Chapter 3: Credit Rating and Environmental Performance: Extending Traditional Assessment

A mid-sized electrolyzer manufacturer in Jiangsu applied for a RMB 300 million term loan in the spring of 2021. Its order book was full, gross margin had held above 22 percent for three years, and leverage sat at 48 percent. Under the bank's standard scorecard it cleared the investment-grade threshold with room to spare. The loan committee rejected the application. Not on financial grounds. A single customer accounted for 61 percent of revenue, and that customer was a state-owned utility whose own credit had been placed on negative watch after a carbon-price stress test. The manufacturer's ratios described a healthy firm. Its environmental exposure described a fragile one. The scorecard could see only the first.

That gap is the subject of this chapter. Traditional credit rating was built to answer a narrow question: what is the probability that this borrower fails to pay? It answers well for firms whose risks are financial and whose cash flows are legible. It answers poorly for green technology enterprises, whose risks are technological, regulatory, and environmental, and whose cash flows depend on policy and on the pace of a transition no single firm controls. Adding a sustainability column to the spreadsheet does not extend the rating. The extension requires deciding which indicators carry information about default, how those indicators are weighted, and where the resulting score enters the rating decision.

3.1 Traditional Credit Rating and Its Limits

3.1.1 What a Rating Is, and What It Is Not

A credit rating is an ordinal opinion about relative creditworthiness. The definition sounds anodyne, and the agencies have insisted on it for a century. Moody's, S&P Global Ratings, and Fitch all state in their published methodologies that a rating is an opinion, not a recommendation to buy, sell, or hold a security, and not a guarantee of any particular outcome. The disclaimer is legally important and analytically instructive. It tells us that a rating ranks borrowers against each other within a defined population and horizon. It does not measure an absolute probability, and it does not price risk.

The distinction between the agencies matters more than practitioners often acknowledge. S&P's issuer credit ratings primarily reflect the probability of default. Moody's ratings tend to reflect expected loss on a specific instrument, which is the probability of default multiplied by loss given default. A firm can carry an S&P issuer rating of BB and a Moody's instrument rating of B2 on the same obligation without either agency being wrong, because they answer different questions. This chapter works with the default-probability interpretation, since that is the quantity a bank's internal rating model must estimate if it is to feed a regulatory capital calculation or a loan-pricing decision. Where loss given default matters, I will say so.

The rating scale itself is familiar. Long-term issuer ratings run from AAA through D on the S&P scale, with modifiers (AA+, AA, AA−) inside the investment-grade band and numerical modifiers (Ba1, Ba2, Ba3) on the Moody's scale. Short-term ratings use a separate classification, A-1 through D. The granularity is coarse by design: a rating should be stable enough that a borrower does not migrate across categories every quarter, which means the model behind it must smooth over noise. That smoothing is a virtue for a bond investor. For a bank trying to detect a deteriorating green technology borrower whose distress shows up first in non-financial signals, it is a problem.

3.1.2 The Standard Rating Process

The rating process at a major agency follows a recognizable sequence. Analysts specialized by industry sector follow a relatively small number of entities, meet management, review audited statements and forecasts, apply a sector-specific criteria framework, and present to a rating committee. The committee, not the individual analyst, assigns the rating. This structure exists because a rating is a judgment that must survive challenge, and because the agency's reputation depends on consistency across issuers and across time.

The strength of this process is informational depth. Rating agencies have offices close to the entities they rate and access to chief financial officers who treat the rating as a capital-markets necessity. A bank analyst may never speak to management and may see only what the borrower chooses to disclose. The weakness is that the criteria frameworks were built for industrial and financial firms whose risk drivers are balance-sheet and cash-flow variables. When the framework encounters a firm whose principal risk is that its core technology becomes obsolete, or that a subsidy regime changes, or that a carbon border adjustment mechanism raises its export costs, the framework has no natural place to put that risk.

3.1.3 Three Frictions for Green Technology Enterprises

Three frictions stand out when traditional rating meets green technology borrowers.

The first is the mismatch between historical data and prospective risk. Rating models are estimated on default histories. A scorecard calibrated on a decade of defaults learns which financial ratios predicted failure in that decade. Green technology firms are young, and the population of green technology defaults is thin. A solar module manufacturer that failed in 2012 because of overcapacity and a collapse in European feed-in tariffs looks, in the data, like a firm with high leverage and negative operating cash flow. The model learns "high leverage predicts default" and misses the fact that the leverage was a symptom of a policy shock the firm could not hedge. The next green technology default may have entirely different proximate causes, and the model will not have learned them.

The second friction is the concentration of green technology risk in a small number of policy variables. A firm's cash flow may depend on a feed-in tariff, a production subsidy, a carbon price, a renewable portfolio standard, or a border adjustment. These variables are set administratively. They can change with a budget cycle or an election. Traditional rating frameworks treat regulatory risk as a qualitative overlay, a "country risk" or "industry risk" adjustment applied after the financial analysis. For a green technology firm, regulatory risk is the financial analysis. A feed-in tariff cut does not merely adjust the rating; it rewrites the revenue line.

The third friction is the intangibility of the asset base. Green technology firms hold intellectual property, process know-how, and customer relationships rather than the plant and equipment that traditional secured lending relies on. A conventional rating model that uses collateral coverage as a mitigant finds little to work with. The firm may be technologically excellent and still fail a loan committee's collateral test, because the collateral is a patent portfolio whose liquidation value is close to zero in the hands of anyone but a strategic buyer.

These three frictions do not make green technology firms unratable. They make them ratable only if the rating model is extended to see the risks that matter. That extension begins with the selection of environmental indicators.

3.1.4 The Evidence That Environmental Performance Carries Credit Information

Before building the extension, it is worth establishing that environmental performance contains information about credit outcomes at all. If it does not, the exercise is decorative.

