RBI's expected credit loss rules: what changes for banks
From 1 April 2027 Indian banks must provision for expected losses before default. What the three stages, the 5% floor and the data demands change.
For decades, an Indian bank set aside money for a bad loan only after the loan had already gone bad. From 1 April 2027 that changes. The Reserve Bank of India has finalised a framework under which banks must provision for the losses they expect, before a borrower misses the 90-day mark that makes a loan non-performing. The idea is simple to state and hard to run. Every loan needs an estimate of how likely it is to default, how much would be lost, and how much would be owed, and those estimates must reflect the economy ahead and not only the past. This deep dive sets out what the rules say, where reports disagree, what the numbers look like, and what a risk or data team has to build. I could not open the RBI's own text, so every figure below is taken from named secondary sources and the points of disagreement are marked.
Current situation
The numbers first, each with the party that stated it. Where sources differ, I say so.
Dates. The RBI Governor said on 1 October 2025 that the framework was proposed to start on 1 April 2027, as [Business Standard](https://www.business-standard.com/amp/finance/news/ecl-framework-proposed-to-be-implemented-from-april-1-2027-rbi-guv-1251001005111.html) and others reported. Draft directions followed on 7 October 2025 for consultation. A rating agency note dated April 2026 and a law-firm newsletter both report that the final directions were issued in April 2026, and Uniqus gives the date as 27 April 2026, with the start date unchanged at 1 April 2027. Reports say the RBI turned down requests from lenders for more time.
Who is covered. One search summary of the ICRA note says the framework applies to scheduled commercial banks and all-India financial institutions, and excludes small finance banks, payments banks and regional rural banks. Another source, JM Financial, says the framework applies to banks and that non-bank lenders are covered separately by the Income Recognition, Asset Classification and Provisioning Directions of 2025, effective from 28 November 2025. Treat the exact scope as something to confirm in the RBI's text.
The three stages. The framework follows the three-stage model of the international accounting standard IFRS 9, which took effect globally in 2018. Stage 1 holds loans with no significant rise in credit risk, and provisions cover the expected loss over the next 12 months. Stage 2 holds loans whose credit risk has risen significantly but which have not defaulted, and provisions cover the expected loss over the whole life of the loan. Stage 3 holds credit-impaired loans, broadly the existing non-performing assets, which stay defined by the 90-day overdue test.
Floors. Banks must apply minimum provisions, called floors, whatever their models say. Uniqus lists a Stage 2 floor of 5 per cent for corporate, MSME, retail (secured and unsecured), farm credit and loans to banks and non-bank lenders, with lower floors for some products: 1.5 per cent for housing and gold loans, 2.5 per cent for loans secured by commercial real estate and 0.4 per cent for loans backed by term deposits. Restructured loans still classed as standard carry 5 per cent in Stage 1 and 10 per cent in Stage 2, according to the same source. Today's provision on standard assets is about 0.4 per cent. CRISIL, JM Financial and the search summaries agree on the 5 per cent figure for most assets.
The size of the hit. CRISIL estimates a one-time net impact of up to 120 basis points on the banking system's core capital (CET-1) ratio, from a gross impact of up to 170 basis points. The gap is explained by provisions banks have already made, including contingency provisions at a few large banks. CRISIL puts the system's CET-1 ratio at about 14 per cent at 31 March 2026 and the sector's return on assets at about 1.25 to 1.3 per cent in the last fiscal year. JM Financial quotes an estimate, with the source not named, of about 60 to 70 basis points for the sector, 60 to 90 for public sector banks and 20 to 50 for private banks. ICRA's earlier October 2025 note on the draft projected a 300 to 400 basis point cut in core capital, and its April 2026 note on the final rules calls the effect likely moderate. The estimates vary because they use different assumptions, which is itself a finding.
The glide path. Banks can spread the cost over four fiscal years, setting the extra provisions against retained earnings, according to CRISIL. Uniqus describes a taper in which the share of the one-time hit that can be added back to capital runs 4/5, 3/5, 2/5 and 1/5 from FY2027-28 to FY2030-31, with the full effect from FY2031-32. JM Financial refers to a transition to March 2031. Neither source reproduces the RBI schedule, so the exact mechanics should be checked in the directions.
What is this research about?
This is not a study in the academic sense. It is a close reading of a regulation and the commentary around it, with an aim: to explain the change in provisioning, the models it demands, and what is reliably known so far.
From incurred loss to expected loss
Under the old rule, the incurred-loss approach, provisioning was driven by classification. A loan was standard, sub-standard, doubtful or loss, depending mainly on how long it had been overdue, and the provision percentage followed the class. A loan 85 days past due and a loan that was fully current might both carry the same small standard-asset provision, even though the first was far more likely to default. The big step up in provisioning came only at the 90-day boundary.
