Use casesFinancial institutions & fintech
Application scorecard for new-to-bank applicants
Most applicants are new to you, the bureau file is thin, and your only outcomes are for the people you approved.
- Kinds of AI
- Models
- Who decides
- Credit policy committee
- Where it runs
- Your servers, banking data kept in Türkiye
Out-of-time test
| Method | Ranking power (Gini) | Stability |
|---|---|---|
| Logistic, WoEchampion | stable | |
| Gradient boostingchallenger | stable | |
| Current rulestoday | — |
ProposedDeploy the logistic scorecard; boosting kept as the monitoring challenger
Swap set at the proposed cutoff
- Approved before → now declineda small group, high expected bad rate
- Declined before → now approveda larger group, bad rate near the current book
- Model
Why a person decidesThe cutoff and the strategy are the credit policy committee's.
DecidesCredit policy committee
1. The desk
Today
- ScoreNew segment or new product: applications are scored on rules and a generic bureau score.
- SortGood young applicants and new sole traders are declined; some bad ones still pass.
- LearnPerformance data exists only for approved customers, so every rebuild learns from a biased sample.
- DecideThe credit policy committee sets cutoff and strategy from that biased picture.
With the system
- ScoreThe scorecard scores each application from the form, the bureau, and consented alternative data, with reason codes.Model
- SortThe bank's engine decides only inside the approved cutoff; the grey zone goes to underwriters, and each decline carries its top reasons.Your rule
- LearnReject inference starts from a bureau retro study, and the development sample states what is inferred.Model
- DecideThe credit policy committee sees the scorecard beside today's rules, the PD by band, and the swap set, and sets cutoff and strategy, as before.Person decides
What lands on the desk
The policy committee sees the new scorecard beside today's rules on applications from a later period, which drivers move the score, and who would change side at the proposed cutoff. Each decline carries its top reasons for the customer notice.
The credit policy committee approves cutoff, strategy, and hard-decline rules, the validation unit approves the model, and the head of retail credit risk signs after independent validation. The bank's engine decides only inside the approved cutoff; overrides are logged, and the override rate is capped and reviewed monthly.
Bad definition
90+ days past due in 12 months
Out-of-time test
| Method | Ranking power (Gini) | Stability |
|---|---|---|
| Logistic, WoEchampion | stable | |
| Gradient boostingchallenger | stable | |
| Current rulestoday | — |
ProposedDeploy the logistic scorecard; boosting kept as the monitoring challenger
Top drivers
- positive: months since oldest facility
- negative: revolving utilization
- positive: check index band
- negative: months with negative net inflow
Swap set at the proposed cutoff
- Approved before → now declineda small group, high expected bad rate
- Declined before → now approveda larger group, bad rate near the current book
- Reject inferencea bureau retro study for most rejects; parcelling as a sensitivity test only
- Decline reasonsthe top 3 negative drivers per decline, for the customer notice
- Model
Why a person decidesThe cutoff and the strategy are the credit policy committee's.
DecidesCredit policy committee
2. The system
For your technical team — skip it if you like.
How it works
In
- Application form
- Credit bureau (KKB)
- Consented alternative data
Code · calculates
Data build
Coarse classing, WoE, monotonic bins, stability checks
Reads
- Everything under In
Out
- Candidate characteristics
Model · estimates
Reject inference
Bureau retro study first; parcelling as a sensitivity test
Reads
- The characteristics
- Approved-book outcomes
- Reject outcomes from the bureau
Out
- A development sample
- A note on what is inferred
Model · estimates
Scorecard and challenger
Logistic on WoE against a monotonic GBM; SHAP reasons
Reads
- The development sample
Out
- A score
- Reason codes
Model · estimates
PD calibration
Long-run central tendency; Platt or isotonic for the GBM
Reads
- The score
Out
- PD by band
- Out-of-time results
- The swap set
Decision
Credit policy committee and validation unit
Reads
- The score and the PD
- The policy rules
Out
- Cutoff and strategy
- The bank's engine decides inside them
- The grey zone goes to underwriters
None is needed: the bank's rules engine and the approved scorecard decide inside policy, which is ordinary practice, and a digital form with bureau data needs no language model either.
How the models are trained and tested
Your expert writes the test first. A step the test does not need is skipped, and we say why.
Show the training steps
Application scorecard
Target
Bad means 90 or more days past due within 12 months (60 or more as an early read), chosen from roll-rate and vintage analysis; indeterminates excluded.
Data
Development, out-of-sample, and out-of-time samples. With no internal history: a bureau-based start plus expert rules, a pooled or bureau-retro sample, or a controlled test cell opened below the cutoff to obtain unbiased outcomes.
Method and challenger
No method is presumed better. The GBM is promoted only if its lift is material and stable out of time, and its explanation cost is accepted.
Validation evidence
For the bank's independent unit: Gini, the area under the ROC curve (AUC), and the Kolmogorov–Smirnov statistic (KS) out of time, calibration by grade (binomial, Jeffreys), population and characteristic stability indices (PSI, CSI), swap-set analysis, and bias tests across sensitive groups.
Monitoring
Weekly PSI and approval rate; monthly early delinquency by band; a quarterly calibration check (expect faster drift under high inflation); annual revalidation.
3. Value and conditions
Where the time goes, and what your institution will measure.
Where the value comes from
- Approval rate at the same bad rate, or bad rate at the same approval rate.
- Fewer manual reviews.
- Instant decisions for clear cases.
What your institution measures
- Gini, AUC, and KS out of time
- Approval rate
- Bad rate by vintage at 3, 6, and 12 months
- Share decided automatically
- Override rate
- Population stability index (PSI)
In about 250,000 observations from a German e-commerce company that ships goods before payment, adding ten simple digital-footprint variables to the credit bureau score raised the AUC from 68.3% to 73.6%, 5.3 percentage points higher than the bureau score alone.
Measured in the Decision Blueprint
Retro-scoring 12–24 months of applications and computing the swap set at today's approval rate sets the baseline; your institution sets the target, confirmed in a controlled test cell.
Conditions
The rules, the data, and the known error, with its guard.
- Scoring natural persons is high-risk under the EU AI Act (Annex III 5(b)); GDPR Art. 22 and KVKK Art. 11 give a right to object to solely automated results.
- Credit bureau (KKB) data use follows member contracts and permitted purposes; alternative data needs a KVKK lawful basis, and the open-banking scope must be checked.
- The banking regulator's (BDDK) consumer-credit caps sit outside the model as policy rules.
- Reject inference is a sensitivity test, not a fix.
Start here
We retro-score your recent applications against today's rules, check which outcomes exist for declined applicants, and size the swap set before any model is built.
Similar desks
- Models
Risk-based pricing with price response
TodayYour price covers risk and capital, yet you lose good customers over a few basis points and discount those who would have signed anyway.
With the systemA price grid that shows how each segment responds to a different rate.
Pricing committee decides
- ModelsSLMsAgentic AI
TodayReferred files wait while analysts rebuild the firm from four systems and last year's committee notes.
With the systemOne file with the rating, its reasons, and the authority level.
Credit manager decides
Equalize the gap.
Bring one decision. Leave knowing the first step.
A short description of the decision is enough; no confidential records.
https://eqlx.ai/use-cases/application-scorecard · Printed copy; the website is the current version.