The evidence is suggestive rather than conclusive, and I will state its limits plainly. A body of work on ESG ratings finds that the major providers disagree substantially on the same firms. Billio, Costola, Hristova, Latino, and Pelizzon (2020) document that heterogeneity in rating criteria can lead agencies to hold opposite opinions on the same evaluated company, and that agreement across providers is low. Escrig-Olmedo, Fernández-Izquierdo, Ferrero-Ferrero, Rivera-Lirio, and Muñoz-Torres (2019), in "Rating the Raters: Evaluating How ESG Rating Agencies Integrate Sustainability Principles" (Sustainability, 11(3), 915), find that ESG rating agencies do not fully integrate sustainability principles into their corporate assessment processes. If the environmental scores themselves are noisy and inconsistent, using them as inputs to a credit model introduces measurement error that can swamp the signal.

The credit-relevant channel does not require a comprehensive ESG score. It requires specific environmental variables that plausibly affect cash flow and asset value. Semet, Roncalli, and Stagnol (2021), in "The ESG–Credit Risk Puzzle: The Impact of ESG Scores on Sovereign Bond Spreads and Ratings" (working paper), examine whether ESG factors are priced in sovereign bond markets and by credit rating agencies, and find evidence of a relationship that varies by pillar and by region. Sautner, van Lent, Vilkov, and Zhang (2020), in "Firm-Level Climate Change Exposure" (working paper, later published in Journal of Finance), construct a firm-level measure of climate change exposure from earnings call transcripts and show that the measure varies meaningfully across firms and predicts option-implied tail risk. These results do not prove that environmental performance predicts default. They do establish that markets and firms behave as if environmental exposure matters, which is enough to justify including it in a rating model and testing whether it improves discrimination.

The honest position is this: environmental performance is a plausible credit risk driver, the evidence for its incremental predictive power is provisional, and any extended rating framework should be built so that the environmental component can be validated, recalibrated, or switched off if the data do not support it. A framework that cannot be falsified is not a model; it is a commitment.

3.2 Environmental Performance Indicators: Selecting and Weighting

3.2.1 The Selection Problem

The universe of possible environmental indicators is large and growing. Carbon emissions, energy intensity, water withdrawal, waste generation, pollutant discharge, land use, biodiversity impact, green revenue share, taxonomy alignment, transition plan credibility, board-level environmental oversight, third-party verification, and supply-chain environmental performance are all candidates. A rating model cannot use all of them. It must select a small set that is material to credit, measurable at reasonable cost, and available for the borrower population in question.

I propose four selection criteria, applied in order.

Materiality to cash flow or asset value. An indicator qualifies only if a plausible transmission mechanism connects it to the borrower's ability to service debt. Carbon emissions qualify because a carbon price or a border adjustment raises operating costs. Water withdrawal qualifies for a textile dyeing firm in a water-stressed basin because a withdrawal restriction halts production. Board-level environmental oversight does not qualify directly; it may matter as a governance signal, but the transmission to default is indirect and the indicator is easy to game.

Measurability and verifiability. The indicator must be quantifiable from data the lender can obtain, and ideally verifiable by a third party. Emissions from metered fuel consumption are measurable. "Environmental commitment" is not. The measurement problem is the binding constraint in most emerging markets, where environmental disclosure is voluntary and inconsistent.

Discrimination across the borrower population. An indicator that takes the same value for almost every borrower adds no information. A regional bank whose borrowers are all small manufacturers may find that none of them report scope 1 emissions, in which case the indicator cannot discriminate and should be replaced by a proxy such as energy expenditure per unit of output.

Stability and comparability over time. The indicator must be defined consistently enough that a score computed this year can be compared with a score computed three years ago. This criterion favors physical quantities (tons of CO2, cubic meters of water) over composite scores whose methodology changes.

Applying these criteria yields a core set of indicators that I will group into three families: emissions and energy, resource use and pollution, and transition positioning. The first two are backward-looking measures of environmental burden. The third is a forward-looking measure of how well positioned the firm is for a constrained-carbon economy.

3.2.2 Emissions and Energy Indicators

The most widely used emissions indicator is scope 1 and scope 2 greenhouse gas emissions, measured in tons of CO2 equivalent and normalized by revenue or by physical output. Normalization matters. An absolute emissions figure rewards small firms and penalizes large ones regardless of efficiency. Emissions intensity—tons of CO2e per million RMB of revenue, or per ton of product—allows comparison across firms of different sizes and is the appropriate input to a rating model.

The measurement hierarchy determines data quality. Scope 1 covers direct emissions from owned or controlled sources: fuel combustion in boilers, process emissions from chemical reactions, fugitive emissions from refrigerants. Scope 2 covers indirect emissions from purchased electricity, heat, or steam. Scope 3 covers everything else in the value chain, including purchased goods, business travel, and the use of sold products. For most manufacturers, scope 1 and 2 are measurable from fuel and electricity bills. Scope 3 requires supplier data that few small firms possess. A rating model for small and medium green technology enterprises should rely on scope 1 and 2 and treat scope 3 as a qualitative adjustment where data exist.

Energy intensity is a useful complement because it is available even when emissions are not. A firm that reports electricity consumption in kilowatt-hours and output in physical units gives the analyst a ratio that can be benchmarked against sector peers. In China, where the national emissions trading system initially covered only the power sector, most manufacturers did not report emissions at all until reporting requirements expanded. Energy intensity filled the gap.

Two boundary conditions apply. First, emissions intensity is only comparable within a sector. A cement plant and a software firm have intensities that differ by orders of magnitude, and pooling them in a single regression would produce a coefficient that describes the sector composition of the sample rather than the effect of emissions. The rating model must therefore benchmark within sector, using sector medians and interquartile ranges. Second, intensity can be reduced by outsourcing. A firm that closes its own boiler and buys steam from a neighboring plant reduces its scope 1 emissions and increases its scope 2 emissions, with no change in total atmospheric burden. The indicator is gameable, and the model should treat a sudden improvement in intensity with suspicion unless it is accompanied by a change in the production process rather than a change in organizational boundaries.

