Chapter 3: Credit Risk Measurement as the Foundation of Green Identification

The Basel Committee's 2020 survey of climate-related financial risk initiatives lists 27 separate disclosure frameworks, stress-test exercises, and supervisory guidance documents in operation across the G20. Not one specifies how a bank should adjust a probability of default for a borrower's emissions trajectory. That gap is the subject of this chapter.

Identification is the first act of lending. Before a bank can price, monitor, or mitigate a credit, it must decide what the credit is. For most of the past three decades, that decision has rested on three numbers: probability of default (PD), loss given default (LGD), and exposure at default (EAD). The three multiply into an expected loss that anchors pricing, provisioning, and capital allocation. I will call this triad the measurement core. It is mature, standardized, and embedded in every loan origination system in the world. Standard PD models do not include environmental variables.

The green transition does not require throwing out the measurement core. It requires extending it. A firm's environmental profile changes its default probability through transition risk (carbon pricing, regulatory shocks, stranded assets) and through physical risk (flood, drought, heat). Both channels are measurable. Neither is measured in standard practice. The result is a systematic mispricing of green and transition-worthy firms. Call it a measurement failure rather than a market failure in the textbook sense. Banks cannot identify what they cannot score. This chapter builds the scoring.

3.1 Traditional Credit Risk Metrics: The Measurement Core

3.1.1 Why One Number Is Not Enough

Arnaud de Servigny and Olivier Renault, in Credit Risk Measurement and Management (机械工业出版社, 2012), open their treatment of credit risk with a simple concession: no single figure tells you whether a transaction is good or bad. Transactions must be analyzed across several dimensions, and those dimensions taken together form a barometer. The four they name are exposure, probability of default, recovery rate, and tenor. The first three multiply into expected loss; the fourth governs how long the exposure is at risk.

This is the right place to start because it disciplines the green-finance conversation. Much of the policy literature treats "green credit" as a category of lending, a label attached to a loan. The measurement tradition treats it as a set of parameters. A loan's greenness depends on how the borrower's environmental performance changes one or more of the parameters in a direction that reduces expected loss, or on how it changes them in a direction the lender has agreed to accept at a lower spread for policy reasons. The distinction matters. Labels can be gamed. Parameters cannot.

The measurement core rests on a two-step derivation of default probability that de Servigny and Renault describe plainly: first assign a rating, then use historical data to map that rating to a default frequency. The rating is a judgment; the mapping is an empirical regularity. The whole edifice of internal ratings-based (IRB) capital under Basel II and III is built on this two-step logic. A bank rates a borrower into a grade, then applies the grade's long-run default rate.

The logic has a hidden assumption. It assumes that the historical default frequency attached to a grade remains valid for the borrower in front of you. For a steel manufacturer in 2005, the assumption held well enough. For the same manufacturer in 2025, facing a carbon price of €80 per tonne and an EU border adjustment mechanism, the historical frequency is a poor guide. The grade may still be BB, but the BB default rate of the past decade was computed on a population that did not face the transition. The mapping has drifted. Nobody has re-estimated it.

That drift is the technical heart of why green identification fails even when banks want to succeed at it. The grade looks stable. The population behind the grade has changed. A model can pass every validation test and still misprice a carbon-intensive borrower, because the validation tests compare the model to a history that no longer describes the world the borrower operates in.

3.1.2 Probability of Default

PD is the likelihood that a counterparty will fail to meet its financial obligations within a defined horizon, usually one year. It is a forward-looking statistical estimate, not a fact. Two firms with identical balance sheets can carry different PDs because their industries, management quality, and competitive positions differ.

The standard derivation runs through a rating scale. S&P's long-term scale runs from AAA to D, with modifiers; Moody's runs from Aaa to C. Each grade carries an observed default rate. Table 4.1 and Table 4.2 in de Servigny and Renault present these scales and their historical frequencies. The frequencies are remarkably stable at the investment-grade end and remarkably unstable at the speculative end. A single-B borrower's one-year default rate has ranged from roughly 2% to 8% depending on the vintage.

For green identification, PD is the most important of the three parameters because it is the one most directly affected by environmental performance. A firm's emissions intensity, its exposure to carbon pricing, its water dependence, and its physical-risk location all bear on its probability of default. None of them appear in a standard PD model. The standard model uses leverage, coverage, liquidity, profitability, size, and sector. Sector is a crude proxy for environmental exposure, and a poor one: within the chemicals sector, a producer of battery electrolytes and a producer of PVC pipe face radically different transition trajectories.

The gap is not theoretical. The European Banking Authority's 2021 report on management and supervision of ESG risks for credit institutions found that fewer than one in five EU banks had incorporated environmental factors into their PD models in any systematic way. The remainder used them as a qualitative overlay, a committee discussion item, or not at all. The EBA's 2021 EU-wide pilot exercise on climate risk reached a similar conclusion: banks could estimate exposure to physical risk at the portfolio level but could not translate it into default probabilities at the obligor level.

3.1.3 Loss Given Default

LGD is the share of exposure that is not recovered when a default occurs. It depends on collateral, seniority, and the depth of the secondary market for the assets. A secured loan against a liquid asset carries an LGD of perhaps 20%; an unsecured loan to a firm with few tangible assets carries 60% or more.

Environmental performance enters LGD through collateral values and through recovery timing. A loan secured against a coal-fired power plant has a collateral value that is falling, and will fall faster under a carbon price. A loan secured against a forest has a collateral value that depends on fire risk, pest risk, and the regulatory treatment of the land. In both cases, the LGD is not a constant; it is a function of environmental variables that the standard model holds fixed.

The point is sharper for transition-worthy firms. A manufacturer retooling a plant to produce heat pumps has collateral (the plant) whose value rises with the transition. A standard LGD model, calibrated on historical recoveries from a population that was not transitioning, will understate the recovery value of the retooled asset. The green firm is penalized twice: once on PD (because the model does not credit its lower transition risk) and once on LGD (because the model does not credit its rising collateral value).

