A specification of what this dataset is bounded by. Every limit below carries a number or a date, because a limit without one is not usable.
What does TatvaRatings not cover?
Four limits bound this dataset. It covers rated borrowers only — most Indian borrowers have never been rated. Only 46.2% of rating actions carry a lender-wise annexure, so every figure is a floor, not a total. Amounts are sanctioned positions, not outstanding balances. And each record is a dated snapshot as at its rating action date, not a feed.
Measured as at 30 August 2026. Read every one of these as a floor, not a total.
All figures as at 30 August 2026.
The two that matter most
The figures are floors, not totals. Every count, aggregate and rupee figure is the amount evidenced by published documents. Read each one as “at least this much, from public record” — never as “this much.”
An annexure is a dated snapshot, not a feed. It reflects the sanctioned position as at the rating action date. Not today. Not an outstanding balance.
Both are specified below.
Limit 1 — Coverage bias. The figures are floors, not totals.
When TatvaRatings reports that SBI lends to 5,476 borrowers, the literally true statement is:
State Bank of India is named as a lender in 5,476 rating annexures that TatvaRatings has ingested.
Two filters sit between that number and the real position.
Filter 1 — 46.2% of rating actions publish an annexure. Disclosure of lender details in a rating press release depends on the issuer’s consent. Where consent is absent, the release publishes without an annexure and there is no lender row to hold. The gap is in the source document, not in the parse.
Measured as at 30 August 2026: 32,498 of 70,337 rating actions carry at least one lender row — 46.2% — holding 179,857 lender-exposure rows in total. The share is not evenly distributed, and the spread matters more than the average:
| Agency | Rating actions | With an annexure | Share |
|---|---|---|---|
| Acuité | 4,856 | 4,467 | 92.0% |
| Infomerics | 8,808 | 5,863 | 66.6% |
| Brickwork | 76 | 45 | 59.2% |
| CARE | 12,403 | 6,501 | 52.4% |
| CRISIL | 22,046 | 10,136 | 46.0% |
| ICRA | 12,658 | 5,486 | 43.3% |
| India Ratings | 9,490 | 0 | 0.0% |
| All seven | 70,337 | 32,498 | 46.2% |
India Ratings’ 0.0% is genuine non-publication, not a gap in our parsing: their annexure is instrument-wise and has no lender column, verified across all 9,490 of their rating actions in this corpus — each one backed by its own source document. If your question depends on India Ratings coverage, this corpus cannot answer it.
Filter 2 — the rated book is a minority of any lender’s borrowers. A public credit rating is a specific event that most lending relationships never involve. Even a perfectly complete annexure corpus would describe only the rated slice of a bank’s book, and that slice is small.
The consequence. Every count, every aggregate and every rupee figure is a floor — the amount evidenced by published documents. The true position is larger. It is larger by a margin that varies by bank, by sector and by period, and that margin is not estimated here. An unquantified estimate would be less useful than a known floor, and a known floor is something you can actually work with.
Limit 2 — Staleness. An annexure is a snapshot, not a feed.
A lender-wise annexure reflects the sanctioned position as at the rating action date. Not today. Not an outstanding balance.
A cooperating issuer’s rating is reviewed periodically rather than continuously, broadly once a year. SEBI does not prescribe a fixed interval: under its Master Circular for Credit Rating Agencies the review due date is the one set in each agency’s own operations manual, and an agency must publicly disclose any review that becomes due and is not completed on time. From that:
- A typical record is a year or more old at the time you read it.
- Where an issuer has stopped cooperating and the rating has migrated to the issuer-non-cooperating category, the underlying information is older still, and how much older is not always evident from the release.
- A facility repaid, refinanced, taken over or cancelled the day after the rating action still reads as sanctioned until the next review publishes.
- A new facility sanctioned between reviews does not appear at all until the next action.
