The short answer
A bureau score estimates default risk from a person’s credit record across lenders. An application score estimates it for one lender’s product at the point of applying, usually by combining bureau data with application details. A behavioural score estimates it for an existing customer, mainly from how the account has run on the lender’s own books. The three differ in who builds them, what goes in and when they are used, so the same number on two of them does not mean the same thing.
Why it matters for underwriters
Each score has a blind spot. Knowing which score is in front of you tells you what it cannot see.
A bureau score does not know the applicant’s current income. An application score is a snapshot and says nothing about what happened after the decision. A behavioural score built only on internal data cannot see a missed EMI at another lender.
A common error is using a score for a question it was not built to answer. Getting this right is part of credit scoring basics for both personal loans and credit cards. Cut-offs, and the weight each score carries, vary by institution.
How it works
A score is a number from a statistical model that ranks borrowers by their likelihood of default. It ranks risk. It does not by itself tell you how many accounts will default. That link has to come from observed outcomes.
| Bureau score | Application score | Behavioural score | |
|---|---|---|---|
| Question it answers | How risky is this person across the credit system? | How risky is this applicant for this product at this lender? | How risky is this existing customer now? |
| Who builds it | The credit information company (CIC), a firm registered with RBI to hold credit data | The lender, or a vendor working for it | The lender, or a vendor working for it |
| Main inputs | Credit data held by the CIC, such as repayment record, balances and account types | Bureau data plus application details such as income, employment and loan terms | Internal account data such as days past due (DPD, days a payment is late), card utilisation and payment patterns |
| When it updates | Each time a report is pulled, on data the CIC holds by then | Fixed at the decision | Recalculated on a regular cycle |
| Who it needs | A person with enough record at the CIC | An applicant the model was built for | A customer with enough history on the book |
| Typical use | Screening, an input to pricing, portfolio monitoring | Approve, decline or refer, plus limit and pricing at origination | Card limit changes, personal loan top-ups, pre-approved offers, collections priority |
| Blind spot | Income and current repayment capacity | Anything after the decision | Other lenders, unless bureau data is added to the model |
The boundaries are not fixed. Some behavioural models take in refreshed bureau data, and some lenders run a separate application model for first-time borrowers. Ask what went into the model before you interpret its output.
India specifics
The source here is the Reserve Bank of India (Non-Banking Financial Companies – Credit Information Reporting) Directions, 2025, updated as on July 01, 2026. The body of these Directions does not define any of the three scores. It governs the data under a bureau score and the use of credit information reports (CIRs, the record a CIC issues on a borrower) in appraisal.
Who it covers. NBFC-D, NBFC-ICC, NBFC-Factor, NBFC-MFI, NBFC-IFC, IDF-NBFC, HFC, NBFC-P2P and Core Investment Companies, which it calls credit institutions (CIs). It excludes Mortgage Guarantee Companies, NBFC-AAs, Standalone Primary Dealers, NOFHCs and NBFCs with no customer interface. Banks are not on the list, so bank staff should check the direction that applies to them. Everything below is for covered NBFCs and HFCs only.
What these Directions say that affects how you read scores:
| Paragraph | Requirement | Why a score reader cares |
|---|---|---|
| 4(2), 6, 10(1) | Four CICs are registered: CRIF, Equifax, Experian and TransUnion CIBIL. A CI must be a member of all of them and submit data to all of them. | Each CIC should receive the CI’s data, though scores remain separate outputs. |
| 10(2) | Data is submitted as on the 9th, 16th, 23rd and last day of the month. The full month-end file is due by the 5th of the next month. Incremental files are due within 4 calendar days of the other dates. | A bureau score reflects what has reached the CIC, so there is always some lag. |
| 12(3) | Part instalments overdue are reported as they are. Filters on amount and period are left to those who use the data. | Deciding which small overdues to ignore is the user’s job. |
| 12(4) | ‘Settled’ is only for a principal or interest waiver given because of financial inability to repay. Wrong debits and contested charges are reported as ‘disputed’. | A CI needs to know of a settlement before extending fresh credit, so read the status as well as the score. |
| 20 | Loan policies or appraisal processes must provide for obtaining CIRs from one or more CICs. | The requirement is about the report. It does not name a score. |
| 38(5) | Listed as a best practice: first-time borrowers’ applications should not be rejected just because they have no credit history. | A missing bureau score is not a low bureau score. |
Worked example, made-up numbers
Everything in this section is invented. All three scores use an invented scale of 0 to 100, where higher means lower risk. Real scales differ. An NBFC has three files on its desk.
| Case | Situation | Bureau | Application | Behavioural | How to read it |
|---|---|---|---|---|---|
| A | New customer, personal loan | 72 | 68 | None | Good external record, with application details pulling the view down a little. Use the application score for the decision and check the report. |
| B | Existing card customer, limit increase | 58 | Not run | 81 | Clean on the lender’s books, weaker elsewhere. The two scores disagree, so pull the report before raising the limit. |
| C | First-time borrower, personal loan | None | 61 | None | No bureau record. Assess on application data and documents, and do not treat the gap as a negative. |
In Case B, the report shows a personal loan at another lender at 30 DPD. The behavioural score could not see it. That is the blind spot at work.
Now take the same score band on two models. Bad rate here means the share of accounts that hit the lender’s definition of default within the period it tracks.
| Score band | Model | Accounts | Went bad | Bad rate |
|---|---|---|---|---|
| 60 to 69 | Application | 1,000 | 60 | 6.0% |
| 60 to 69 | Behavioural | 1,000 | 25 | 2.5% |
The same band carries 2.4 times the bad rate on the application model (6.0 / 2.5). The populations differ: behavioural accounts have already run on the book for a while, while applicants have no such track record. A score only means something against outcomes from its own model.
Pitfalls
- Comparing across models. A number on one model cannot be set against the same number on another, including bureau scores from different CICs.
- Reading the score and skipping the report. Under paragraph 12(4) of the Directions above, a ‘settled’ status is something a covered NBFC needs to know before lending again. A single number can hide it.
- Counting bureau data twice. If the application score already uses bureau data, a good bureau score and a good application score are not two independent confirmations.
- Using a behavioural score alone for fresh exposure. Check what the model sees outside your own book first.
- Reusing an old application score. It describes the applicant at the decision date. Later decisions need current data.
- Ignoring reporting lag. A missed payment just after a reference date appears only with a later submission. Rejected records can also lag, since paragraph 15 of the Directions lets a CI resubmit them along with the next reference date’s data.
- Trusting a score built on wrong data. Under paragraphs 19 and 35 of the Directions, errors are corrected at source by the CI that submitted them. The CI gets 21 calendar days, the overall limit is 30, and delay beyond that earns the complainant ₹100 per calendar day.
- Assuming a model stays accurate. Populations and products change, so compare scores with actual outcomes on a regular basis.
Checklist
- Which score is this, and who built it?
- What data went in, and does it include bureau data?
- What is the score date, and is it still valid for this decision?
- What can this score not see, and have I covered that gap some other way?
- Have I read the CIR itself, including any ‘settled’ or ‘disputed’ status?
- Am I comparing this number only with outcomes from the same model?
- If the bureau and internal views disagree, have I found out why?
- If there is no bureau record, is the file being assessed on other data rather than rejected for that reason alone?
- At a covered NBFC or HFC, does the appraisal process provide for CIRs from one or more CICs, as paragraph 20 of the Directions requires?
- Am I applying my institution’s own limits, given that they vary by institution?
Sources
- Reserve Bank of India (Non-Banking Financial Companies – Credit Information Reporting) Directions, 2025, rbi.org.in, accessed 2026-11-06