3.1.4 Exposure at Default

EAD is the amount of money at risk at the moment of default. For a term loan, it is roughly the outstanding balance. For a revolving credit line, it is the drawn amount plus an estimate of the undrawn portion that the borrower will draw down before defaulting. For derivatives, it is a mark-to-market plus a potential future exposure.

Environmental performance affects EAD less directly than PD or LGD, but the effect is real for project finance and for trade finance. A project-finance loan to a solar developer has an EAD that depends on construction completion; a delay caused by permitting disputes over environmental impact assessments extends the period of maximum exposure. A trade-finance facility to an agricultural exporter has an EAD that depends on harvest timing; a drought shifts the drawdown pattern.

These are second-order effects. The first-order work is on PD and LGD. I will concentrate there.

3.1.5 The Expected Loss Identity

The three parameters combine into expected loss:

This identity is the anchor of credit risk management. It is also the point at which green identification must be inserted. If environmental performance changes PD, or LGD, or EAD, then it changes EL. If it changes EL, it changes the price at which the loan is profitable. A bank that ignores the change is either overpricing green firms (and losing them to competitors) or underpricing brown firms (and accumulating losses that will emerge when the transition accelerates).

The identity also clarifies what "green identification" means in measurement terms. It does not mean a separate green lending book. It means a PD model that credits lower transition risk, an LGD model that credits rising collateral value, and an EAD model that accounts for environmental timing. The green book emerges from the measurement; it cannot be assembled by relabeling existing loans.

3.1.6 The Tenor Dimension

De Servigny and Renault add tenor as a fourth dimension, and it deserves separate treatment because it is where the green transition bites hardest. Most corporate lending has a tenor of one to five years. Most transition investments have a payback period of seven to fifteen years. The mismatch is structural.

A five-year loan to a firm making a fifteen-year transition investment is a loan whose PD must be estimated over a horizon shorter than the investment's payoff. If the transition succeeds, the firm's credit quality improves after year five; if it fails, the firm defaults before the loan matures. The lender captures neither the upside nor the full downside. This is a classic maturity-mismatch problem, and it is why green lending needs instruments with longer tenors or with equity-like features. I will return to this in Chapter 12, on exit mechanisms.

For measurement, tenor matters because it determines which environmental variables are relevant. A one-year loan to a chemical producer is exposed to carbon price changes within the year; a ten-year loan is exposed to the entire regulatory trajectory. The same firm can carry a low PD on a short loan and a high PD on a long one. A measurement system that reports a single PD per obligor is reporting an average that misleads at both ends.

3.2 The Structural Gap: Why Standard Models Miss Environmental Risk

3.2.1 The Historical Calibration Problem

Every PD model is calibrated on history. The history of corporate defaults from 1980 to 2020 contains few events that were caused by environmental factors. The Superfund legislation of 1980 in the United States created liability for pollution and prompted banks to guard against credit risk from environmental contamination, and it is often cited as the origin of green finance. But the defaults it caused were idiosyncratic, not systemic. The modelers who built the modern PD models in the 1990s and 2000s had no reason to include environmental variables because the historical default data did not require them.

This is the calibration trap. The absence of environmental variables in the historical record is not evidence that they are unimportant; it is evidence that the record predates the transition. A model trained on a period when carbon was unpriced will underestimate the default risk of a carbon-intensive firm in a period when carbon is priced. The underestimation is not a bias that can be corrected by adding a dummy variable. It is a structural misspecification that requires rethinking the model's inputs.

The European Central Bank's 2020 guide on climate-related and environmental risks made this point explicitly: banks' risk management frameworks "do not yet adequately reflect" climate risk because the historical data on which they are built does not contain the relevant events. The ECB's 2021 risk assessment found that none of the significant institutions it reviewed had a "fully-fledged" climate risk methodology. The Bank for International Settlements' 2020 Green Swan report reached the same conclusion from a different angle: climate risk is a "green swan" precisely because it falls outside the historical distribution on which risk models are calibrated.

3.2.2 The Sector Proxy Problem

The standard workaround is to use sector as a proxy for environmental exposure. A coal-fired utility gets a worse PD than a solar developer; a cement producer gets a worse PD than a software firm. This is better than nothing, and it is what most banks do. It is also crude in ways that matter.

Within the cement sector, a producer that has invested in carbon capture and a producer that has not face different transition trajectories. Within the software sector, a firm that provides grid-management software and a firm that provides advertising technology face different transition opportunities. The sector proxy assigns them the same environmental adjustment, which is to say none.

The sector proxy also fails at the boundary between green and brown. A chemicals firm that produces battery electrolytes is classified as chemicals; its transition trajectory is closer to a battery manufacturer than to a PVC producer. A utility that generates 40% of its power from renewables is classified as a utility; its trajectory depends on the remaining 60%. Sector codes were designed for industrial classification, not for transition assessment. Using them for the latter is a category error.

3.2.3 The Data Problem

Even if a bank wanted to incorporate firm-level environmental data into its PD model, the data is uneven. Scope 1 and Scope 2 emissions are reported by large listed firms under the EU's Non-Financial Reporting Directive and its successor, the Corporate Sustainability Reporting Directive (CSRD). Scope 3 emissions, which for many sectors dominate the total, are estimated with wide error bands. Small and medium enterprises, which make up the bulk of bank lending in most economies, report little or nothing.

The data problem is not an excuse for inaction. It is a design constraint. A measurement system that requires complete Scope 3 data will exclude most of the lending book. A system that requires only what is reported will be biased toward large listed firms. The workable approach is tiered: use reported data where it exists, use sector-and-size benchmarks where it does not, and flag the difference in model confidence. I will develop this approach in Section 3.3.

3.2.4 The Regulatory Patchwork

The regulatory environment for green credit measurement is fragmented. The EU's Capital Requirements Regulation, as amended by the 2021 proposal on credit risk, requires banks to disclose ESG risks but does not yet specify how to incorporate them into capital calculations. The EBA's 2020 guidelines on loan origination and monitoring require banks to consider environmental factors in credit assessment but leave the methodology to the bank. The ECB's guide sets supervisory expectations but not prescriptive rules.