Under the expected-loss approach, the provision is the expected loss itself. In the words used in our Credit Risk Analytics track, it is PD multiplied by LGD multiplied by EAD, where PD is the probability of default, LGD is the loss given default and EAD is the exposure at default. JM Financial confirms that banks must use probability-weighted estimates across several macroeconomic scenarios, such as base, stress and optimistic. The result is that provisions should start to build as risk rises, well before default.
What the three stages do
Stage · Meaning · Loss horizon · Trigger (as reported) · Provision floor (most loans)Stage 1 · No significant rise in credit risk · 12 months · Performing loans · About 0.4% (floor broadly like today's)
Stage 2 · Significant rise in credit risk, not yet defaulted · Lifetime · 30 days past due is a rebuttable presumption for term loans, per Uniqus · 5%
Stage 3 · Credit-impaired · Lifetime · 90 days past due, unchanged · Broadly the current NPA provisioning
The table shows two places where the sources disagree. On the trigger, Uniqus says that 30 days past due on a term loan is a rebuttable presumption of a significant increase in credit risk, with a 60-day equivalent for revolving facilities. A rebuttable presumption means the lender can argue against it, but only with evidence approved by its board and applied consistently. JM Financial, by contrast, says Stage 2 will be "largely" 60 to 90 days overdue, and gives no explicit 30-day trigger, while one other summary also describes the 30-day presumption. I cannot settle this from the material I opened, and the difference matters, as the simulator below shows. On Stage 3, CRISIL notes that current provisioning on non-performing assets is already about 76 per cent, so the step-up there is smaller than at Stage 2.
The step that matters most: Stage 2
The headline change is the move from about 0.4 per cent to a 5 per cent floor at Stage 2, a rise of 4.6 percentage points, or more than twelve times. CRISIL notes that the impact is contained because loans that are overdue by 1 to 60 days, called SMA-1 and SMA-2, make up only about 2 to 2.2 per cent of the banking system. A big increase on a small base is a small number overall. The number is far from small for a lender with a weaker book.
At the defaults, a bank with a ₹1,00,000 crore book needs about ₹97 crore of extra provisions, which is about 9.7 basis points of the book or 7.7 per cent of a year's profit. Now raise the Stage 2 share to 8 per cent, as a weaker unsecured lender might have, and the cost climbs to about 37 basis points, almost 30 per cent of a year's profit. The same rule bites very differently depending on the book.
How will this impact you?
If you borrow. The rules do not directly change what you owe. The indirect effects come through pricing and appetite. A loan that carries a higher expected loss costs the lender more to hold, so unsecured lending, which gets a high floor and a high risk weight, may look relatively more expensive. A reported 125 per cent risk weight on unsecured personal loans and card dues, from the capital rules that also take effect on 1 April 2027, points the same way. Whether lenders pass this on is a business choice that the sources do not settle.
If you work at a bank or lender. The rules are a large project. Uniqus describes banks needing a board-level committee including the finance and risk heads, an independent model validation function and, for the first time at many banks, lifetime PD curves. A bank must also value its whole loan book at fair value on 1 April 2027, with the difference going to opening retained earnings, and Uniqus says quantification is expected by December 2026. For risk analysts and data scientists, this is demand for exactly the skills taught in the Credit Risk Analytics track.
If you invest. The effect on bank profits is one-time and spread over years, but the structural effect continues. CRISIL says credit costs are likely to rise somewhat for good, because delinquent assets not yet classed as non-performing also need higher provisions. Some listed public sector banks fell on the confirmation of the date, as [5paisa reported](https://www.5paisa.com/news/psu-bank-stocks-fall-as-rbi-confirms-ecl-norms-from-april-2027), though one day's move says little about the long-run effect.
If you run a non-bank lender. Larger non-bank lenders that follow Indian accounting standards have long used an expected-loss approach under the accounting standard, and the bank-focused framework is not written for them. JM Financial points to separate directions of 2025 for non-bank lenders that bring their norms closer to the banks'. Smaller lenders should check which rules apply, because the picture differs by size.
What does it demand from the data and the models?
This section matters most for analysts. Uniqus's account of the implementation challenges lines up neatly with the pieces of a credit model.
Component · What the framework needs · The difficulty describedPD · A point-in-time PD that reflects today's conditions, and lifetime curves for Stage 2 · Banks that use through-the-cycle PD need a separate point-in-time layer, and lifetime curves ideally need at least one full credit cycle of internal default data
LGD · Recovery estimates, by product · Recovery data is split across legal, restructured and written-off pools; workout recoveries on corporate loans can take over five years; proxy curves and external benchmarks will be needed in the first one or two reporting cycles
EAD · Exposure including undrawn limits · Standard conversion factors from the capital rules are unlikely to be enough for revolving and committed facilities
Staging · A rule that flags a significant rise in risk · Needs a PD-based or days-past-due test with documented evidence for any rebuttal
Data and systems · A borrower-level view across products · Stage 3 is assessed per borrower, so a bank with separate systems per product needs a combined borrower view before go-live
Governance · Independent validation, a model inventory, formal overlays · Gaps include validation teams that report through the development chain and adjustments documented as memos
Two ideas from this table deserve a plain explanation.
