Application fraud checks for deciding which applications to refer

You open a cash loan application from a laborer working for a self-employed employer, and the stated income sits well above what you would expect for someone in the same occupation and education group. The salary against position rule has fired. The file is now in your work queue, one more inconsistency asking to be turned into a decision about what happens next. You can send it to referral, where someone will spend time reviewing it, or decide the flag does not warrant that step. In a single application, how much should this mismatch carry?

Check which contradictions the form can record

Run hard contradictions once across the book, then retire checks the form prevents from firing. A job start before plausible working age and conflicting family details are candidate hard contradictions. On this lender’s form, however, some fields were computed from one another or limited at a boundary, preventing those mismatches from entering the book.

Here, default is the lender’s own flag of payment difficulty, set when an early instalment is late beyond its threshold. A fraud loss and a credit loss both land in that flag. The same lender’s book is also introduced in the article on applications after a refusal.

Keeping every plausible contradiction can make a checklist look thorough, but a check that cannot fire sends no useful applications to review. Several contradictions flagged zero applications across about 310,000 applications.

Count the expected hard contradictions on the whole book before putting any of them into a rule.

The form computes or limits some of the fields these checks compare. Retire the contradictions it prevents from appearing.

High stated income predicted lower default

On this book, a salary above the usual range for the applicant’s occupation and education was a poor reason to refer. Compare income within occupation-and-education cells, with a fallback for small groups, rather than against one salary for the whole book.

Applicants at three to five times their cell median defaulted at 5.9%, below the book’s 8.1%. The rate varied across thresholds, but remained below the rest of the book at each tested threshold. At the cell’s 99th-percentile threshold, the result also held in both random halves and in each outside-score tercile.

A team can reasonably read a sharp mismatch as a fabricated detail and send those applications to referral. On this book, the salary check instead selects applicants with lower default. The check points the wrong way.

Leave high stated income out of the referral count on this book.

Test income against comparable occupation-and-education cells on your own book.

Test recent changes address gaps and loan ratios

The checks that held on this book were recent changes, address gaps and a high credit amount relative to goods price. They merit testing because they captured higher default, while the familiar salary mismatch selected applicants with lower default.

• Recent changes. ID and registration changes within 30 days had higher default than the rest of the book. Registration changes within seven days gave the sharpest result in the recency sweep.
• Address gaps. Registration in a different city from the applicant’s living or working city separated groups with higher default.
• Credit over goods price. A high credit-to-goods ratio had higher default. The six-check count used the 99th-percentile cut, fixed before the ratio sweep.

Each threshold needs a comparison that fits the check: occupation and education for income, the same product for a loan ratio, and a sweep of day windows for recent changes. The riskiest segments tool can help set thresholds within comparable cells.

Choosing a cut because it gives the most striking default rate can make a check look stronger than it is. The cut changes which applications go to review and how many defaults those applications include. Set the comparison and cut before judging the outcome.

Test each candidate against its own comparable applications before counting it.

The seven-day registration result is distinct from the 30-day registration check used in the six-check count.

Check whether results hold across halves and age groups

Validate a check’s direction and repeatability before relying on its lift. A higher rate in one slice can reflect age, loan term or a threshold chosen to fit that slice.

On this book, the checks shown in the half comparison had higher default in both random halves. For each check on your own book, set the threshold on one half and apply it to the other, then set it on the second half and apply it back to the first. That evaluates each threshold on the opposite half, using applications held out from threshold setting. The reported cross-half credit-over-goods result here is for a cut set on one half and applied to the other.

Compare each flagged group with the rate expected from its five-year age bands. A ratio near 1.0 means age accounts for the apparent lift: car age relative to adult life and registration in another region both came to 1.01 here. The other checks ranged from 1.18 to 1.59. For a loan-ratio check, also compare groups formed by credit amount over payment, a stand-in for loan term. The credit-over-goods result remained elevated in several such groups.

A full-book rate can look persuasive even when a result varies across halves or age accounts for its lift, sending review capacity toward a weak check. Keep checks whose direction holds across both halves and whose lift survives the proxy comparisons.

Make the threshold earn its place on the other half of the book.

Use the fraud rule backtest tool to test the exact check and threshold on your own book.

Put these checks into a decision flow

Create rules for credit, fraud or collections, then run them on live decisions.