3.2.3 Resource Use and Pollution Indicators

Water withdrawal intensity, measured in cubic meters per unit of output, matters for firms operating in water-stressed basins. The transmission mechanism is direct: a withdrawal restriction, a discharge standard, or a water price increase raises costs or halts production. The indicator should be weighted by basin stress. A textile firm in the Taihu basin faces a different water risk from an identical firm in a water-abundant region, and a rating model that ignores the basin treats them as equivalent when they are not.

Pollutant discharge—chemical oxygen demand, sulfur dioxide, nitrogen oxides, particulate matter—is regulated through discharge permits and effluent standards. A firm operating close to its permit limit faces a real risk that a tightening of the standard forces capital expenditure or curtailment. The indicator is the ratio of actual discharge to permitted discharge, which measures headroom. A firm at 95 percent of its permit has less room to grow than a firm at 60 percent, and the rating model should reflect that.

Waste generation and hazardous waste handling are relevant for manufacturing firms but are harder to benchmark because definitions vary. I would include hazardous waste as a binary indicator (generates hazardous waste: yes or no) and rely on permit compliance rather than volume, because volume data are unreliable.

The common thread across these indicators is that they measure regulatory headroom. Environmental regulation works through permits, standards, and prices. A firm's exposure to a tightening of regulation is a function of how close it operates to the current limit. This is the credit-relevant quantity, and it is more informative than the absolute level of pollution.

3.2.4 Transition Positioning Indicators

The backward-looking indicators describe where a firm has been. The forward-looking indicators describe where it is going. For a green technology enterprise, the forward-looking dimension is where most of the credit risk and most of the credit opportunity reside.

The first transition indicator is green revenue share: the proportion of revenue attributable to products or services that are environmentally beneficial. The definition should follow a published taxonomy where one exists. The EU taxonomy provides a classification of economic activities that contribute substantially to environmental objectives, and the Chinese Green Bond Endorsed Projects Catalogue (2021 edition) provides an analogous list. Using a published taxonomy rather than an ad hoc definition makes the indicator comparable across firms and defensible to a loan committee. A firm with 80 percent green revenue is a different credit from a firm with 15 percent green revenue and a declining brown business, even if their current financial ratios are identical.

The second indicator is transition plan credibility. This is the hardest to measure and the easiest to fake. A credible transition plan specifies targets, dates, capital expenditure commitments, and governance. A firm that announces net-zero by 2050 without interim targets or capital allocation has announced an aspiration, not a plan. I would score credibility on three observable features: whether the plan includes interim targets at five-year intervals, whether capital expenditure is explicitly allocated to the transition, and whether progress is reported annually against the targets. A plan that satisfies all three is credible; a plan that satisfies none is not.

The third indicator is technology obsolescence risk, which is specific to green technology firms. A firm producing a component for a technology that is being displaced—a supplier of silicon-based photovoltaic cells as perovskite tandem cells approach commercialization, for instance—faces a risk that its asset base loses value faster than its depreciation schedule assumes. This risk is difficult to quantify but can be proxied by the firm's research and development intensity and by the rate at which its product specifications are being superseded in the market. A rating model can include R&D intensity as a proxy and flag firms whose core product is on a published technology roadmap's displacement list.

3.2.5 Weighting: The Case Against Equal Weights

Once indicators are selected, they must be weighted. The default in many ESG scoring systems is equal weighting, which is a choice, not a neutral one. Equal weighting implies that each indicator is equally informative about the outcome of interest. That implication is almost certainly false.

Consider two indicators for a steel mill: emissions intensity and board-level environmental committee. Equal weighting gives them the same influence on the score. But emissions intensity transmits to cash flow through carbon costs and compliance expenditure, while a board committee transmits only if it changes decisions, and the evidence that it does is weak. The credit-relevant weight on emissions intensity should be higher.

I propose weighting by the strength of the transmission mechanism to default, estimated in three steps.

The first step is to specify the mechanism for each indicator. Carbon intensity affects default through operating cost under a carbon price, through compliance capital expenditure, and through asset stranding risk. Water withdrawal affects default through curtailment risk and water price. Green revenue share affects default through demand growth and through access to green finance at preferential rates. Each mechanism gives a predicted sign and a predicted magnitude.

The second step is to estimate the relationship empirically where data permit. A bank with a portfolio of borrowers and a default history can regress default on indicator values, controlling for financial ratios, and use the estimated coefficients as weights. Where default histories are too thin, as they are for green technology firms, the bank can use a broader proxy: the indicator's relationship with rating migration, with interest coverage deterioration, or with the probability of a covenant breach. These are not defaults, but they are correlated with default and are more frequently observed.

The third step is to impose judgment where the data are silent. This is unavoidable and should be transparent. A weight that is set by expert judgment rather than estimated should be documented as such, with the reasoning stated, so that it can be revised when data accumulate.

The resulting weights should satisfy two constraints. They should sum to one within the environmental component, and the environmental component as a whole should carry a weight in the overall rating that reflects its incremental information. If environmental indicators explain 5 percent of the variation in default after financial ratios are controlled for, giving them 30 percent of the overall rating overweights them and will degrade the model's accuracy. The weight should be calibrated, not asserted.

3.2.6 The Measurement Problem in Emerging Markets

A framework designed in Frankfurt or London must survive contact with a county-level branch in a province where the borrower's environmental data consist of a self-reported form and an occasional inspection record.

Three problems recur. The first is missing data. Many small firms do not report emissions, water use, or waste. The rating model must handle missingness without treating it as a zero, which would reward non-disclosure. The standard approach is to impute from sector and size benchmarks and to flag the imputation, so that the loan committee knows the score rests on an assumption. A firm that refuses to disclose should not receive the sector median by default; it should receive a penalty that reflects the informational risk.

