Chapter 3: From Credit Scoring to Environmental Performance Rating

A loan officer at a rural commercial bank in Zhejiang receives two applications on the same morning. The first comes from a textile dyeing firm with twelve years of operating history, stable receivables, and a carbon emissions report that its own accountant assembled from utility bills and a consultant's template. The second comes from a precision-components manufacturer that has just installed a rooftop solar array, holds two invention patents on low-temperature plating, and has never borrowed more than RMB 3 million. The bank's scorecard, a logistic model built in 2016 and recalibrated twice since, returns scores of 612 and 588 respectively. The cutoff for unsecured lending is 600. The dyeing firm is approved; the components maker is declined, or routed to a collateral-backed product at a rate 180 basis points higher. Neither outcome reflects anything about environmental performance. The scorecard does not contain an environmental variable, and even if it did, the officer has no validated way to weight it.

Chapter 1 argued that green finance fails at the point of recognition—the lender cannot distinguish a genuinely transitioning firm from one that has learned the vocabulary. Chapter 2 traced that failure to information asymmetry and to the architecture of credit assessment, which rewards observable collateral and penalizes what it cannot verify. If the rating is the instrument through which creditworthiness is made legible, then the rating is where environmental performance must be made legible too. The question is how.

What follows is a construction problem: how to specify environmental indicators that are observable at reasonable cost, how to estimate their incremental predictive power for default and for transition outcomes, and how to combine them with financial factors in a way that survives validation. The chapter proceeds in three movements. The first examines what traditional credit scoring actually does and where its limits lie—not as a prelude to dismissing it, but to locate precisely the joints at which environmental information can be inserted. The second works through the incorporation of environmental metrics: the data, the transformations, the collinearity, and the timing. The third sets out a hybrid rating framework that treats financial and environmental factors as joint predictors of a single latent variable—the probability that a firm completes its transition without defaulting—and specifies how such a framework is estimated, validated, and governed.

3.1 Traditional Credit Scoring: What the Model Does and What It Cannot See

3.1.1 The scorecard as a compressed forecast

A credit score is a forecast compressed into a number. The forecast concerns a defined event—default within twelve months, or delinquency beyond ninety days—and the number is calibrated so that a given score maps to a known frequency of that event in a reference population. Everything else about the scorecard is machinery: the selection of variables, the transformation of each into a weight of evidence, the fitting of a logistic function, the setting of cutoffs, and the monitoring that detects drift. The machinery matters because it determines what the score can and cannot represent.

The canonical form is a logistic regression. Let \(y_i = 1\) if borrower \(i\) defaults within the horizon and \(0\) otherwise. The model asserts

where \(\mathbf{x}_i = (x_{i1}, \ldots, x_{iK})\) is the vector of borrower characteristics. In retail and small-business scoring the \(x\)'s are typically binned and re-expressed as weights of evidence,

so that each variable enters linearly with a coefficient that is usually constrained to be positive. The constraint is a governance device as much as a statistical one: a scorecard in which a higher debt-service ratio lowers the default probability is easier to defend to a supervisor than one in which the sign emerged from the data and happens to be perverse.

The variables that populate such scorecards are remarkably stable across jurisdictions and decades. They fall into five families. Financial ratios: leverage, liquidity, coverage, profitability, and the growth rates of each. Behavioral data: repayment history on prior facilities, utilization of revolving lines, and the seasoning of the relationship. Firmographics: age, sector, size, legal form, and region. Collateral and guarantee characteristics: type, haircut, and the identity of the guarantor. And, in consumer and micro-enterprise scoring, bureau attributes: the number of inquiries, the age of the oldest account, and the delinquency count across all lenders.

What is absent from this list is any variable that describes how the firm produces, what it emits, or how its production process is expected to change. That absence is not an oversight. It is a consequence of three design constraints that have governed scorecard development since the 1980s, and each constraint has a specific environmental cost.