This patchwork has a predictable effect. Large banks with the resources to build proprietary models do so; smaller banks wait for standards. The result is a measurement gap that widens with bank size, and a green lending market that concentrates in the institutions best able to identify green firms. That concentration is not necessarily bad—specialization has its advantages—but it means the transition is financed unevenly across the banking system.

The Basel Committee's climate-risk work has moved slowly by design. The 2020 survey was a stocktaking, not a standard-setting exercise. The Committee's 2021 principles for the effective management and supervision of climate-related financial risks are principles, not rules. A binding methodology for incorporating environmental factors into PD and LGD remains years away. Banks that wait for it will be late.

3.3 Integrating Environmental Factors into Credit Risk Models

3.3.1 Two Channels: Transition and Physical

Environmental performance affects credit risk through two channels, and they require different measurement approaches.

Transition risk is the risk that a firm's business model loses value as the economy decarbonizes. It arrives through carbon pricing, regulation, technology substitution, and changing consumer preferences. It is concentrated in carbon-intensive sectors and in firms with long-lived assets. It is measurable through emissions intensity, asset age, and regulatory exposure.

Physical risk is the risk that environmental change damages a firm's assets or disrupts its operations. It arrives through flood, drought, heat, storm, and sea-level rise. It is concentrated in agriculture, real estate, infrastructure, and utilities. It is measurable through geolocation, hazard maps, and insurance data.

The two channels interact. A coal-fired power plant in a flood zone faces both transition risk (falling demand for coal power) and physical risk (flood damage). A farm in a drought-prone region faces physical risk (crop failure) and transition risk (water pricing). A measurement system that handles only one channel will misclassify firms that face both.

flowchart TD
    A[Firm environmental profile] --> B[Transition risk channel]
    A --> C[Physical risk channel]
    B --> D[Emissions intensity]
    B --> E[Asset age and lock-in]
    B --> F[Regulatory exposure]
    C --> G[Geolocation]
    C --> H[Hazard exposure]
    C --> I[Insurance coverage]
    D --> J[PD adjustment]
    E --> J
    F --> J
    G --> K[LGD adjustment]
    H --> K
    I --> K
    J --> L[Expected loss]
    K --> L
    L --> M[Pricing and capital]

The diagram is a simplification, but it captures the essential structure. Environmental performance enters the measurement core through two doors, and both doors lead to the same room.

3.3.2 The Transition-Risk Adjustment to PD

Transition risk enters the PD model most naturally as a shift in the firm's cash-flow distribution. A firm facing a carbon price has a lower expected cash flow and a higher variance of cash flow than the same firm without the carbon price. Lower expected cash flow raises PD; higher variance raises it further. The adjustment can be modeled as a multiplicative factor on the baseline PD:

where \(T\) is a transition-risk score scaled to the unit interval and \(\alpha\) is a calibration parameter. The score \(T\) is constructed from emissions intensity relative to the sector benchmark, asset age, and regulatory exposure. The parameter \(\alpha\) is estimated from historical episodes in which a transition shock materialized—the EU Emissions Trading System's price spikes, the collapse of coal demand in Western Europe, the phase-out of diesel in European cities.

The multiplicative form has the advantage of preserving the ranking of the baseline model while shifting it in the direction of transition risk. A firm with a baseline PD of 1% and a transition score of 0.8 might see its PD rise to 1.6% if \(\alpha = 0.75\). A firm with a transition score of 0.2 might see its PD rise to 1.15%. The green firm is not rewarded with a lower PD; it is spared the penalty that the brown firm receives. This asymmetry is deliberate. The baseline model already captures the green firm's lower leverage and higher profitability; the transition adjustment should not double-count those advantages.

An alternative is to construct a separate transition-risk rating and combine it with the financial rating through a scorecard. This is the approach taken by several European banks that have built proprietary green rating models. It has the advantage of transparency: the transition score is visible and can be challenged. It has the disadvantage of requiring a second rating infrastructure, which smaller banks cannot afford.

3.3.3 The Physical-Risk Adjustment to LGD

Physical risk enters LGD through collateral value. A loan secured against a property in a flood zone has an LGD that depends on the flood probability and the expected damage. A loan secured against a forest has an LGD that depends on fire and pest risk. The adjustment can be modeled as a reduction in the recovery value:

where \(P\) is the physical-risk score and \(\delta\) is the expected fractional loss of collateral value conditional on the hazard. The form is deliberately conservative: it reduces recovery only to the extent that the hazard is expected to materialize within the loan's tenor.

The physical-risk score \(P\) is constructed from geolocation and hazard maps. The EU's Copernicus climate data, the US Federal Emergency Management Agency's flood maps, and the insurance industry's catastrophe models are all inputs. The score is not a forecast of a specific event; it is an estimate of the probability that a hazard of a given intensity will occur within the tenor.

The insurance data is particularly valuable because it is priced. A property that is cheap to insure is a property with low physical risk; a property that is expensive to insure, or uninsurable, is a property with high physical risk. The insurance market has already done the measurement work. Banks that use insurance premiums as an input to LGD are free-riding on the underwriters' models, which is a legitimate and efficient use of available information.

3.3.4 The Data Tiering Approach

The data problem is real, and the solution is tiering. I propose three tiers of environmental data, each with its own confidence level and its own treatment in the model.

Tier 1 covers reported data. Large listed firms report Scope 1 and Scope 2 emissions under CSRD, and many report Scope 3. The data is audited, at least in principle. For these firms, the transition score can be constructed from reported emissions intensity relative to sector benchmark, with adjustments for asset age and regulatory exposure. The confidence level is high.

Tier 2 covers estimated data. Mid-sized firms that do not report can be assigned emissions estimates from sector-and-size benchmarks. The estimates are not firm-specific, but they are better than nothing. The confidence level is medium, and the model should widen the PD confidence interval accordingly.

Tier 3 covers proxy data. Small firms with no environmental data can be assigned a sector-level transition score. The score is crude, but it captures the first-order exposure. The confidence level is low, and the model should apply a conservative adjustment (i.e., assume the firm is at the sector average, not below it).