The second problem is unreliable data. Self-reported figures can be audited or not. Where third-party verification exists, the indicator should be weighted more heavily. Where it does not, the indicator should be discounted. A simple approach is to multiply the indicator's contribution by a verification factor: 1.0 for third-party verified data, 0.7 for self-reported data with supporting documentation, 0.4 for self-reported data without documentation, and 0 for no data.

The third problem is that environmental data are often collected at the facility level while the rating is at the entity level. A firm with three plants may report emissions for one. The model must aggregate across facilities, and where facility-level data are incomplete, it should use production-weighted averages rather than simple sums, which would understate the entity's intensity.

These are unglamorous problems. They determine whether the framework works in practice. A model that is elegant in theory and unusable in a county branch is not a model that will be adopted.

3.3 An Extended Rating Framework: Integrating Environmental Scores into Ratings

3.3.1 The Architecture

The extended framework has four layers: a financial score, an environmental score, a transition-risk adjustment, and a rating mapping. The financial score is the conventional scorecard, unchanged. The environmental score aggregates the selected indicators using the calibrated weights. The transition-risk adjustment captures the forward-looking risks that the indicator set does not fully measure. The rating mapping translates the combined score into a rating category and, where the bank uses an internal ratings-based approach, into a probability of default.

The key design decision is where the environmental score enters. There are three options.

The first option is a separate environmental rating displayed alongside the credit rating. This is the approach most ESG rating agencies take. It preserves the integrity of the credit rating and provides information without altering the decision. Its weakness is that the environmental rating may be ignored by a loan committee under time pressure.

The second option is to fold the environmental score into the financial score as an additional input, so that the combined score determines the rating. This gives the environmental score real influence but risks double-counting: a firm with high carbon costs already shows elevated operating expenses in its financial statements, and adding an emissions penalty on top penalizes it twice.

The third option is a notching approach: compute the conventional rating, then notch it up or down by a limited number of grades based on the environmental score and the transition-risk adjustment. This is the approach I favor, for three reasons. It preserves the transparency of the conventional rating, so the committee can see what the environmental adjustment did. It bounds the influence of the environmental component, which is appropriate given the provisional state of the evidence. And it is operationally familiar, since rating agencies already use notching for structural subordination and other adjustments.

The notching should be asymmetric. A strong environmental score should notch up by at most one grade, because environmental strength is a mitigant, not a driver of repayment. A weak environmental score should notch down by up to two grades, because environmental weakness can accelerate distress. The asymmetry reflects the fact that environmental risk is primarily a downside risk for a lender.

3.3.2 The Environmental Score

The environmental score \(E\) is a weighted sum of normalized indicator values:

where \(x_k\) is the raw value of indicator \(k\), \(\tilde{x}_k\) is its normalized value, and \(w_k\) is its weight. Normalization maps each indicator onto a 0–100 scale using sector-specific benchmarks:

with \(x_k^{\min}\) and \(x_k^{\max}\) set at the 5th and 95th percentiles of the sector distribution, so that outliers do not compress the scale. For indicators where a higher value is better (green revenue share), the normalization is applied directly. For indicators where a higher value is worse (emissions intensity), the scale is inverted:

The weights \(w_k\) are calibrated as described in section 3.2.5 and sum to one. A reasonable starting configuration, before empirical calibration, gives 30 percent to emissions intensity, 15 percent to energy intensity, 15 percent to water and pollution headroom, 25 percent to green revenue share, and 15 percent to transition plan credibility. These weights are a prior, not an estimate. They should be replaced by estimated weights as data accumulate.

The environmental score is then adjusted for data quality:

where \(q\) is the verification factor described in section 3.2.6, bounded between 0.5 and 1.0 so that poor data quality discounts the score without eliminating it.

3.3.3 The Transition-Risk Adjustment

The environmental score captures the firm's current position. The transition-risk adjustment captures its exposure to a change in the policy or technology environment.

I model transition risk as the product of exposure and resilience:

where \(p_j\) is the probability of scenario \(j\) (a carbon price increase, a subsidy reduction, a technology standard tightening), \(\Delta C_j\) is the change in the firm's annual cost under that scenario as a fraction of EBITDA, and \(R \in [0,1]\) is a resilience index measuring the firm's ability to absorb the cost through pricing power, cost pass-through, or capital reserves.

The scenario probabilities are the weakest link. They are policy forecasts, and policy forecasts are unreliable. I would anchor them to published policy trajectories where they exist—a carbon price path consistent with a jurisdiction's stated target, a subsidy phase-down schedule already legislated—and treat scenarios without a legislative basis as low-probability. The point is not to predict policy but to ensure that the rating reflects the firm's sensitivity to plausible changes.

The resilience index \(R\) is estimated from three observable features: the firm's gross margin relative to sector peers (higher margin implies more absorptive capacity), the share of costs that can be passed through under existing contracts (a regulated utility with cost-of-service pricing has high pass-through; a commodity manufacturer has low), and the firm's undrawn committed credit facilities as a fraction of annual operating costs. Each feature is scored 0–1 and averaged.

The transition adjustment enters the notching decision. A firm with high exposure and low resilience receives a downward notch. A firm with high exposure and high resilience receives no adjustment. A firm with low exposure receives no adjustment regardless of resilience.

3.3.4 The Notching Rule

The combined adjustment is:

with the following structure. If \(E^{*} \geq 70\) and \(T \geq -0.05\) (transition risk below 5 percent of EBITDA), notch up by one. If \(E^{*} < 40\) or \(T < -0.20\), notch down by two. If \(E^{*} < 50\) or \(T < -0.10\), notch down by one. Otherwise, no adjustment.

The thresholds are illustrative and should be calibrated to the bank's portfolio. Their purpose is to make the adjustment rule explicit and auditable, so that two analysts applying the framework to the same borrower reach the same notch.

3.3.5 Worked Example: The Electrolyzer Manufacturer

Return to the Jiangsu manufacturer. Its financial scorecard places it at BBB− on the internal scale. The environmental assessment proceeds as follows.