3.1.2 Constraint one: the outcome must be observed at scale

A scorecard is estimated on a sample of borrowers whose outcomes are known. The binding requirement is that the outcome—default—be observed for a population large enough to estimate \(K+1\) parameters with acceptable standard errors, and that the population be homogeneous enough that a single model is defensible. In a mid-sized commercial bank, the small-business portfolio may contain 8,000 obligors with perhaps 240 defaults over a three-year window. That is adequate for a twelve-variable model and nothing more. Every variable added consumes degrees of freedom; every variable added that is available for only a subset of borrowers shrinks the estimation sample further.

Environmental variables are, almost by construction, available for a subset. A firm that participates in an emissions trading scheme has verified emissions data. A firm that does not has utility bills, if the lender can obtain them, or nothing. A firm with a green bond has a third-party verification report. A firm without one has a supplier questionnaire, if the lender's relationship manager completed it, which depends on the relationship manager's workload and training. The coverage rate for any environmental variable in a Chinese city commercial bank's small-business portfolio in 2021 was, in the samples I have examined, between 11 and 34 percent, depending on the variable and the sector. A scorecard cannot be estimated on a variable that is missing for two-thirds of the population without either imputing the missing values—which reintroduces the very information problem the variable was meant to solve—or restricting the model to the covered subset, which changes the population to which the score applies.

The environmental variable need not enter the default model directly. It can enter a first-stage model that predicts the transition outcome—whether the firm completes a defined environmental upgrade—and the predicted transition probability can then enter the default model as a covariate. This two-stage structure, which I develop in Section 3.3, allows the environmental data to be sparse at the level of the individual borrower while the transition model is estimated on a broader population of firms, including those that have not borrowed. The cost is that the first-stage model must be validated on its own terms, and its errors propagate into the second stage.

3.1.3 Constraint two: the variable must be stable within the horizon

A scorecard with a twelve-month horizon assumes that the borrower's characteristics at the observation date predict the outcome over the following year. Variables that fluctuate at higher frequency than the horizon introduce noise that dilutes their predictive power. Equity prices are the classic case: the Moody's Analytics EDF model, which derives a distance-to-default from market capitalization and book debt, produces estimates that move substantially day to day, and Moody's itself notes that "the equity market is volatile, so EDFs will be volatile as well" and that "some events that move the stock price of a firm may not be relevant from a credit perspective." The EDF is forward-looking and current, but its volatility exceeds what the fundamentals of the credit exposure merit, and experienced credit professionals must exercise judgment when using it.

Environmental performance has the same property, for different reasons. A firm's carbon intensity in a given year depends on its output mix, its energy source, and the weather. A hydroelectric-dependent manufacturer in Yunnan emits less in a wet year than in a dry one, not because it has transitioned but because the grid was cleaner. A firm that installs a waste-heat recovery unit records a step change in energy intensity in the installation year, followed by a slow drift as the equipment degrades. Neither the year-to-year noise nor the step change is well suited to a twelve-month default model. The noise averages out; the step change is a structural break that the model will misread as a level shift unless it is dated.

The practical implication is that environmental variables must be smoothed, and the smoothing window must be justified by the physics of the underlying process rather than by statistical convenience. A three-year rolling average of energy intensity is defensible for a manufacturer with a stable capital stock; a single-year reading is not. For firms undertaking a dated upgrade, the variable should be split into a pre-upgrade level and a post-upgrade indicator, so that the model can learn whether the upgrade itself—independent of the level—predicts survival.

3.1.4 Constraint three: the variable must be verifiable at low cost

The third constraint is economic. A scorecard variable earns its place only if the cost of collecting and verifying it is small relative to the improvement in discriminatory power it delivers. This is why bureau attributes dominate consumer scoring: they arrive electronically, at near-zero marginal cost, from a data intermediary that has already solved the verification problem. It is also why collateral characteristics appear in small-business scorecards in coarse form—property, equipment, receivables—rather than in the detail a valuer would produce. The scorecard does not need the value; it needs a variable correlated with the value at a cost the lender can bear.

Environmental data fails this test in its current form. Verified emissions under an emissions trading scheme are reliable but cover a minority of firms. Third-party green certification is reliable but expensive and, in the Chinese context, has been undermined by a period in which certification bodies issued certificates against weak criteria. Utility bills are cheap but measure electricity, not total energy, and reveal nothing about process emissions. Satellite-derived estimates of activity—nighttime lights, thermal anomalies, land-use change—are cheap and universal but measure the wrong unit: they observe the facility, not the firm, and the mapping from facility to obligor is often unknown to the lender.