The tiering approach has a political economy advantage. It does not penalize small firms for not reporting; it assigns them a benchmark and moves on. It also creates an incentive to report: a firm that moves from Tier 3 to Tier 1 gets a more accurate score, which, if its environmental performance is good, lowers its cost of capital.

3.3.5 The Scorecard Structure

The integration of environmental factors into the measurement core is best done through a scorecard that sits alongside the financial rating. The scorecard has four components:

The weights are illustrative. A bank lending to agriculture would raise the physical-risk weight; a bank lending to heavy industry would raise the emissions-intensity weight. The scorecard produces a transition-risk score \(T\) and a physical-risk score \(P\), which feed the PD and LGD adjustments described above.

The scorecard is not a substitute for the financial rating. It is a complement. A firm with strong financials and a poor transition score is still a better credit than a firm with weak financials and a good transition score. The scorecard adjusts the financial rating; it does not replace it.

3.3.6 Validation and Backtesting

A model that cannot be validated is a model that cannot be trusted. The standard validation tools—backtesting, benchmarking, discrimination tests—apply to the environmental adjustments as much as to the baseline model.

Backtesting is the hard part. The historical default data does not contain enough transition-driven defaults to validate the adjustment statistically. The available episodes are few: the collapse of German coal utilities after the nuclear phase-out, the distress of European automotive suppliers after the diesel scandal, the failure of several solar manufacturers after the 2011 price collapse. Each episode is instructive, but none provides a large enough sample for statistical validation.

The workaround is to use stress testing as a validation tool. A bank can ask: if the carbon price rises to €150 per tonne by 2030, how many of our borrowers default? The answer is a scenario, not a forecast, but it tests whether the adjustment behaves sensibly under extreme conditions. The ECB's 2022 climate stress test did exactly this for the largest euro-area banks, and it found that transition risk concentrated in a small number of sectors could produce credit losses comparable to a severe recession.

Benchmarking is easier. A bank can compare its environmental adjustments to those of peers, to the adjustments implied by credit rating agencies, and to the adjustments implied by market prices (credit default swap spreads, bond spreads). Divergence is not necessarily error, but it requires explanation. A bank whose green adjustments are far more generous than the market's is either ahead of the market or wrong. The burden of proof is on the bank.

3.3.7 The Rating Agency Experience

The credit rating agencies have moved faster than the banks, though not fast enough. Moody's, S&P, and Fitch all publish ESG assessments alongside their credit ratings. The assessments are separate from the ratings, which is a deliberate choice: the agencies argue that ESG factors are already incorporated into the ratings where they are material, and that a separate score would double-count.

The argument is defensible in principle and questionable in practice. S&P's 2021 review of its ESG integration found that environmental factors were material to the rating in about 10% of corporate cases, concentrated in utilities, oil and gas, and mining. For the other 90%, the factors were deemed immaterial. Whether that judgment is correct depends on the time horizon. Over a one-year horizon, a software firm's emissions are immaterial to its default probability. Over a ten-year horizon, they may not be, if the firm's data centers face carbon pricing or if its customers demand decarbonization.

The agencies' approach has the virtue of discipline: they incorporate environmental factors only where they affect cash flow within the rating horizon. It has the vice of short-termism: the rating horizon is typically one to three years, which is shorter than the transition. A firm with a ten-year transition problem can carry an investment-grade rating for nine of those years and then default. The rating was correct at each point and wrong overall.

Moody's Analytics' EDF (Expected Default Frequency) metric, which de Servigny and Renault discuss in their chapter on alternative estimations of credit quality, is built on Merton's structural model. The model treats equity as a call option on the firm's assets; default occurs when asset value falls below the debt threshold. Environmental factors enter only through their effect on asset volatility and asset value. A carbon price that reduces expected cash flow reduces asset value and raises EDF. The model can accommodate environmental factors if they are translated into cash-flow adjustments, but the translation is not automatic.

3.3.8 The Merton Framework as a Bridge

The Merton framework deserves more attention than it usually gets in the green-finance literature, because it provides a natural bridge between environmental performance and default probability.

In the Merton model, the firm's asset value \(V\) follows a stochastic process, and default occurs if \(V\) falls below the debt threshold \(D\) at maturity. The distance to default is:

where \(\mu\) is the expected return on assets, \(\sigma\) is asset volatility, and \(T\) is the horizon. The probability of default is the probability that a standard normal variable falls below \(-DD\).

Environmental performance enters the model through three channels. First, it affects \(V\): a firm facing a carbon price has lower expected cash flow and therefore lower asset value. Second, it affects \(\sigma\): a firm facing transition risk has higher cash-flow volatility. Third, it affects the correlation between \(V\) and the market: a firm whose fortunes are tied to the carbon price has a different beta than a firm whose fortunes are not.

The Merton framework is attractive because it is transparent and because it uses market data. For listed firms, equity prices and equity volatility are observable, and the model can be calibrated without historical default data. The framework's weakness is that it assumes efficient markets and continuous trading, which are approximations at best. For unlisted firms, the framework requires accounting-based proxies, which reintroduces the data problem.

The bridge is worth building because it connects the green transition to the most widely used structural model in credit risk. A bank that uses Moody's KMV or a similar EDF model can incorporate environmental factors by adjusting the asset value and volatility inputs. The adjustment is not trivial, but it is tractable.

3.4 Case Studies of Green Rating

3.4.1 Case Study: A Technology Firm's Transition Rating

Consider a Chinese manufacturer of lithium-ion battery separators. The firm is listed on the Shenzhen Stock Exchange, has annual revenue of RMB 4.2 billion, and supplies separators to CATL and BYD. It is, by any reasonable definition, a green firm: its product is a critical component of electric vehicle batteries, and its growth is tied to the electrification of transport.

The standard credit rating for this firm would use leverage, coverage, profitability, and sector. The sector is "electronic components," which carries a moderate risk weight. The firm's leverage is 1.8x debt-to-EBITDA, coverage is 4.5x, and profitability is 12% net margin. On these metrics, it rates in the A range.