Emissions intensity is 12 percent below the sector median, normalized to 62. Energy intensity is 8 percent below median, normalized to 58. Water and pollution headroom is adequate, normalized to 65. Green revenue share is 74 percent, normalized to 78. Transition plan credibility is moderate: the firm has a five-year target and reports annually, but capital expenditure is not explicitly allocated, so it scores 55. The weighted environmental score is:

Data quality is good: emissions are third-party verified, so \(q = 1.0\) and \(E^{*} = 64.3\).

The transition-risk assessment identifies two scenarios. A carbon price of RMB 100 per ton (probability 0.4) raises annual costs by 6 percent of EBITDA. A reduction in the green hydrogen subsidy (probability 0.3) reduces revenue by 9 percent of EBITDA. The expected cost is \(0.4 \times 0.06 + 0.3 \times 0.09 = 0.051\), or 5.1 percent of EBITDA. Resilience is moderate: gross margin is at the sector median (0.5), cost pass-through is limited by fixed-price contracts (0.3), and undrawn facilities cover 40 percent of annual operating costs (0.4). The resilience index is 0.4, and \(T = -(1 - 0.4) \times 0.051 = -0.031\).

With \(E^{*} = 64.3\) and \(T = -0.031\), the firm falls in the "no adjustment" band. Its rating remains BBB−.

Now change one parameter. Suppose the firm's largest customer—the state-owned utility on negative watch—accounts for 61 percent of revenue, and that customer's carbon exposure implies a 15 percent probability of a contract renegotiation that would cut revenue by 20 percent of EBITDA. This is a concentration risk that the financial scorecard already flags but does not quantify environmentally. Adding it to the transition scenarios raises the expected cost to \(0.051 + 0.15 \times 0.20 = 0.081\). With resilience 0.4, \(T = -0.049\), which crosses the \(-0.05\) threshold only marginally. The notch down by one applies if we also judge \(E^{*}\) to be below 50 after reweighting for the customer's exposure—which it is not. The example shows the framework's discipline: a real risk that intuition says should matter produces a small adjustment because the firm's resilience absorbs most of it. Whether that is the right answer is an empirical question, and the framework is built so that the question can be tested.

3.3.6 Calibration and Validation

A rating model that cannot be validated is a scoring exercise. The extended framework should be validated on three dimensions.

Discriminatory power. Does the model separate defaulters from non-defaulters better than the financial-only model? The standard measure is the area under the receiver operating characteristic curve, or AUC. An AUC of 0.5 is no discrimination; 1.0 is perfect. A financial-only model for mid-sized firms typically achieves 0.70–0.80. The extended model should be tested for incremental AUC, and if the increment is not statistically significant, the environmental component should be reweighted or dropped.

Stability. Does the model's output migrate in response to changes in environmental performance, or is it dominated by the financial score? A model whose environmental component never changes the rating is decorative. A model whose environmental component changes the rating every quarter is noisy. The target is a model where environmental adjustments are infrequent but consequential.

Calibration. Do borrowers with a given score default at the predicted rate? This requires a default history, which for green technology firms is thin. The practical approach is to calibrate on a broader portfolio that includes brown firms with environmental data, then test the green technology subset separately as data accumulate.

The validation should be conducted out of sample and out of time. A model estimated on 2015–2019 data and tested on 2020–2022 data will reveal whether the relationships are stable. The COVID-19 period is a useful stress test because it combined a demand shock with a policy shock, and firms with different environmental profiles responded differently.

Concretely, the calibration exercise can be structured as a three-stage test. In the first stage, the bank estimates the financial-only model on a training window and records the score distribution. In the second stage, it adds the environmental score and transition adjustment and re-estimates. In the third stage, it compares the two models on a holdout window using three statistics: the AUC, the Kolmogorov–Smirnov statistic (which measures the maximum separation between the cumulative default distributions of good and bad borrowers), and the Hosmer–Lemeshow test (which checks whether the predicted default rates match the observed rates across score deciles). A model that improves the AUC but fails the Hosmer–Lemeshow test is discriminating better without being calibrated; that is a warning sign, not a success.

The population problem deserves a separate note. Green technology firms in a bank's portfolio are few, and defaults among them are fewer. A direct estimation of the environmental component's coefficient on this subsample will have wide confidence intervals. The practical solution is to estimate the coefficient on the full portfolio, where brown firms provide variation in environmental indicators, and then test whether the coefficient is stable in the green technology subsample. If the coefficient estimated on the full portfolio does not hold in the subsample, the framework is not transferable and should be re-specified for the green technology segment.

A worked calibration example makes the arithmetic concrete. Suppose a bank has 4,000 mid-sized corporate borrowers, of which 300 are classified as green technology enterprises. Over a five-year window, 80 borrowers default, 6 of them green technology firms. The financial-only model achieves an AUC of 0.74 on the holdout. The extended model achieves 0.77. The increment of 0.03 is modest but, with 4,000 observations and 80 events, is statistically significant at conventional levels (the standard error of an AUC difference of this size is roughly 0.012). The environmental component earns its place. If the increment had been 0.01, with a standard error of 0.012, the component would not have earned its place, and the honest response would be to report that the data do not support the extension in this portfolio.

The calibration of the notching thresholds follows the same logic. The thresholds \(E^{*} \geq 70\), \(E^{*} < 50\), \(T \geq -0.05\), \(T < -0.10\) are initial values. They should be chosen so that the borrowers who receive a downward notch have observed default rates consistent with the notched rating. If borrowers notched from BBB− to BB+ default at 2.5 percent over three years, and the bank's historical BB+ default rate is 1.8 percent, the notch is too shallow; the threshold should be tightened. If they default at 0.9 percent, the notch is too deep. This is a calibration loop, not a one-time exercise, and it requires the bank to track notched borrowers separately in its default database.