The consequence. The corpus is refreshed continuously; the facts inside it are not. Ingest cadence is not data freshness, and no amount of the first compresses the second. This is a property of a periodic disclosure regime, not an engineering gap.
Universe boundaries
What is outside the dataset by construction.
| Boundary | Specification |
|---|---|
| Unrated borrowers | Excluded entirely. The universe is borrowers with a published rating action from one of these seven SEBI-registered agencies — 45,528 of them. |
| Unrated facilities of a rated borrower | Excluded. Only facilities covered by a rating action appear. A rated borrower’s unrated lines are invisible here. |
| Non-bank lending relationships | Present only where a rating annexure names the lender. Trade credit, intra-group funding and unrated private debt are generally absent. |
| Outstanding balances | Not held. The annexure gives sanctioned amounts. Where a release states average limit utilisation, that is captured as published and is not the same thing as an outstanding figure. |
| Non-public sources | None used. No supervisory returns, no credit bureau data, no confidential lender submissions. |
| Pre-May 2019 actions | The earliest action currently held is from May 2019. Agency archives run longer; backfill is incremental. |
Field-level cautions
Proposed and unallocated limits. Rating releases routinely include proposed or unallocated facilities, sometimes without a named lender. These are flagged distinctly in the data. A lender total that includes them overstates the position, and the flag exists so that it does not have to.
Amounts are as published. Where an agency restated or reclassified a figure, the figure is captured as the agency printed it, with the basis and period stated. TatvaRatings does not reconcile agency figures to audited statements.
Notation mapping is a mapping, not an equivalence. Ratings from different agencies are mapped to a common scale for comparability, with each agency’s native symbol retained. A mapped-equivalent rating from a different agency is not an interchangeable opinion.
Entity resolution carries residual error. Borrower and lender identities are reconciled across variant spellings, initialisms, branch suffixes and legacy names of merged banks — 1,501 distinct lenders as at 30 August 2026. Resolution is deliberately tuned to prefer splitting a single entity over merging two distinct ones. The effect of that choice is a mild undercount, which is consistent with floors-not-totals; the alternative would produce an overcount that a reader cannot detect. Method →
Extraction failures exist and are tracked. Some source documents are scanned or malformed and defeat extraction. Failures are recorded with a reason and replayed against improved parsers. Crucially, “no annexure was published” and “the annexure could not be extracted” are stored as different states and are never merged.
Rights and status
No regulatory status. TatvaRatings holds no registration, licence, recognition or approval from the RBI or SEBI, and claims none. It is not a credit rating agency, not a credit information company, and not a reporting institution under any supervisory framework.
No exclusive access. The source documents are public disclosures published by the agencies themselves. What is added here is parsing, entity resolution and the lender-side index — not privileged access to anything.
No agency relationship. There is no partnership, endorsement, authorisation or affiliation with any credit rating agency, and none is implied. Agency narrative prose is not reproduced at length. Sources and position →
What this data is not for
It is not exposure monitoring. There is no continuous observation of any lender’s book, no alerting layer, and no mechanism by which a change is detected before the next rating action publishes. The periodic review cycle rules that out.
It is not a system of record. It should not be relied on as a control in any process requiring a current or complete exposure position, nor for regulatory reporting, capital computation, or any determination where completeness is a requirement.
It is not a credit opinion. No score, no rank, no rating of our own. The ratings held are the agencies’ opinions, attributed to the agency that issued them.
It does not replace verification on a named borrower. Use it to see the shape of the market, find counterparties, size a sector, and generate questions. Then confirm a specific name through your own channels before you act on it.
Why this page exists
Because a credit team will find these limits in the first hour of the first evaluation, and the only variable is whether they find them here or find them themselves. A dataset whose denominator is stated is one you can reason about. One whose denominator is implied is not.
If a question you need answered runs into one of the limits above, we would rather establish that in the first conversation than in the third.
Related
- Coverage — what is in the corpus
- Methodology — how it gets there
- Sources — where it comes from