The cost constraint is not immovable. It defines a research agenda, which Section 3.4 takes up: which environmental variables can be constructed from data the lender already holds or can obtain at the point of loan application, and how much predictive power do they add? The answer will differ by sector. For a cement producer, the relevant variable is clinker ratio and kiln technology, both of which are observable from the production permit. For a textile dyer, it is water discharge volume and treatment method, observable from the environmental impact assessment. For a software firm, there is no material environmental variable at the facility level, and the rating should say so rather than manufacture one.

3.1.5 What the scorecard sees, and what it does not

The three constraints produce a scorecard that is a good forecast of default within its reference population and a poor instrument for distinguishing transition trajectories. This is not a criticism of the scorecard's designers. It is a statement about what the instrument was built to do. A scorecard answers the question: given the financial and behavioral record of this borrower, how likely is repayment? It does not answer the question: is this borrower's business model viable under a carbon constraint that is tightening at a known rate?

The gap matters because the two questions diverge precisely where green finance operates. Consider the dyeing firm and the components maker from the opening scene. The dyeing firm's financial ratios are strong because its process is mature, its customers are locked in, and its capital stock is depreciated. Its environmental profile is poor: high water discharge, coal-fired boilers, no treatment upgrade. Under a scenario in which discharge standards tighten over five years, its cash flows are exposed to a compliance cost that the scorecard does not see. The components maker's ratios are weak because it is young and its receivables are concentrated in two customers. Its environmental profile is strong: electrified process, low-temperature chemistry, and a patent portfolio that positions it to supply the electrification supply chain. Under the same scenario, its cash flows are protected.

A scorecard calibrated on the historical default experience of the dyeing firm's sector will rank the dyeing firm above the components maker. It will do so correctly with respect to the past and incorrectly with respect to the transition. The error is not in the model's arithmetic; it is in the model's omission. The omitted variable is correlated with the outcome in the forward-looking sense that matters for a lender with a five-year exposure, and it is uncorrelated with the outcome in the backward-looking sense on which the model was estimated.

An omitted variable that is uncorrelated with the included variables biases nothing in the estimated coefficients; it simply leaves the model's predictions less accurate for the population in which the omitted variable varies. As the transition accelerates, the variance of the omitted variable increases, and the model's discriminatory power in the transition-relevant population declines. The decline is silent. The scorecard's backtests, run on historical defaults, will continue to show acceptable performance because the historical defaults were generated by an economy in which the omitted variable mattered less.

I am not arguing that every scorecard is about to fail, or that environmental variables will dominate financial ones. Financial ratios predict default because they measure the firm's capacity to service debt, and that capacity is real regardless of the firm's emissions. The narrower claim: environmental variables carry incremental information about default and transition that financial variables do not, the increment is larger in sectors exposed to transition policy, and the increment is currently excluded from the rating. The rest of this chapter is about how to include it without breaking the instrument.

flowchart TD
    A[Borrower application] --> B[Financial data: ratios, behavior, collateral]
    A --> C[Environmental data: energy, emissions, permits, patents]
    B --> D[Traditional scorecard]
    C --> E[Environmental performance model]
    D --> F[Financial credit score]
    E --> G[Transition probability]
    F --> H[Hybrid rating]
    G --> H
    H --> I[Decision: price, limit, covenant]
    I --> J[Realized outcomes: default, transition completion]
    J --> K[Validation and recalibration]
    K --> D
    K --> E

Figure 3.1: The hybrid rating inserts an environmental performance model between the raw environmental data and the final rating. Validation feeds back to both models separately.

3.2 Incorporating Environmental Metrics

3.2.1 Three families of environmental variables

The environmental information available to a lender falls into three families, distinguished by what they measure and by their failure modes. The first family measures emissions and resource use: Scope 1 and Scope 2 greenhouse gas emissions, energy consumption by source, water withdrawal and discharge, and waste generation. These are the variables that regulation increasingly requires firms to report, and they are the variables that enter the denominator of any intensity ratio. Their strength is that they are physical and, where verified, reliable. Their weakness is coverage and comparability: a firm that reports emissions on a different boundary from its peer cannot be compared to that peer without adjustment, and a firm that does not report cannot be placed on the scale at all.