The transition rating tells a more complicated story. The firm's product is green, but its manufacturing process is energy-intensive and uses organic solvents. Its Scope 2 emissions are high relative to the sector. Its customers are demanding lower-carbon separators, and the firm has committed to a 30% reduction in emissions intensity by 2030. The commitment is credible—it has installed rooftop solar at two plants—but it is not yet reflected in the financials.

The transition score has four components. Emissions intensity is 1.4x the sector benchmark, which is a negative. Asset age is low—the plants were built in the past five years—which is a positive. Regulatory exposure is moderate: the firm operates in a jurisdiction with a nascent carbon market, and its exports to the EU will face the border adjustment mechanism. Physical risk is low: the plants are in inland Jiangsu, away from the coast.

The composite transition score is 0.45 on a 0-to-1 scale, where 1 is the worst. The PD adjustment is modest: the baseline PD of 0.8% rises to 0.92% with \(\alpha = 0.3\). The LGD adjustment is negligible because the collateral is specialized equipment with a ready market.

The interesting result is not the size of the adjustment but its direction. The firm is green in its output and brown in its process. A sector-based model would classify it as green and give it a pass. A firm-level model catches the process emissions and prices them. The difference is small in this case—12 basis points of PD—but it is the right kind of difference. It rewards the firm for its solar installation and penalizes it for its solvents. Over time, as the firm executes its emission-reduction plan, the score will improve and the PD will fall. The rating becomes a dynamic instrument, not a static label.

This is the first case in which the measurement core does what the green-finance literature promises: it identifies a transition-worthy firm, distinguishes it from a green-labeled firm, and creates an incentive for improvement.

3.4.2 Case Study: An Agricultural Cooperative's Credit Rating

The second case is harder, and it exposes the limits of the measurement core.

Consider a farmers' professional cooperative in Shandong province. The cooperative has 340 member households, grows apples and vegetables, and sells to a regional supermarket chain. It has applied for a working-capital loan of RMB 8 million to finance cold-storage capacity.

The cooperative has no credit rating in the conventional sense. It has no audited financials, no listed equity, and no bond history. The bank's PD model, calibrated on corporate borrowers, cannot score it. The standard workaround is to score the cooperative's members individually or to require a corporate guarantee. Neither is satisfactory: the members are smallholders with volatile incomes, and the guarantee is only as good as the guarantor.

The measurement core breaks down at the cooperative because the unit of analysis is wrong. The cooperative is not a firm; it is a collective of households whose incomes depend on weather, prices, and the cooperative's storage and marketing decisions. Its default probability is a function of the apple harvest, the vegetable price, and the cold-storage utilization rate. None of these appear in a corporate PD model.

This is where the environmental adjustment becomes essential rather than optional. The cooperative's physical risk is the dominant driver of its credit quality. A drought reduces the apple harvest; a flood damages the storage facility; a heat wave reduces the vegetable yield. The bank that ignores these factors is not measuring the cooperative's credit risk at all.

A workable approach has three steps. First, score the cooperative's physical risk using geolocation and historical yield data. The cooperative is in a region with moderate drought risk and low flood risk. The physical-risk score is 0.35. Second, score the cooperative's transition risk. The transition is less relevant for apples and vegetables than for carbon-intensive goods, but water pricing and pesticide regulation are real exposures. The transition score is 0.25. Third, combine the two scores with the cooperative's financial metrics—members' average income, the cooperative's leverage, the cold-storage utilization rate—into a composite rating.

The composite rating is not a PD in the conventional sense. It is a score that can be mapped to a PD using a calibration sample of similar cooperatives. The calibration sample is small, and the mapping is uncertain. But it is better than the alternative, which is to exclude the cooperative from credit altogether.

The Chinese context makes this case urgent. The People's Bank of China's Measures for the Administration of Loan Cards specify application requirements for individuals, enterprise legal persons, and public-institution legal persons, and do not cover farmers' professional cooperatives. The requirement that cooperative legal persons hold a loan card therefore effectively excludes them from borrowing through that channel. The measurement gap has teeth: it operates as a regulatory exclusion, not merely as a modeling problem. Building a rating methodology for cooperatives is a precondition for including them in the financial system at all.

3.4.3 Case Study: A Forestry Enterprise's Green Rating

The third case tests the measurement core at the boundary between credit risk and ecosystem risk.

Consider a state-owned forestry enterprise in Fujian province. The enterprise manages 120,000 hectares of plantation forest, produces timber and bamboo, and has issued a RMB 500 million green bond to finance reforestation. It is applying for a complementary bank loan of RMB 200 million.

The enterprise's credit quality depends on the forest's growth, the timber price, and the fire and pest risk. The standard PD model would use the enterprise's leverage, coverage, and the timber price. The environmental adjustment must capture the forest's physical risk and the carbon sequestration value.

The physical risk is substantial. Fujian is exposed to typhoons, and the plantation is monoculture Chinese fir, which is vulnerable to pests. The enterprise has insurance, but the coverage is partial. The physical-risk score is 0.55, which is high.

The transition risk is unusual. The forest is a carbon sink, and the enterprise earns carbon credits under China's national emissions trading scheme. The carbon revenue is a new income stream that offsets some of the timber-price risk. The transition score is 0.30, which is moderate.

The LGD adjustment is where the forestry case differs from the technology case. The collateral is the forest itself, and the forest's value depends on its health. A fire or pest outbreak destroys the collateral. The LGD model must therefore incorporate the physical-risk score directly, not just as a PD adjustment. The baseline LGD of 35% rises to 48% with \(\delta = 0.6\) and \(P = 0.55\).

The case illustrates a general point: for natural-capital collateral, the physical-risk adjustment to LGD is larger than the transition-risk adjustment to PD. The forest's credit quality is determined less by the carbon price than by the fire risk. A measurement system that focuses only on transition risk would miss the dominant driver.

The forest insurance market is the natural complement. I will return to it in Chapter 8. For now, the point is that the credit rating and the insurance are two measurements of the same risk. The bank's LGD model and the insurer's premium should be consistent. If they diverge, one of them is wrong.