3.3.7 What the Framework Does Not Claim

Three limitations should be stated plainly.

First, the framework does not claim that environmental performance causes default. It claims that environmental performance is associated with default through identifiable transmission mechanisms, and that including it improves the rating's discrimination. Causation would require instruments that the data do not provide.

Second, the framework does not replace judgment. The notching rule is a decision aid, not a decision rule. A loan committee that sees a reason to override the notch should override it and document why. The framework's value is that it forces the committee to confront the environmental dimension explicitly rather than ignoring it.

Third, the framework is calibrated for green technology enterprises and may not transfer to other sectors. A cement company, a steel mill, and a software firm have different environmental risk profiles and different data availability. The indicator set and the weights should be re-estimated for each sector, and a framework applied across sectors without re-estimation will produce spurious precision.

3.3.8 Implementation: The Data Pipeline

The framework's operational feasibility depends on a data pipeline that connects the borrower's disclosures to the rating model. Figure 3.1 sketches the pipeline.

flowchart TD
    A[Borrower disclosure: fuel, electricity, water, waste, permits] --> B[Data validation: verification factor q]
    C[Third-party verification: emissions audit, permit records] --> B
    D[Sector benchmarks: 5th and 95th percentiles] --> E[Normalization to 0-100]
    B --> E
    E --> F[Weighted environmental score E]
    G[Policy scenarios: carbon price, subsidy schedule] --> H[Transition cost as % EBITDA]
    I[Resilience inputs: margin, pass-through, facilities] --> H
    F --> J[Data-quality adjustment: E* = E x q]
    H --> K[Transition adjustment T]
    J --> L[Notching rule]
    K --> L
    M[Financial scorecard: conventional rating] --> L
    L --> N[Final rating and PD mapping]

Figure 3.1: Data pipeline for the extended rating framework.

The pipeline has three entry points—borrower disclosure, third-party verification, and sector benchmarks—and two analytical branches that converge on the notching rule. The design keeps the environmental score and the transition adjustment separate until the final step, so that a committee can see which component drove the decision.

The cost of operating the pipeline is not trivial. A bank must maintain sector benchmarks, train analysts, and audit a sample of disclosures. For a large bank with a dedicated sustainability team, the marginal cost is manageable. For a small rural commercial bank, it is not. This is an argument for shared infrastructure: a credit rating agency, an industry association, or a regional green finance platform can maintain the benchmarks and provide the scores, and the bank can consume them. The framework is modular enough to allow that division of labor.

3.3.9 The Rating Agency Question

The framework described here is a bank's internal rating model. It is not a rating agency's published rating. The two serve different purposes and face different constraints.

A rating agency's rating is public, comparable across issuers, and used by investors who have no relationship with the issuer. It must be stable, transparentand defensible to a broad audience. A bank's internal rating is private, used for credit decisions and capital calculations, and can be more granular and more responsive to borrower-specific information.

The agencies have moved toward incorporating environmental factors, but cautiously. S&P, Moody's, and Fitch all publish ESG assessments alongside credit ratings, and all insist that the credit rating itself reflects environmental risk only where it is material to creditworthiness. Semet, Roncalli, and Stagnol (2021) find that ESG factors are partially priced in sovereign bond markets and by rating agencies, with variation across pillars and regions. The agencies' caution is warranted: a public rating that moves with an environmental score whose methodology is contested would import that contestation into the credit markets.

The division of labor that makes sense is this. Rating agencies provide the sector benchmarks and the environmental assessments that banks cannot produce themselves. Banks use those assessments, plus borrower-specific data, to compute an internal rating that reflects their own exposure and their own risk appetite. The two ratings need not agree, and their disagreement is informative: a borrower whose internal rating is lower than its public rating is telling the bank something the market has not yet priced.

3.3.10 The Chinese Context

China's green bond market reached approximately RMB 1.2 trillion in cumulative issuance by the end of 2020, ranking second globally. The pace has not been smooth. China Chengxin Credit Rating reports that green bond issuance in 2020 totaled RMB 222.161 billion, a year-on-year decline of 24.0 percent attributed to the COVID-19 pandemic. The decline is a reminder that green finance is not immune to the credit cycle, and that a rating framework for green borrowers must work in bad years as well as good.

The Chinese institutional context shapes the framework in three ways.

First, environmental data infrastructure is improving but uneven. The national emissions trading system, launched in 2021 for the power sector, created reporting obligations for covered entities. The 2021 Green Bond Endorsed Projects Catalogue provides a taxonomy for green revenue classification. But small and medium enterprises outside the covered sectors still report little. The framework's missing-data handling and verification factors are therefore essential, not optional.

Second, the banking system is dominated by large state-owned banks with the capacity to build internal models, alongside thousands of rural commercial banks and village banks without that capacity. The modular design in section 3.3.8 allows the large banks to build and the small banks to consume. A one-size-fits-all model imposed on both would fail at one end or the other.

Third, policy guidance is a major driver. The People's Bank of China's green finance evaluation scheme for banks, introduced in 2018 and revised since, assesses banks on their green credit share and their environmental risk management. That evaluation scheme is, in effect, an external environmental score applied to each bank's portfolio. The extended rating framework described here is the micro-level complement: the evaluation scheme tells a bank how its portfolio should look, and the framework tells the bank which borrowers belong in that portfolio and at what price.

The data problem in China has a specific shape that deserves detail. The Second National Pollution Source Census, published in 2020, provides facility-level data on pollutant discharge for roughly 3.6 million industrial sources, but the data are aggregated to the county level and lag by several years. A bank cannot use the census to rate an individual borrower. What the census does provide is a benchmark: a county-level discharge intensity that the bank can compare against a borrower's self-reported figures to detect implausible disclosures. A textile firm reporting water withdrawal far below the county average for its subsector is either unusually efficient or misreporting, and the rating model should flag the discrepancy rather than accept the number.

