Reject inference for a credit scorecard

Run reject inference for a credit scorecard by entering goods, bads and rejects for each score band.

Score bands
Score bandGoodsBadsRejects

The demo uses eight made-up score bands.

Enter one band per line with the label, goods, bads and rejects separated by tabs, commas or semicolons.

Reject inference

Rejects in each band are split into goods and bads at the accepts' bad rate times the uplift, in whole applicants.

Through-the-door bad rate

18.3%
Accepts' bad rate5.0%
Rejects' inferred bad rate47.8%
Share accepted69.0%

With rejects inferred at 47.8%, all applicants go bad at 18.3%, compared with 5.0% for accepts alone.

Bad rates by score band

Accepts onlyAll applicants, inferred

Inferred counts

Score bandGoodsBadsAccepts' bad rateRejectsInferred goodsInferred badsRejects' bad rateAll applicants' bad rate
< 520402235.5%1,9005521,34871.0%69.8%
520–5591806125.3%1,40069170950.6%46.9%
560–59962011816.0%90061228832.0%24.8%
600–6391,3501429.5%5204219919.0%12.0%
640–6792,1001155.2%2602332710.4%5.7%
680–7192,600782.9%12011375.8%3.0%
720–7592,300351.5%605823.0%1.5%
760+1,800120.7%303001.3%0.7%
Total10,9905835.0%5,1902,7102,48047.8%18.3%

Get the result by email

Build the scorecard inside the credit decision and test it on the lender's own past applications, accepted and rejected, before it goes live. Start your trial

Show formulas
Accepts' bad rate = Bads ÷ (Goods + Bads)
Parcelling: Rejects' bad rate = min(1, Accepts' bad rate × Uplift factor)
Fuzzy augmentation: Rejects' bad rate = Accepts' bad rate
Inferred bads = Rejects × Rejects' bad rate
Inferred goods = Rejects − Inferred bads
All applicants' bad rate = (Bads + Inferred bads) ÷ (Goods + Bads + Rejects)

The rejects' outcomes cannot confirm the assumption used to infer them, so the uplift is the modeller's judgement. Parcelling with an uplift of 1 gives the same totals as fuzzy augmentation.

Reject inference for scorecards

Reject inference adds rejected applicants to scorecard data using an assumed outcome for each reject. The tool works out inferred goods and bads by score band using parcelling or fuzzy augmentation.

A scorecard trained on accepts learns from applicants the old policy accepted. Compare the accepts' bad rate with the through-the-door bad rate, then use the inferred counts when building and testing the scorecard.

Reject inference questions

What is reject inference in credit scoring?

Reject inference adds rejected applicants to scorecard data with an assumed outcome. The assumption cannot be confirmed from their outcomes because those outcomes are unknown.

Which reject inference methods can I compare?

This tool compares parcelling and fuzzy augmentation. Parcelling applies an uplift to the accepts' bad rate, while fuzzy augmentation uses the accepts' bad rate for each band.

How does parcelling work in reject inference?

Parcelling estimates the rejects' bad rate in each band by multiplying the accepts' bad rate by the uplift, up to a bad rate of 1. It then assigns inferred goods and bads in whole applicants.

How do I include rejected applicants in a scorecard?

Enter goods, bads and rejects for each score band, then choose parcelling or fuzzy augmentation. The tool gives you inferred counts to use when building the scorecard.

What does the uplift factor mean in reject inference?

The uplift multiplies the accepts' bad rate in each band to estimate the rejects' bad rate under parcelling. Choosing the uplift is the modeller's judgement.