3.4.4 Cross-Case Lessons

The three cases share a structure and differ in their parameters. In each, the measurement core provides the skeleton: PD, LGD, EAD, and tenor. In each, the environmental adjustment provides the flesh: a transition score, a physical-risk score, and a mapping to the parameters. In each, the data is imperfect and the calibration is uncertain.

The differences are instructive. For the technology firm, the transition adjustment dominates and the physical adjustment is negligible. For the cooperative, the physical adjustment dominates and the transition adjustment is secondary. For the forestry enterprise, both matter, and the LGD adjustment is larger than the PD adjustment. A single green rating methodology cannot handle all three. The methodology must be modular, with weights that vary by sector and by collateral type.

The cases also show that green identification is not the same as green labeling. The battery separator firm is green in output and brown in process; the cooperative is green in process and exposed to physical risk; the forestry enterprise is green in both and exposed to ecosystem risk. A labeling system would classify all three as green. A measurement system distinguishes them and prices them differently. The distinction is the value that credit risk measurement adds to the green transition.

3.5 The Identification Boundary: What Measurement Can and Cannot Do

3.5.1 The Limits of Quantification

I have argued that environmental factors can be incorporated into the measurement core. I have not argued that they can be incorporated precisely. The distinction matters.

The precision of a PD estimate depends on the size and quality of the calibration sample. For corporate borrowers in developed markets, the sample is large and the estimate is precise to within a few basis points. For cooperatives in emerging markets, the sample is small and the estimateis precise to within a few hundred basis points. The environmental adjustment adds a second layer of uncertainty on top of the first.

A bank that reports a PD of 0.92% for the battery separator firm is overstating its precision. The honest statement is a range: 0.8% to 1.1%, with a central estimate of 0.92%. The range reflects the uncertainty in the transition score, not just the sampling error in the baseline. A measurement system that reports point estimates without ranges is a system that hides its own limitations.

This has a practical consequence for lending decisions. A loan officer who sees a PD of 0.92% will price the loan at the corresponding spread. A loan officer who sees a range of 0.8% to 1.1% will price at the upper end, or will ask for more collateral, or will escalate the decision. The range changes behavior. It also changes the conversation with the borrower: a firm that wants a lower spread can improve its transition score by installing solar or switching solvents, and the improvement will show up in the range before it shows up in the point estimate.

3.5.2 The Greenwashing Risk

The measurement core is also vulnerable to gaming. A firm that knows its environmental performance affects its cost of capital has an incentive to report favorably. The incentive is strongest where the data is self-reported and unaudited, which is to say in Tier 2 and Tier 3.

The standard defenses against greenwashing are disclosure standards, third-party verification, and supervisory oversight. The EU's CSRD requires audited sustainability reporting for large firms; the ISSB's IFRS S1 and S2 provide a global baseline; the EBA's disclosure requirements for banks create a paper trail. These defenses are necessary but not sufficient. They raise the cost of misreporting, but they do not eliminate it.

The measurement response to greenwashing is triangulation. A firm's reported emissions can be checked against its energy purchases, its input quantities, and its output quantities. A cement plant that reports low emissions but purchases large quantities of coal is reporting inconsistently. A battery separator plant that reports low Scope 2 emissions but has no renewable energy contract is reporting implausibly. The triangulation does not require a forensic audit; it requires a model that expects consistency across data sources.

The credit rating agencies have a role here. Their ESG assessments are separate from their credit ratings, but the underlying data is the same. A bank that uses the agencies' data is relying on the agencies' verification. A bank that collects its own data is relying on its own. The choice is a trade-off between cost and control.

3.5.3 The Transition-Worthy Firm Problem

The hardest identification problem is not the green firm or the brown firm. It is the transition-worthy firm: the firm whose current environmental performance is poor but whose trajectory is improving. The battery separator firm is a mild case; a steel producer investing in electric-arc furnaces is a stronger one.

The measurement core struggles with transition-worthy firms because PD models are calibrated on current financials, not on future trajectories. A steel producer that spends heavily on decarbonization has lower current profitability and higher leverage than a peer that does not. The standard model penalizes it. The transition adjustment should reward it, but the reward is speculative: the investment may fail, the technology may not scale, the carbon price may not rise as expected.

The solution is to separate the transition-worthy firm's rating into two components: a current rating and a trajectory rating. The current rating uses the standard financials and the current environmental performance. The trajectory rating uses the firm's capital expenditure plan, its emission-reduction commitments, and its management's track record. The two ratings are reported separately, and the loan is priced on a weighted average that depends on the loan's tenor. A short loan is priced on the current rating; a long loan is priced on the trajectory rating.

This is not a standard practice. It requires a bank to make judgments about management quality and technology viability that go beyond the financial statements. But it is the only way to avoid penalizing the firms that are doing the transition. A measurement system that cannot distinguish a steel producer investing in decarbonization from a steel producer that is not is a system that will starve the transition of capital.

3.5.4 The Aggregation Problem

Individual PDs are not enough. A bank's portfolio risk depends on the correlation of defaults across borrowers. Two firms with the same PD can produce very different portfolio risk if one is a solar developer and the other is a coal miner, because their defaults are correlated with different macro factors.

Environmental factors change the correlation structure. A carbon price shock raises the PD of every carbon-intensive firm at once. A drought raises the PD of every farm in a region at once. The standard portfolio models—CreditMetrics, CreditRisk+, KMV Portfolio Manager—estimate correlations from historical equity returns or from sector factors. Neither captures the environmental correlation.

The measurement response is to add environmental factors to the correlation model. A carbon-price factor, a water-stress factor, and a physical-hazard factor can be estimated from market data (for listed firms) or from scenario analysis (for unlisted firms). The factors are not independent of the standard macro factors—a carbon price shock is correlated with an energy price shock—but they add explanatory power.

The ECB's 2022 climate stress test found that the correlation of transition risk across euro-area banks was higher than the banks' models assumed. The finding is not surprising: the banks' models were calibrated on a period when transition risk was not a common factor. The stress test forced the banks to model it as one.

3.5.5 What Measurement Cannot Do

I will be direct about the limits. Credit risk measurement cannot identify a green firm with certainty. It cannot price a transition-worthy firm's trajectory with precision. It cannot eliminate greenwashing. It cannot substitute for policy.

