India · commercial banks
The ECL directions are final. Here is what has to be built.
The Reserve Bank issued its final directions on 27 April 2026. They take effect on 1 April 2027, with the provisioning impact spread to 31 March 2031. This page sets out what they ask for, and — line by line — what this product does about each one, including the parts it does not do.
Dates and scope taken from the Reserve Bank of India (Commercial Banks — Asset Classification, Provisioning and Income Recognition) Directions, 2026, checked on 5 September 2026. This is a reading of the directions, not advice on them.
What it covers
- 01Loans
- 02Debt securities not measured at fair value through profit or loss
- 03Trade receivables
- 04Lease receivables
- 05Loan commitments, including undrawn commitments
- 06Off-balance-sheet credit exposures
- 07Other financial assets with a contractual right to receive cash
Undrawn commitments and off-balance-sheet exposures being in scope is the part most spreadsheets get wrong, because a spreadsheet built around a loan tape has no row for a limit nobody has drawn.
A provisioning method has to survive being taken apart by somebody who did not build it.
Line by line, including the gaps
10 of these are built, 2 partly, 1 built but not yet reconciled to the published text, and 0 not built at all. The last three are in the same table as the rest, at the same weight, because a checklist that is all ticks tells you only that the vendor wrote the checklist.
| What the directions ask for | Here | What that means |
|---|---|---|
| Three-stage classification with a significant-increase-in-credit-risk test | Built | Staging on quantitative and qualitative criteria with the 30-day and 90-day presumptions, each exposure carrying the reason it sits where it does.See it |
| 12-month and lifetime ECL from PD, LGD and EAD | Built | The general model, discounted at the effective interest rate, with undrawn commitments converted at a credit conversion factor rather than ignored.See it |
| Forward-looking, probability-weighted scenarios | Built | Scenarios weighted to one, with the weighted result shown against each scenario so the effect of the weighting is visible rather than buried — and the systematic factor behind each one fitted to a published series rather than typed in, with the strength of that relationship reported and a weak one called weak.See it |
| Product-wise minimum provisioning floors, reported alongside the estimate | Built, not verified | Floors are a first-class concept, not an overlay: the modelled figure and the floor are held apart and the uplift is reported, because netting them hides the comparison a supervisor is asking for. Stage 3 minimums escalate with time in default.Gap. Read from the instrument — RBI/DOR/2026-27/398, paragraph 82 — and every rate is now located in its text by a test, so a transcription error fails the build rather than waiting to be noticed. Three things remain that no test can settle: whether it is still the current law on the day you use it, which of your product codes belongs to which category, and who answers to a supervisor if it is wrong. The engine keeps blocking review until somebody puts their name to it.See it |
| Purchased or originated credit-impaired assets | Built | POCI is tracked separately and disclosed as its own row rather than folded into Stage 3, which is what IFRS 7 asks for. |
| Effective interest rate as the discount rate | Built | Each exposure discounts at its own EIR rather than at a portfolio average. |
| Model validation — discrimination, calibration and stability | Built | Back-test against a period that has already finished, graded per grade rather than in aggregate, because a portfolio-level check hides an understated grade behind a conservative one.See it |
| Governance over model changes, with independent approval | Built | Parameters are under change control: edits collect in a draft and take effect only once somebody other than the author approves them. The approver and date are stamped onto every result computed under that version.See it |
| Movement in the allowance, separating the book from the model | Built | Opening to closing with each period rebuilt on the model that was in force then, so a change in the allowance caused by a change to the model cannot be presented as a change in credit risk.See it |
| Reconciliation of the population to the general ledger | Built | The loan tape is tied to the ledger before anything is measured, and an unreconciled population stops the measurement rather than being measured anyway.See it |
| PD estimation where the book has little or no default history | Partly built | Loss rates from your own history — cohort loss rates, roll rates from a series of ageing snapshots, or a survival curve — including a defensible figure for a book that has never defaulted.Gap. Calibration from external or pooled data is not built: a book with no history and no proxy still needs judgement applied outside the tool. The forward-looking half is now estimated rather than assumed — the systematic factor is backed out of the default rates the book actually saw and regressed on a series the entity did not choose.See it |
| Glide path: the transitional provisioning impact spread to 31 March 2031 | Built | The day-one adjustment — ECL at transition less the IRAC provisions carried the day before — taken to opening retained earnings rather than through profit or loss, and the declining fraction addable back to CET1 across the four years, net of tax, with the effect on the ratio in basis points.See it |
| Regulatory reporting templates in the supervisor's own format | Partly built | Returns are definitions rather than code, so a revised template or a second jurisdiction is data and not a release. All four Annex 4 disclosures are transcribed from the Directions themselves — credit quality of loans aged into the prescribed bands, of investments by rating grade, of commitments at notional, the summary by stage, and the loss-allowance roll-forward. Every total declares what it is the sum of and the evaluator checks it, because a return whose columns do not come to its total is the one that gets sent back.Gap. Annex 4 repeats its credit-quality rows per product type; what is produced is the total across all of them, and which products a bank discloses separately is its own judgement. The roll-forward cannot split write-offs from other derecognitions on its own — that is a fact about the ledger, not the model — so it takes them as an input and says so rather than reporting nil as though it were measured.See it |
Elsewhere
The same shape of problem, at different stages, in the other markets this is built for.
- India — NBFCsIn force
- Already in force and has been for years. NBFCs are not waiting for 2027 — they are doing this now, usually in a spreadsheet somebody inherited.
- PakistanIn force 1 January 2024
- Live since January 2024. The same shape of problem, two years further along.
Try it on your own book, without asking anyone.
The file is read in this browser and never uploaded, so there is nothing for procurement to review and no data-sharing agreement to sign before you can find out whether it works on your data. That is the part no hosted competitor can offer you.