when the share rises above that location’s usual pattern, using comparable products and decision periods and checking that the volume is large enough to assess. For each store or agent, review the paired applications, changed personal facts, retry timing, recorded refusal reasons and who handled the attempts. Track first-payment default as an early outcome alongside later payment difficulty. Treat the location pattern as a reason to inspect handling, not as an automatic decline rule.
