Part 08 — Credit Risk Deep Dive

Scorecards Models

Scorecards turn credit judgement into repeatable ranking. They do not replace underwriting; they force consistency in how information is weighted. In Indian SME lending, a practical model stack usually has three layers: hard policy rules, application scorecard, and behavioural or portfolio scorecard. The first says “can we lend at all”; the second says “what is the risk at origination”; the third says “how is the borrower behaving after disbursement”.

This page links the scorecard output to cash-flow assessment, portfolio analytics and provisioning under IRAC.

Hard Rules Before Scores

Hard rules should be few and explicit. Examples: RBI or internal negative list, current non-performing asset (NPA), wilful defaulter match, fraud match, forged document, prohibited geography, minimum age, minimum vintage, active GSTIN where product requires GST, and product-level exposure cap. A scorecard should not “average out” a fraud alert with strong turnover.

Bureau variables are central. TransUnion CIBIL’s Commercial Rank classifies commercial borrowers into CMR-1 to CMR-10, where CMR-1 is strongest and CMR-10 weakest; the newer CV CMR uses a 36-month observation window and is positioned for underwriting, pricing and portfolio management (TransUnion CIBIL CV CMR, Commercial CIR description). CRIF High Mark describes commercial reports with MSME Rank and a 300-900 commercial score designed to predict default over the next 12 months (CRIF commercial credit information). A lender should store the bureau name, score version and pull date, not merely “bureau score”.

Application Scorecard Illustration

Below is an illustrative unsecured SME business-loan scorecard. It is not a production model; it shows how variables and weights can be structured.

Variable block	Weight	Good profile	Weak profile
Bureau and repayment conduct	25	CMR 1-4 or consumer score above policy cut-off; no recent 30+ DPD	Thin file, CMR 8-10, recent settlement, multiple enquiries
Banking quality	20	Stable operating credits, low bounces, healthy average balance	Circular credits, cash stuffing, frequent cheque/NACH returns
GST or sales stability	15	Regular filing, stable monthly sales, e-invoice consistency where applicable	Delayed filings, sales spike before application, related-party invoices
Leverage and obligations	15	Low fixed obligation to income ratio (FOIR), low unsecured debt	High EMI load, many small fintech loans, credit-card stress
Business vintage and stability	10	Same premises and line for 3 or more years	New location, changed line of business, weak references
Industry and geography	5	Stable local demand, low policy risk	Negative list, commodity volatility, stressed cluster
Promoter quality	5	Own house, stable family support, clean KYC	Address instability, legal disputes, adverse market feedback
Relationship and collateral	5	Repeat borrower, security or guarantor support	New-to-lender unsecured exposure

Assume points are scaled to 1,000. A borrower gets:

Block	Max	Score
Bureau	250	180
Banking	200	145
GST/sales	150	120
Leverage	150	80
Vintage	100	75
Industry/geography	50	35
Promoter	50	40
Relationship/collateral	50	15
Total	1,000	690

Policy maps scores as:

Score band	Decision posture	Indicative pricing
780 and above	Auto-approve if no deviation	Base rate plus 0-2%
700-779	Approve with normal checks	Base rate plus 2-4%
620-699	Manual review; amount cut likely	Base rate plus 4-7%
560-619	Senior deviation only	Base rate plus 7% or secured structure
Below 560	Decline	Not applicable

The example score of 690 should not be an automatic decline. It says the file needs manual review because leverage and lack of relationship are weak. Credit may approve a lower ticket, require guarantor, shorten tenor or move to secured product.

Behavioural Scorecards

Behavioural models start after disbursement. Useful variables include first EMI bounce, days past due (DPD) cure speed, mandate type, average balance erosion, GST filing delay, reduction in monthly credits, bureau new enquiries, new loans from other lenders, utilisation of cash-credit limit, stock-statement delay, and failed collection contacts. A behavioural model should trigger actions: limit freeze, renewal review, field visit, early collection, pricing reset for future top-up, or decline top-up.