The emissions trading system creates a second data source. Covered entities must report verified emissions annually, and the verification is conducted by third-party institutions accredited by the Ministry of Ecology and Environment. For a bank lending to a covered entity, the verified emissions report is the highest-quality environmental data available, and it should receive the full verification factor of 1.0. For a borrower outside the covered sectors, the bank falls back on energy bills and permit records, with the lower verification factors described in section 3.2.6.

A third source is the green bond disclosure regime. Issuers of green bonds must report on the use of proceeds and the environmental impact of funded projects, and the reports are published. A bank can use these reports as a cross-check on a borrower's green revenue share. If a firm claims 70 percent green revenue but its green bond proceeds went to a general working capital facility, the claim deserves scrutiny.

The practical implication is that Chinese banks have more environmental data than they typically use, but the data are scattered across census records, ETS reports, bond disclosures, and permit databases that no single system integrates. The framework's data pipeline is, in part, an integration exercise. The bank that builds the integration first will have a rating advantage that is difficult to replicate, because the advantage lies in the data plumbing rather than in the model.

3.3.11 Interaction with Loan Pricing

The rating feeds pricing. A bank that assigns a borrower a lower rating must charge a wider spread or hold more capital, or both. The transmission from the environmental score to the loan price is therefore where the framework's incentive effect lives.

Consider two borrowers with identical financial scores, both rated A− on the conventional model. Borrower A has an environmental score of 72 and a transition adjustment of zero; borrower B has an environmental score of 38 and a transition adjustment of −0.14. Under the notching rule, A receives no adjustment and B is notched down by two grades to BBB. If the bank's pricing grid charges 90 basis points over the funding rate for an A− borrower and 140 basis points for a BBB borrower, the environmental difference costs B 50 basis points a year. On a RMB 200 million facility, that is RMB 1 million a year. The borrower now has a reason to invest in its environmental performance that is denominated in currency rather than in reputation.

Two cautions apply. The first is that the pricing differential must be credibly attributable to the environmental component, or borrowers will perceive it as arbitrary. The bank should disclose the notching rule and the score to the borrower, at least in summary form, so that the borrower can see what would change its rating. The second is that a borrower with a weak environmental score but a strong financial position may simply refinance with a bank that does not apply the framework. The incentive effect works only if the framework is applied widely enough that the borrower cannot escape it. This is an argument for industry-wide or regulator-endorsed adoption, which is beyond the scope of a single bank's model but is the natural trajectory.

The capital channel reinforces the pricing channel. Under the internal ratings-based approach of the Basel framework, a lower rating maps to a higher risk weight, which raises the capital the bank must hold against the exposure. For a corporate exposure at BBB, the risk weight is roughly 100 percent; at BB, it is roughly 120 percent. On a RMB 200 million facility, the difference in required capital at an 8 percent capital ratio is roughly RMB 3.2 million. The bank that prices the environmental notch correctly recovers both the expected loss and the capital charge. The bank that ignores it subsidizes the brown borrower at the expense of the green one, and the subsidy compounds over the life of the loan.

3.3.12 Portfolio Effects and the Aggregation Problem

A rating is assigned to a borrower. A portfolio is a collection of borrowers whose risks are correlated. The extended framework improves borrower-level discrimination, but it does not by itself solve the aggregation problem, which is the subject of Chapter 5. Two points are worth flagging here because they affect how the rating pillar should be designed.

First, environmental risk is correlated across borrowers within a sector and a region. A carbon price increase affects every steel mill in the province at once. A water withdrawal restriction affects every textile firm in the basin. If the bank's portfolio is concentrated in one sector or one region, the improvement in individual borrower ratings does not reduce portfolio risk; it may even increase it, because the bank is now more confident in its individual assessments and may lend more to the same concentrated group. The rating framework should be paired with a concentration limit that is set at the portfolio level, not derived from the sum of individual ratings.

Second, the transition adjustment is a common factor. If the bank applies the same carbon price scenario to every borrower, the resulting notches are correlated, and the portfolio's rating distribution shifts as a block when the scenario is updated. This is a feature, not a bug: it means the bank's portfolio rating responds to policy developments, which is what a risk-sensitive framework should do. But it means that the scenario assumptions must be reviewed centrally and consistently, not left to individual analysts, or the portfolio distribution will be an artifact of inconsistent assumptions.

The practical implication for the rating pillar is that the framework must produce two outputs: a borrower-level rating, used for credit decisions, and a set of scenario-conditional ratings, used for portfolio stress testing. The same environmental score can feed both, but the transition adjustment must be computed under a consistent scenario set for the portfolio to be aggregable.

The magnitude of the concentration effect is worth illustrating. Suppose a provincial bank has a loan book of RMB 50 billion, of which 30 percent is lent to manufacturing firms in three sectors: chemicals, textiles, and non-ferrous metals. All three sectors are carbon-intensive and water-intensive, and all three are concentrated in the same two river basins. A carbon price of RMB 100 per ton, applied to the portfolio, raises the expected annual cost across the three sectors by an estimated RMB 1.2 billion, or 8 percent of the sectors' combined EBITDA. If the bank's transition adjustment is applied borrower by borrower, each borrower may receive a notch of one or two grades, and the portfolio's average rating may decline by half a grade. But the correlated component of the risk—the part that materializes when the carbon price rises for everyone at once—is not captured by the average rating decline. It is captured by the concentration limit and by the stress test, which must be run on the portfolio as a whole rather than on the sum of individual borrowers.

A second illustration makes the point about scenario consistency. Suppose two analysts in the same bank apply different carbon price assumptions: one uses RMB 100 per ton, the other RMB 200. A borrower assessed by the first analyst receives a transition adjustment of −0.04; the same borrower assessed by the second receives −0.09. The first borrower stays at BBB−, the second is notched to BB+. The portfolio's rating distribution now depends on which analyst saw which borrower, not on the borrowers' actual exposure. This is why the scenario set must be set centrally and applied uniformly. The framework's credibility depends on the consistency of its inputs, and the inputs that vary most across analysts are the policy scenarios.