Start free See pricing

Six checks separate default within score bands

The tested six-check count separated default inside every outside-score tercile, including the best-scored third. The six checks are:

• Credit over goods price. Cash-loan credit amount divided by goods price, with the cut at the 99th percentile of that ratio among cash loans.
• Registered elsewhere than living. The registration city differs from the living city.
• Registered elsewhere than working. The registration city differs from the work city.
• Recent registration change. The registration changed within 30 days.
• Recent ID change. The ID document changed within 30 days.
• Phone changed on application day. The phone was recorded as changed on the day of application.

First check whether default rises as more of these checks fail.

Default rose from 7.0% with no failed checks to 16.5% with three or more. Three or more applied to 1.0% of applications, so the higher rate sat in a small group.

Then test each check inside the outside-score terciles, as well as the count. ID changes, registration changes and both city-gap checks had higher default in all three terciles. Credit over goods price showed higher default in the low and middle terciles, with the lift confined to those terciles. Phone changes were nearly level with the rest in the low tercile and higher in the middle and high terciles in the raw comparison. At the same outside score, however, the phone check had no additional separation.

In the best-scored third, applications with three or more failed checks defaulted at 8.2%, against 2.7% for clean applications. That is about three times the clean-application rate in the same tercile. Across the whole book, adding the count moved ranking power only from 0.717 to 0.718.

The outside score already prices the phone check, and the credit-over-goods check shows its separation in the low and middle terciles. The six-check count as tested still separates default in every tercile. Keep the two address checks separate in that count.

Test the individual checks inside each outside-score tercile, then assess the count separately.

Set referral thresholds to fit review capacity

Use the count to route applications for human review, and set the threshold against the team’s capacity. On this book, referring at one or more failed checks would send about 3,500 applications per 10,000 to review. At two or more, it would send about 950.

The trade-off is straightforward: a higher threshold sends fewer cases and selects a group with higher default. At two or more failed checks, 87% of referred applications repaid. Even at three or more, most repaid.

Choosing the threshold with the highest default rate reduces review volume, but also leaves more of the count’s risk separation outside the referral group.

Compare the referral volume and repayment rate at each threshold before choosing one.

On this book, two or more failed checks refers about 950 applications per 10,000. The reviewer reads which checks fired and the details the count uses: the changed ID or registration and its date, the registered city and the living city, and the credit amount against the goods price.

Use the tested checks to guide referrals

Robin’s high stated income triggered the salary check, but Robin repaid and failed none of the six counted checks. Alex passed the outside-score cutoff, had phone, ID and registration changes in the preceding month, and defaulted.

Robin was a laborer whose stated income was about 3.1 times the cell median. Alex was in the middle outside-score tercile, where the score cutoff alone would have passed the application. The distinction is the set of checks that survived testing.

When a salary flag fires on your own book, validate the checks and thresholds before using a six-check count to refer applications at two or more failed checks.

1. Run the hard contradictions once on the whole book, count each one, and retire those the form prevents from firing.
2. Define each remaining check as a yes or no, and fix its threshold before examining default.
3. Set thresholds within the relevant cell or product, then sweep the day windows for recency checks.
4. Check direction before strength. Remove a check that points toward lower default from the risk count.
5. Validate on two random halves. Set each half’s threshold on the other half, then apply it back.
6. Compare each flagged group with its age-matched expected rate, and check loan term where a loan ratio could reflect it.
7. Test each check inside outside-score terciles, and identify checks the score already prices.
8. Count the six checks as tested, keeping the two address checks separate. Refer at two or more failed checks for human review, and leave the decline decision to the reviewer.

Questions about application fraud checks

Does a salary that does not fit the job mean a fake application?

On this book, higher stated income relative to the occupation-and-education cell pointed toward lower default at every tested threshold. The 99th-percentile check also pointed that way in both halves.

Why do hard contradictions in an application almost never show up?

The form computes or limits fields such as family size and employment details, so the mismatch those checks look for cannot be entered.

How recent is a suspicious ID or registration change?

The tested count uses changes within 30 days. Registration changes within seven days gave the sharpest result in the recency sweep on this book.

Is a loan far above the goods price a fraud sign?

The ratio predicted higher default on this book. The lender’s data does not show what makes up the gap, so the ratio alone does not establish fraud.

Do these checks add anything to a credit bureau score?

The count separated default inside every third of the lender’s outside scores and moved whole-book ranking power only from 0.717 to 0.718.

Should a failed check decline the application?

No. Use the count to refer the application for human review. Even the highest-count groups include applicants who repaid.

Put these checks into a decision flow

Create rules for credit, fraud or collections, then run them on live decisions.

Start free See pricing