Measuring loan book emissions: where the quality breaks and what to fix first

Most banks can produce a loan book emissions figure in an afternoon. Producing one that a relationship manager, a risk committee or an auditor can use takes longer, and the difficulty is concentrated in one place: the business customers that have never published an emissions figure, which is most of them.

Segment the book by asset class before anything else

The PCAF standard measures each asset class differently, so the first job is to split the book. Business loans and unlisted equity attribute a counterparty's emissions by outstanding amount over total equity plus debt. Listed equity and corporate bonds attribute by outstanding amount over enterprise value including cash. Mortgages, motor loans, project finance and sovereign debt each have their own method and their own data sources, and none of them involves a counterparty inventory.

This post is about the first two groups, the ones where the counterparty is a company. That is where the tonnage usually sits in a commercial bank, and it is where data quality is decided by whether the company itself has disclosed.

Where the number breaks

At the point where the counterparty stops being listed. Listed borrowers and issuers nearly all publish an inventory, so those lines arrive at PCAF data quality score 2, or score 1 if the inventory was assured. Below that line the standard fallback is the economic activity method: the customer's revenue multiplied by an emissions intensity for its sector and country, score 4, or score 5 where the revenue is estimated too. The PCAF data quality scores guide sets out the scale.

The problem with a score 4 line is not that it is wrong. It is that it is the same for every customer in the sector. Two haulage firms with the same turnover carry the same emissions whatever their fleets look like, and neither figure moves when one of them electrifies. A loan book built that way shows progress only when exposure shifts or when a sector average is revised, which is a description of the market rather than of the customers. It also cannot support a conversation with any of them, because the number was never theirs.

Which counterparties to improve first

The ones that carry the tonnage and are still on a sector factor. Sort the modelled lines by attributed emissions and the shape is familiar: a small number of larger mid-market customers in carbon-intensive sectors account for most of the estimated total, and a long tail of small exposures accounts for very little. The tail can legitimately stay on the economic activity method. The head cannot, because that is where a sector average is most likely to be wrong by the widest margin, in either direction.

For the head, the question is not how to model better. It is whether the customer has already published. Many mid-market and large private companies have: in an annual report, a sustainability statement, a national filing or a response to a large customer's Scope 3 programme. That figure is primary data under the GHG Protocol and score 2 or 1 under PCAF, and it exists whether or not the bank has found it. Finding it moves the line up two or three data quality levels with no request to the customer at all.

What each improvement costs

Stated plainly, because the rungs do not cost the same. Moving a line from score 5 to score 4 costs the bank a revenue figure and the customer nothing. Moving it from score 4 to score 2 or 1 by finding a published inventory costs the customer nothing and the bank the work of finding, matching and normalising the disclosure, which is where a data layer earns its place. Moving a line to score 3 with activity data costs real work on both sides, and belongs on the handful of counterparties where it will change a decision. The cheapest large improvement in a loan book is the second one, and it is the one most banks have not made.

DitchCarbon does that finding at scale. It holds verified emissions data for over 2 million organisations, built on primary emissions data wherever it exists, private companies included, matched to the bank's counterparties through entity resolution against DUNS, LEI and ISIN identifiers. Each record states which method produced its figure, with industry data fourth of four, and every figure carries its source and change history, so the line an analyst moves from score 4 to score 2 links to the report the number came from.

How this fits the PCAF model the bank already runs

Additively. The model, the factor set and the published baseline stay where they are. The counterparty records slot into the fallback ladder the model already uses: a disclosed figure first, a figure modelled from the company's own disclosures and actual revenue where there is none, a regionalised industry factor last and shown as such. Lines move up the hierarchy one at a time as disclosures are found, and the bank decides when to recalculate. The baseline does not jump, and the source for every changed line is on the record when the auditor samples it.

Where a counterparty in the head of the book has not published, a request goes out from the portfolio view prepopulated with whatever it has disclosed elsewhere, so its finance team confirms and corrects rather than starts from a blank form. The answer lands on a profile the company owns and reuses with the next bank that asks, which is the reason a business customer will do it. Where nothing comes back, the line stays on a regionalised industry factor and the record says so.

A loan book measured this way carries a figure for each counterparty that is that counterparty's own wherever one exists, a stated method for every line that is not, and a document behind each figure. That is a book the bank can lend against, set targets on and defend, and it is built from disclosures that were already there.