ESG rating divergence

Almost all of it is self-reported. A company measures its own emissions, chooses its own boundary, picks its own base year, and publishes the result in a document that is assured to a lower standard than its accounts, if it is assured at all. India's Business Responsibility and Sustainability Report is mandatory for the largest listed companies, and a subset of its indicators carries an assurance or assessment requirement on a glide path that has been revised more than once (check the current SEBI circular).

Europe's regime under the Corporate Sustainability Reporting Directive and the associated standards is broader, and it has been rescoped since it was first adopted (check the current scope and timetable).

Boundary choice alone can move a headline number by more than any real-world change. Kalinga Cement Works (synthetic) emits 6,048,000 tonnes of carbon dioxide across scopes one and two. Add a scope-three estimate at 38% of that and the figure becomes 8,346,240 tonnes and the intensity rises from 0.720000 to 0.993600 tonnes per tonne of cement. Neither number is wrong. They answer different questions.

A rating that blends companies reporting on different boundaries has compared two of those different questions and called the difference performance.

How do you measure the gap between two ESG ratings?

Rating divergence is therefore structural rather than accidental, and it can be decomposed exactly. Two providers scoring the same issuer differ in the weights they attach to each pillar and in the scores they assign within a pillar, and the gap between their totals splits into three terms that sum to the gap with no residual:

gap = SUM_i (wQ_i - wP_i) * sP_i          <- weight effect
      + SUM_i wP_i * (sQ_i - sP_i)         <- score effect
      + SUM_i (wQ_i - wP_i) * (sQ_i - sP_i) <- interaction

That is the same allocation-and-selection algebra the performance-attribution work uses, applied to a rating instead of to a portfolio, and the identity check is the same: the three terms must reproduce the gap to the last decimal. Worked Example 4 runs it on a US utility where the two providers are 1.9500 points apart on a ten-point scale. Of that gap, 1.5000 points is disagreement about the pillar scores, 1.2500 of it in the environmental pillar alone, and 0.9500 points comes from the weights, with an interaction of -0.5000 pulling back.

The number that matters to a portfolio, though, is not the score gap. It is what the gap does to a rank. On a stated universe distribution with mean 5.60 and standard deviation 1.40, a score of 5.10 sits at the 36.05th percentile and a score of 7.05 sits at the 84.98th. The same company is a bottom-third name to one provider and a top-sixth name to the other, 48.93 percentile points apart, which is the difference between exclusion and overweight in most systematic mandates.

How can you test for greenwashing?

Greenwashing tests follow from the same logic and are cheap to run. Compare the narrative claim against the disclosed quantity in the same document. Compare this year's base year against last year's. Check whether an intensity improvement came from the numerator or from a denominator that grew. Ask whether the assured subset happens to contain exactly the indicators that improved. None of that requires a data vendor.

{
  "skill": "Filling a governance checklist from primary filings without transcription error",
  "use": "Have a document assistant pull the specific governance facts out of an annual report, corporate governance report and proxy into a fixed schema (board size, independent count, chair status, committee membership, auditor and tenure, related-party totals, pledge percentage, share classes), and have it return the page number for every single field so each one can be checked against the source.",
  "tools": ["Document or PDF assistants grounded ONLY on the filing you already hold", "A chat assistant with code execution, for the scoring arithmetic", "The regulator's own text: sebi.gov.in for the listing regulations, mca.gov.in for the Companies Act, sec.gov for the proxy and the exchange listing standards"],
  "prompt": "From the attached annual report and corporate governance report, extract into a table: total directors; names and count of independent directors; whether the chairperson is executive and whether related to the promoter; audit committee members with independence flags and stated financial expertise; nomination and remuneration committee members with independence flags; statutory auditor name, first year of appointment and current term; total related-party transaction value and the revenue figure it should be compared with; promoter shareholding and the pledged percentage; every class of share with its voting rights. Give the page number for every field. Where the document does not state a field, write NOT DISCLOSED rather than inferring it.",
  "verify": "Open the filing and check every extracted field against the page the tool cited, because dense governance tables are exactly where document models transpose rows and an independent-director count that is wrong by one changes the compliance conclusion. Recompute the independent share, the related-party percentage and the checklist total yourself. Match the related-party figure to the revenue basis it belongs with, standalone against standalone or consolidated against consolidated, since mixing them is the commonest error in this exercise. Date-check the filing against the period you are valuing, and check whether a later disclosure has superseded it. For any question about what the rule REQUIRES rather than what the company DISCLOSED, go to the regulator's own current text, never to the model's recollection of a regulation, which is the single most confidently wrong output in this whole area.",
  "diy": "You must be able to read a corporate governance report and a proxy unaided and say, in under fifteen minutes, how many independent directors there are, whether the audit committee is compliant, how long the auditor has been in the chair and what fraction of revenue moves through related parties. The gate rewards that because a tool cannot tell you that a compliant committee has nobody numerate on it, and the scoring judgement is the part that carries the valuation.",
  "market": "IN"
}