3.3.13 Governance and the Rating Committee

A rating framework is only as good as the process that applies it. The extended framework introduces new sources of discretion—indicator selection, weighting, scenario probabilities, resilience scoring—and each is an opportunity for the process to drift.

Three governance provisions reduce the drift.

The first is a documented methodology. Every weight, threshold, and scenario assumption should be written down, versioned, and dated. When the methodology changes, the change should be applied prospectively and the effect on the portfolio's rating distribution should be disclosed. A methodology that changes silently cannot be validated, because the validation sample no longer corresponds to the model.

The second is a separation between the analyst who computes the environmental score and the committee that assigns the rating. The analyst should present the score, the underlying data, and the data-quality flags. The committee should decide the notch. This separation mirrors the conventional rating process, where the analyst recommends and the committee decides, and it serves the same purpose: it prevents a single individual's judgment from becoming the rating.

The third is an override log. When the committee overrides the notching rule, it should record the reason. Overrides are informative: a pattern of overrides in one direction suggests that the rule is miscalibrated, and a pattern of overrides concentrated in one sector suggests that the sector benchmarks are wrong. The log turns overrides from an exception into a source of calibration data.

A fourth provision, less common but worth considering, is a periodic review of the indicator set itself. Environmental regulation changes, and an indicator that was material five years ago may be irrelevant now. The sulfur dioxide permit headroom that mattered for a coal-fired boiler in 2015 matters less for a firm that has converted to natural gas. The indicator set should be reviewed on a fixed cycle, with the review documented and the changes justified. Without the review, the framework accumulates indicators that no longer transmit to cash flow, and the weights become a historical artifact rather than a current assessment.

3.3.14 The Limits of the Extended Framework

I want to close this chapter by being explicit about what the framework can and cannot do, because the temptation to oversell a new methodology is strong and the cost of overselling is borne by the institutions that adopt it.

The framework cannot make a bad credit good. A firm with negative cash flow, an unsustainable capital structure, and a management team that has missed every forecast will not be rescued by a high environmental score. The environmental component is a modifier, not a substitute for financial analysis. The notching rule's asymmetry—one notch up at most, two notches down—is designed to enforce this.

The framework cannot resolve the measurement problem in markets where environmental data are absent. It can handle missing data with imputation and verification factors, but imputation is an assumption, and a portfolio built on assumptions is only as good as the assumptions. Banks operating in data-poor environments should treat the environmental score as a screening tool, not a pricing tool, until the data improve.

The framework cannot predict which technology will be displaced. It can flag firms whose products are on a published displacement roadmap, and it can measure R&D intensity as a proxy for adaptive capacity. It cannot tell the analyst whether perovskite tandem cells will commercialize in 2027 or 2032. The transition adjustment is a sensitivity analysis, not a forecast.

What the framework can do is make the environmental dimension of credit risk explicit, measurable, and auditable. It can force a loan committee to confront the question it would otherwise defer. It can price the difference between a firm that is positioned for a constrained-carbon economy and one that is not. And it can be validated, revised, and—if the evidence does not support it—narrowed. That is a modest claim. It is also the most that a rating methodology can honestly offer.

3.3.15 From Rating to the Next Pillar

The rating pillar identifies. It tells the bank which borrowers are positioned for the transition and which are exposed. It does not tell the bank what to do about the exposed ones. A borrower with a weak environmental score and a high transition risk may still be a good credit if the risk can be transferred, and that is the work of the insurance pillar.

The connection between the two pillars is direct. The transition adjustment \(T\) measures the firm's exposure to a policy or technology shock net of its resilience. Insurance raises \(R\), and the rating model reads the covered exposure at a lower value. A weather index insurance policy on an agricultural borrower, or a forest insurance policy on a forestry enterprise, converts an unhedged exposure into a covered one, and the covered exposure enters the rating model at a lower value. The framework should be built so that the purchase of insurance is visible to the rating model as a change in \(R\).

Chapter 6 works through the mechanics: how a weather index is constructed from meteorological station data, how the payout schedule is calibrated to the borrower's revenue loss, and how the insurance contract is represented in the rating model as a reduction in the expected transition cost. Before that, one more task remains within the rating pillar. The rating is an ordinal category. The credit decision requires a number. Chapter 4 takes up the estimation of default probability and loss given default from the rating, and shows where the provisional character of the environmental adjustment becomes most visible—in the mapping from a notched rating to a probability that the bank must defend to its regulator.

3.4 Conclusion: What the Rating Pillar Contributes

The electrolyzer manufacturer in Jiangsu eventually received its loan, at a spread 40 basis points wider than its financial scorecard alone would have justified. The additional spread was attributed to customer concentration, not to environmental performance, because the bank's framework at the time did not separate the two. Under the framework described here, the concentration risk would have entered the transition adjustment, the environmental score would have been computed and disclosed, and the committee would have seen which component drove the price. The borrower would have known what to fix. The committee would have known what it was pricing.

Whether the framework actually improves discrimination is an open question, and it will remain open until a bank runs the validation described in section 3.3.6 on a portfolio large enough to produce a statistically meaningful increment. The evidence reviewed in section 3.1.4 is suggestive, not conclusive. The measurement problems in section 3.2.6 are real, not hypothetical. A bank that adopts the framework on the strength of this chapter's argument, without running its own validation, is making a bet on the transmission mechanisms rather than on the data. That bet may be right. It should be tested.

The next step is concrete. A bank with a mid-sized corporate portfolio and at least five years of default history can run the three-stage calibration in section 3.3.6 within a single credit cycle. It needs the financial scorecard it already has, the environmental indicators it can obtain from ETS reports and permit records, and a holdout window. The output will be an AUC increment, a calibration table, and a decision: keep the environmental component, reweight it, or drop it. The framework is built to survive that decision, whichever way it goes.