For SME lenders, the strongest early signal is often not 30 DPD; it is the first failed mandate combined with falling bank credits. If a borrower bounces once but cures in two days and banking remains healthy, risk is different from a borrower who cures after 25 days while taking three new digital loans.

PD, LGD and EAD

Probability of default (PD) estimates the chance of default over a horizon, usually 12 months or lifetime. Loss given default (LGD) estimates the percentage loss if default occurs after recovery and collateral. Exposure at default (EAD) estimates outstanding exposure when default happens. Expected loss is:

expected loss = PD x LGD x EAD

Worked example:

Case	Value
Sanctioned term loan	₹20 lakh
Outstanding after 6 months	₹18.4 lakh
12-month PD by score band	5.5%
LGD unsecured after recoveries	62%
EAD	₹18.4 lakh
Expected loss	₹62,744

Calculation: 5.5% x 62% x ₹18.4 lakh = ₹62,744. This is a portfolio expected loss, not a prediction that this exact borrower will default. Pricing must recover expected loss plus capital, operating cost and funding cost. See pricing and ALM.

For a cash-credit or overdraft line, EAD is not just today’s outstanding. If limit is ₹50 lakh and utilisation is ₹32 lakh, the borrower may draw more before default. A credit conversion factor might estimate EAD at current utilisation plus 40% of undrawn limit: ₹32 lakh + 40% x ₹18 lakh = ₹39.2 lakh.

Ind AS 109 ECL vs IRAC

Ind AS 109 impairment uses expected credit loss (ECL), with 12-month ECL where credit risk has not increased significantly and lifetime ECL where it has. The Companies (Indian Accounting Standards) Rules text requires expected credit losses to reflect probability-weighted outcomes, time value of money and reasonable supportable forward-looking information (Ind AS 109 text via Indian Kanoon mirror); ICAI’s current compendium is the accounting reference point for Ind AS material (ICAI Ind AS compendium 2025-26).

RBI prudential classification is different. As of July 2026, NBFCs following Ind AS still need to respect RBI prudential floors under the NBFC Scale Based Regulation framework; RBI’s NBFC standard-asset provisioning circular for Upper Layer NBFCs states that Ind AS impairment allowances are subject to the prudential floor (RBI NBFC standard asset provisioning, 6 June 2022). RBI has also issued a final ECL framework for commercial banks effective 1 April 2027, while retaining NPA discipline during transition; the public RBI page was not readily retrievable through search in this session, so this point is cited to a regulatory mirror and should be checked against the RBI notification repository before production policy implementation (RBI Commercial Banks Asset Classification, Provisioning and Income Recognition Directions, 2026 mirror).

Topic	Ind AS 109 ECL	RBI IRAC / prudential view
Objective	Recognise expected credit losses in financial statements	Prudential income recognition, asset classification and minimum provisioning
Stages/buckets	Stage 1: 12-month ECL; Stage 2: lifetime ECL after significant credit-risk increase; Stage 3: credit-impaired	Standard, SMA-0/1/2, NPA, sub-standard, doubtful, loss
Trigger	Change in credit risk since origination plus forward-looking information	Overdue status, out-of-order status, restructuring and borrower-level classification rules
Default horizon	12-month or lifetime depending on stage	90 DPD NPA rule remains an operational anchor
Provision math	PD x LGD x EAD, discounted and probability-weighted	Prescribed rates/floors by asset class and entity type
Income	Effective interest method; Stage 3 interest treatment differs	NPA income generally on realisation basis
Use in pricing	Directly useful through expected loss	Useful as regulatory capital/provision floor and delinquency discipline

Model Governance

Production models need governance. Minimum controls:

documented variable definitions and score version;
development, validation and approval records;
reject inference policy for declined applications;
monitoring of approval rate, bad rate, stability index and override performance;
bias and fair-treatment checks where alternate data is used;
reason codes for declines and cutbacks;
manual override caps by authority.

The model should be humble. It ranks risk using historical patterns; it cannot detect every fabricated invoice, political-event shock, commodity crash or promoter dispute. Strong lenders combine scorecards with document controls, field verification, fraud review and continuous vintage monitoring.