What it can do is make environmental performance legible to the financial system. A firm that reports its emissions and its transition plan can be scored. A firm that does not can be benchmarked. A bank that scores its borrowers can price them differentially. The differential is the signal that directs capital.

The signal is imperfect. It is noisy, gameable, and uncertain. But it is better than the alternative, which is a financial system that prices all firms as if the transition were not happening. The mispricing of the alternative is not zero; it is the accumulated cost of ignoring a risk that is already materializing. The measurement core, extended with environmental factors, reduces that mispricing. It does not eliminate it.

3.6 From Measurement to Identification: The Institutional Layer

3.6.1 Why Models Need Institutions

A PD model is a piece of mathematics. It becomes a lending decision only when it is embedded in an institution: a credit committee, a risk appetite framework, a pricing engine, a monitoring process. The institutional layer determines whether the model's output changes behavior or sits in a report that nobody reads.

The institutional layer is where most green-finance initiatives fail. A bank can build a sophisticated environmental scorecard and still lend to brown firms if the scorecard is not wired into the approval process. The wiring requires three things: a mandate, a limit, and a price.

The mandate is a credit policy that says environmental performance is a material factor in credit assessment. The EBA's 2020 guidelines on loan origination and monitoring provide a template: banks must consider environmental factors in the creditworthiness assessment, must document the consideration, and must monitor it over the life of the loan. The mandate is necessary but not sufficient; a policy that is not enforced is a policy that does not exist.

The limit is a portfolio constraint that caps exposure to high-transition-risk sectors or to high-physical-risk geographies. A limit converts the model's output into a hard constraint. It also creates an incentive for the business units to find green borrowers: if the brown exposure is capped, the only way to grow is to lend green.

The price is a spread differential that reflects the environmental adjustment. A bank that adjusts PD but not price is leaving money on the table for brown firms and overcharging green ones. The price differential is the signal that reaches the borrower. Without it, the borrower has no reason to improve.

3.6.2 The Chinese Green Credit Framework

China's green credit framework is the largest institutional experiment in this space, and it is worth examining closely because it shows both the promise and the limits of the institutional layer.

The framework dates to 2012, when the China Banking Regulatory Commission issued the Green Credit Guidelines. The guidelines require banks to establish environmental and social risk management systems, to identify and measure environmental risks in credit activities, and to restrict lending to polluting industries. The 2012 guidelines were followed by the 2014 Green Credit Key Performance Indicators, which made green credit a scored item in bank evaluations, and by the 2016 guidance from the People's Bank of China on green bond issuance.

The framework's strength is its integration into the supervisory process. Green credit is not a voluntary initiative; it is a scored item in the annual assessment of every bank. The score affects the bank's regulatory standing, which affects its access to liquidity and its ability to expand. The institutional wiring is in place.

The framework's weakness is the measurement underneath it. The green credit statistics are based on a catalogue of eligible projects, not on a firm-level rating. A loan is classified as green if it finances a project on the list; the borrower's overall environmental performance is not assessed. The result is a classification system that can be gamed: a loan to a coal company can be classified as green if it finances a wind farm, even if the coal company's overall emissions are rising.

The People's Bank of China has recognized the weakness. The 2021 Green Finance Evaluation Scheme for Banking Institutions introduced a more granular assessment, including a measure of the carbon intensity of the loan book. The scheme is a step toward firm-level measurement, but it is not yet a full rating methodology.

The Chinese experience suggests a general lesson: the institutional layer is necessary but not sufficient. A mandate without a measurement is a mandate that produces statistics, not a change in capital allocation. The measurement must be firm-level to change behavior at the firm level.

3.6.3 The EU Taxonomy as a Measurement Standard

The EU Taxonomy Regulation, which entered into force in 2020, is the most ambitious attempt to define what counts as green. It sets technical screening criteria for economic activities, requiring substantial contribution to one of six environmental objectives and no significant harm to the others.

The taxonomy is a classification standard, not a credit risk model. It tells a bank whether an activity is green; it does not tell a bank whether a borrower is likely to default. The two questions are different, and conflating them is a mistake.

The taxonomy's contribution to measurement is indirect. It provides a common language for defining green activities, which reduces the ambiguity in green lending. It also provides a data standard: a firm that reports its taxonomy-aligned revenue and capital expenditure gives a bank a measurable input. The CSRD requires large firms to report taxonomy alignment, which means the data will become available.

The taxonomy's limitation is that it is binary. An activity is either aligned or not; there is no partial credit. A steel producer that reduces its emissions by 40% but is still above the threshold is not aligned, even though it is on a transition path. The taxonomy has a transition category under development, but it is not yet operational.

For credit risk measurement, the binary classification is a problem. A transition-worthy firm that is not yet aligned gets no credit in the taxonomy, even though its trajectory is improving. The measurement core needs a continuous score, not a binary label. The taxonomy can inform the score, but it cannot substitute for it.

3.6.4 The Supervisor's Role

Supervisors have a distinctive role in the measurement layer. They do not lend, but they set the rules under which lending occurs. Their choices determine whether environmental measurement is a compliance exercise or a risk-management tool.

The ECB's approach has been to set expectations and then assess against them. The 2020 guide on climate-related and environmental risks sets out 13 expectations, covering governance, risk management, and disclosure. The 2022 stress test assessed the banks against the expectations and found significant gaps. The gaps were published, which creates reputational pressure to close them.

The EBA's approach has been to build the infrastructure. The 2021 report on management and supervision of ESG risks provides a methodology for assessing ESG risks, including a mapping of transmission channels from environmental factors to credit risk. The 2021 pilot exercise on climate risk tested the methodology on a sample of banks. The results were mixed: the banks could estimate exposure but not default probabilities.

The Chinese approach has been to score and rank. The Green Finance Evaluation Scheme rates banks on a set of indicators, including the share of green credit and the carbon intensity of the loan book. The ratings are published, which creates a league table. The league table creates competition, which drives behavior.