Point-in-time versus through-the-cycle. A through-the-cycle PD is an average over good and bad years, stable and slow to move, and it suits capital planning. A point-in-time PD reflects current conditions, so it rises in a downturn and falls in a boom. An accounting provision should reflect the economy now and in the near future, so it needs a point-in-time view. A bank that has only a through-the-cycle model must scale it, with evidence from back-testing that the scaling works.
Lifetime versus 12-month. A loan in Stage 1 needs only the loss expected in the next year. A loan in Stage 2 needs the loss expected over its entire remaining life, which for a 20-year home loan is a very long horizon. The difference between the two is the cliff in provisioning. Our Credit Risk Analytics track covers PD, LGD and EAD in Module 1, vintage curves and migration matrices in Module 4, and the regulation itself in Module 10.
How the staging trigger changes the picture
The reported 30-day presumption versus a looser 60-day reading is not a technicality. The following simulator uses an illustrative snapshot of an unsecured book in which 4.4 per cent of loans are more than 30 days past due, 2.5 per cent are more than 60, and 1.2 per cent are more than 90 and so in Stage 3. Stage 2 is the band between the trigger and 90 days.
Move the trigger from 30 to 60 days and the Stage 2 share falls from 3.2 per cent to 1.3 per cent, and the extra provision from about ₹14.7 crore to about ₹6.0 crore on a ₹10,000 crore book. A day-count rule is a lever on provisions, which is why the board-approved evidence needed to rebut it matters.
Current limitations
Several things cannot be established from what is publicly available at the time of writing.
The RBI text itself. I could not open the directions, so every number comes through secondary sources. Some pages were unreadable or blocked, including an ICRA note whose slides did not extract.
Conflicting accounts of the trigger. One source says 30 days past due is a rebuttable presumption, and another says Stage 2 is largely 60 to 90 days. The correct reading is in the directions.
Floors by product. Only one source gave the full product list. The Stage 1 floor is described differently across sources: one says 0.4 per cent, another gives a range of 0.25 to 0.40 per cent under current norms, and CRISIL says only that it is broadly similar to the present norms.
Impact estimates. They range from about 60 to 120 basis points for net impact in the sources I found, and the earlier draft estimate was far higher. The difference is explained by different assumptions about portfolio mix, tax and existing provisions, and by the move from draft to final rules.
Scope. Sources differ on which banks are included and on how non-bank lenders are treated.
Limitations and open questions
Even with the rules in hand, the evidence cannot yet tell us several things.
How models will behave. Most Indian banks have never had to produce lifetime PD curves at scale. Whether their first models will be stable, back-test well and survive independent validation is unknown until parallel runs report. Uniqus says a parallel run begins in the second quarter of FY27, which is now.
How much will be rebutted. If the 30-day presumption applies, how often banks will rebut it, with what evidence, and how supervisors will respond, will decide the real size of Stage 2.
Effects on lending. It is a reasonable expectation that unsecured and small-business lending become costlier, but no source I found measures it, and expectations are not evidence.
Procyclicality. Because provisions follow the economy, they can rise in a downturn just when banks have least capital. The glide path and the floors are the regulator's tools against this, and whether they are enough is a matter of judgement.
Data quality. The framework relies on years of clean default and recovery data. How much of that exists, in usable form, is not public.
What is coming next
Dated items, with forecasts labelled.
Day-1 fair valuation and parallel run. Uniqus describes quantification being expected by December 2026 and independent valuation evidence in place by the third quarter of FY27. A parallel run starts in the second quarter of FY27.
1 April 2027. The framework starts. The same date applies to revised capital risk weights for certain assets, including 75 per cent for eligible retail loans and 125 per cent for unsecured personal loans and revolving credit card dues, reported by [Outlook Business](https://www.outlookbusiness.com/economy-and-policy/banks-cant-delay-bad-loan-recognition-now-heres-what-rbis-2027-credit-loss-norms-mean). CRISIL notes that any capital released by revised risk weights is not in its impact estimate and would partly offset the cost.
FY2027-28 to FY2030-31. The reported transition years, with the framework fully loaded in FY2031-32 according to Uniqus.
31 March 2030. The date Uniqus gives for the existing loan book to move to effective-interest-rate income recognition, with new loans on that basis from 1 April 2027.
Forecasts, labelled as such. CRISIL expects credit costs to rise somewhat permanently, and analysts expect a one-time hit that varies widely across banks. These are views, not facts.
The wider backdrop for models
Expected-loss models sit inside a wider debate about how much to trust complex models. Three points from outside India help frame it, each from a single source and to be confirmed before use. A practitioner article reports that US regulators revised their model risk guidance in April 2026 and withdrew the 2011 framework known as SR 11-7, replacing it with a risk-based, non-binding approach; other sources still describe SR 11-7 as current. A European policy paper argues that the explanation methods widely used to justify credit scoring decisions are after-the-fact, and that regulators should treat them as part of the infrastructure of credit markets. And one source says EU lawmakers agreed in May 2026 to push back the high-risk obligations for credit scoring systems under the AI Act to December 2027, while others still give an earlier date. The common thread is that model governance, not only accuracy, is becoming the limit on how fast lenders can adopt more advanced models.
Closer to home, bureau data feeds the same models. The RBI has required lenders to report credit information to credit bureaus fortnightly, with reference dates of the 15th and the last day of the month, since 1 January 2025. One report says reporting moves to weekly from 1 April 2026, but I could not confirm that from an RBI document. Fresher bureau data matters because delinquency signals reach a bank's model sooner.
Key takeaways
From 1 April 2027, banks must provision for expected credit losses in three stages, before the 90-day mark that defines an NPA.
The most discussed change is the reported 5 per cent floor at Stage 2, up from about 0.4 per cent, with lower floors for housing, gold and some secured loans.
CRISIL puts the one-time net hit on banks' core capital at up to 120 basis points, gross up to 170, spread over four fiscal years. Other estimates differ.
Sources disagree on whether 30 days past due is a rebuttable trigger for Stage 2. That one detail changes the size of Stage 2 severalfold.
The hard work is analytical: point-in-time PD, lifetime curves, LGD from fragmented recoveries, EAD for undrawn limits, borrower-level data and independent validation.
Evidence assessment
I rated confidence Medium. The date of 1 April 2027 and the headline 5 per cent Stage 2 floor are reported by several independent outlets and an official statement by the Governor, which supports the main story. The glide-path mechanics, the Stage 2 trigger, the scope and the product floors rest on secondary sources that sometimes disagree, and I could not read the RBI's directions or the ICRA note. Impact estimates come from rating agencies and a source that is not named, and they differ by a factor of two. The model and governance material is partly from a vendor and a consulting firm. Items from outside India each come from one source. The article therefore reports what is stated, flags where it conflicts, and avoids claims beyond it.
Suggested follow-up
Read the RBI's final directions on expected credit loss and asset classification on rbi.org.in, and check the trigger, the floors table and the glide path against this article.
Read the ICRA and CRISIL notes in full for the sector estimates and their assumptions.
Watch the second and third quarter results of listed banks for early disclosures of parallel-run impact or extra provisions.
If you build credit models, work through Modules 1, 4 and 10 of the Credit Risk Analytics track on this site.
Sources
- [Business Standard: ECL framework proposed to be implemented from April 1, 2027, RBI Governor (1 October 2025)](https://www.business-standard.com/amp/finance/news/ecl-framework-proposed-to-be-implemented-from-april-1-2027-rbi-guv-1251001005111.html)
- [CRISIL Ratings: For banks, ECL norms to have one-time net impact of up to 120 bps (April 2026)](https://www.crisilratings.com/en/home/newsroom/press-releases/2026/04/for-banks-ecl-norms-to-have-one-time-net-impact-of-up-to-120-bps.html)
- [Uniqus: The implementation reality check (stages, floors, trigger, glide path)](https://uniqus.com/the-implementation-reality-check/)
- [JM Financial: RBI unveils final ECL framework for banks](https://www.jmfinancialservices.in/blogs-and-articles/rbi-unveils-final-ecl-framework-for-banks)
- [K&S Partners: RBI finalises income recognition and asset classification directions](https://ksandk.com/newsletter/rbi-ecl-framework/)
- [ICRA: Indian banking industry, RBI releases final ECL framework (April 2026; slides could not be read)](https://www.icra.in/Rating/DownloadResearchSpecialCommentReport?id=6908)
- [Outlook Business: What RBI's 2027 credit loss norms mean](https://www.outlookbusiness.com/economy-and-policy/banks-cant-delay-bad-loan-recognition-now-heres-what-rbis-2027-credit-loss-norms-mean)
- [Taxmann: RBI directs credit information companies to update credit information fortnightly](https://taxmann.com/post/blog/rbi-directs-credit-information-companies-institutions-to-update-credit-information-fortnightly-or-at-shorter-intervals)
- [SAFE Policy Letter 114: explainability in credit scoring](https://safe-frankfurt.de/fileadmin/userupload/editorcommon/PolicyCenter/SAFEPolicyLetter114.pdf)
Tags: Expected Credit Loss RBI Credit Risk Bank Provisioning IFRS 9