The three approaches are complementary. The ECB sets expectations, the EBA builds methods, and the PBOC scores outcomes. A complete supervisory framework would combine all three: expectations, methods, and scores. No jurisdiction has yet done so.

3.7 The Measurement Core Extended: A Synthesis

3.7.1 The Architecture

I will now synthesize the chapter's argument into an architecture. It runs from data to scores, from scores to parameter adjustments, and from adjustments to decisions.

flowchart TD
    A[Layer 1: Data] --> B[Layer 2: Scores]
    B --> C[Layer 3: Adjustments]
    C --> D[Layer 4: Decisions]
    A1[Reported emissions] --> A
    A2[Estimated emissions] --> A
    A3[Geolocation] --> A
    A4[Asset register] --> A
    A5[Insurance data] --> A
    B1[Transition score T] --> B
    B2[Physical score P] --> B
    B3[Financial rating] --> B
    C1[PD adjustment] --> C
    C2[LGD adjustment] --> C
    C3[Correlation adjustment] --> C
    D1[Pricing] --> D
    D2[Limits] --> D
    D3[Monitoring] --> D

The data layer is tiered by availability and quality, as described in Section 3.3.4. The scoring layer converts data into a transition score and a physical score. The adjustment layer maps the scores into PD, LGD, and correlation parameters. The decision layer converts the adjusted parameters into prices, limits, and monitoring triggers.

The architecture is modular. A bank can implement the data and scoring layers without the adjustment layer, using the scores as a qualitative overlay. It can implement the adjustment layer for one parameter (PD) before the others. The modularity allows incremental adoption, which is important because no bank has the resources to build the whole architecture at once.

3.7.2 The Calibration Discipline

The architecture's credibility depends on calibration. A score that is not calibrated to outcomes is a score that cannot be trusted. Calibration requires three things: a definition of the outcome, a sample of observations, and a method for mapping scores to outcomes.

The outcome for PD calibration is default. The sample is the population of borrowers with known environmental scores and known default outcomes. The method is a statistical model—logistic regression, survival analysis, or machine learning—that estimates the relationship between the score and the default probability.

The sample problem is severe. Few banks have a history of environmental scores linked to default outcomes, because the scores are new. The workaround is to use external data: the rating agencies' ESG assessments linked to their default histories, the CDP's climate disclosures linked to bond defaults, the academic literature's studies of environmental performance and financial distress. The external data is imperfect, but it is better than no calibration.

The calibration should be revisited annually. The relationship between environmental scores and defaults will change as the transition accelerates. A model calibrated on 2020 data will be stale by 2025. The recalibration is not a technical nicety; it is a condition for the model's continued validity.

3.7.3 The Governance of Measurement

A measurement system is a governance system. The choices about what to measure, how to weight it, and how to use it are political choices, not technical ones. They determine which firms get capital and which do not.

The governance of measurement requires three things: transparency, challenge, and accountability. Transparency means the methodology is documented and disclosed. Challenge means the methodology is subject to internal and external review. Accountability means the methodology's outputs are linked to decisions for which someone is responsible.

The governance failure mode is the black box. A proprietary model that nobody outside the model team understands is a model that cannot be challenged. It may be right, but its rightness is unverifiable. The green-finance literature has spent too much time on model design and too little on model governance. The design is the easy part; the governance is the work.

3.7.4 The Feedback Loop

The final element of the architecture is the feedback loop. The measurement system's outputs—prices, limits, monitoring triggers—change borrower behavior. A firm that faces a higher spread for its emissions has an incentive to reduce them. A firm that faces a lower spread for its transition plan has an incentive to execute it. The measurement system is not a passive observer; it is an active participant in the transition.

The feedback loop is the mechanism by which measurement becomes identification. A firm is identified as green not because a label is attached to it but because the measurement system prices it as green and the market responds. The identification is dynamic: as the firm improves, its score improves, and its cost of capital falls. The dynamic is the point.

The feedback loop also creates a risk. If the measurement system is wrong—if it rewards a firm that is not actually green—the feedback loop reinforces the error. A firm that games the score gets a lower cost of capital and expands, crowding out firms that are genuinely green. The governance of measurement is the defense against this risk. Transparency, challenge, and accountability are not bureaucratic formalities; they are the conditions for the feedback loop to work in the right direction.

3.8 Conclusion: Measurement as the First Act of Identification

The chapter began with a gap: the Basel Committee's 27 frameworks and the absence of a PD methodology. It ends with an architecture that fills the gap, imperfectly.

The architecture's core claim is that environmental performance can be incorporated into the measurement core without replacing it. The three parameters—PD, LGD, EAD—remain the anchor. The environmental adjustment is a layer on top, not a substitute. A bank that adopts the architecture does not abandon its existing models; it extends them.

The extension is not costless. It requires data, calibration, governance, and institutional wiring. The costs fall disproportionately on smaller banks, which is why the supervisory layer matters: a common methodology, or at least a common data standard, reduces the fixed cost of adoption.

The extension is also not complete. The calibration is uncertain, the data is uneven, and the transition-worthy firm problem is unsolved. I have been honest about these limits because a measurement system that hides its limits is a system that will fail when the limits bind. The green transition is a long event; the measurement system will need to evolve with it.

The next chapter takes the architecture and applies it to the specific case of technology firms. The battery separator firm is a preview: a firm whose output is green and whose process is brown, whose rating depends on the resolution of that tension. The resolution requires a rating methodology that is firm-specific, dynamic, and transparent. Chapter 4 builds it.

The institutions exist: the PD models, the rating scales, the supervisory frameworks. Extending those institutions to environmental performance is the work this chapter has described, and it is unfinished. The steel producer investing in electric-arc furnaces shows why. The current rating penalizes it; the trajectory rating rewards it. The two ratings are not a contradiction; they are a measurement of a firm in transition. A financial system that can hold both ratings in view is a system that can finance the transition. A system that can hold only one is a system that will finance either the past or the future, not the passage between them.

The measurement core is the first act of identification. It is not the last. Insurance, intermediation, and exit mechanisms follow. But without measurement, the later acts have no script. The script begins with a number: the probability of default, adjusted for the transition.