The second family measures transition capacity: the firm's capital stock, its technology vintage, its patent portfolio in low-carbon technologies, its share of revenue from products that are compatible with a decarbonized economy, and its stated and budgeted transition plan. These variables are forward-looking in a way that emissions are not. A firm with a young, electrified capital stock has low transition risk even if its current emissions are high, because its emissions trajectory is already declining. A firm with a coal-fired boiler and fifteen years of remaining useful life has high transition risk even if its current emissions are within permit, because its trajectory is locked in. The strength of this family is that it captures the dynamics the rating needs. The weakness is measurement: capital stock vintage is observable from the depreciation schedule, but technology vintage is not; patent counts are observable but patent quality is not; transition plans are observable but their credibility is not.

The third family measures environmental governance and disclosure: whether the firm has an environmental management system, whether its emissions data are third-party verified, whether it has been subject to environmental penalties, and whether its board has assigned responsibility for environmental performance. These variables are institutional rather than physical. They predict not the firm's emissions but its capacity to manage them—and, critically, its capacity to report them accurately. The strength of this family is that it is cheap to observe and, in the Chinese context, increasingly available through public credit registers and environmental penalty databases. The weakness is that governance variables are gameable. A firm can adopt an environmental management system certificate without changing its process, and the rating that rewards the certificate will be rewarding the certificate.

The three families are complementary, and a rating that uses only one will be biased in a predictable direction. Emissions-only ratings penalize firms with high current emissions regardless of trajectory, which disadvantages precisely the heavy-industry firms that most need transition finance. Transition-capacity-only ratings reward plans regardless of current performance, which invites greenwashing. Governance-only ratings reward disclosure regardless of substance, which invites the certification arbitrage that has already occurred in the Chinese green bond market.

A defensible rating uses all three, with weights that reflect their incremental predictive power. The next three subsections take up the construction of each family, and the problems that arise when they are combined.

3.2.2 Emissions and intensity: the denominator problem

The standard measure of a firm's environmental performance is emissions intensity: emissions divided by output. The numerator is physical—tonnes of CO₂ equivalent—and the denominator is economic—revenue, value added, or physical output. The choice of denominator determines the ranking.

Consider two cement producers with identical clinker capacity. Firm A produces 2.0 million tonnes of cement annually at an emissions intensity of 0.85 tonnes CO₂ per tonne of cement. Firm B produces 1.2 million tonnes at 0.62 tonnes CO₂ per tonne. Measured per tonne of output, B is cleaner. Measured per tonne of clinker, the two may be identical, because B's advantage comes from blending more fly ash into its cement rather than from a cleaner kiln. Measured per unit of revenue, A may appear cleaner if its product mix commands a higher price. The three rankings are all defensible, and they disagree.

The resolution is to choose the denominator that matches the policy variable the rating is trying to anticipate. If the transition policy targets the production process—as the EU emissions trading scheme does, by allocating allowances to installations rather than to firms—then the correct denominator is physical output at the installation level, and the rating should be built at that level. If the policy targets the product—as building codes and product standards do—then the correct denominator is the emissions embedded in the product, and the rating should be built at the product level. If the policy targets the firm's overall footprint—as disclosure regimes do—then revenue is the relevant denominator, because it is the denominator on which investors and lenders compare firms.

In practice, a lender will not have installation-level data for most borrowers. The compromise that has emerged in the Chinese green bond market is to use revenue as the denominator and to adjust for sector. The adjustment is necessary because revenue intensity conflates two things: the emissions per unit of physical output, which is a technology characteristic, and the price per unit of output, which is a market characteristic. A firm with a clean process but a commodity product will have high revenue intensity; a firm with a dirty process but a branded product will have low revenue intensity. Ranking them by revenue intensity rewards the brand.

The adjustment I favor is a two-step normalization. First, express emissions per unit of physical output within the sector, using the physical denominator that the sector's regulators use. Second, express the firm's position relative to the sector distribution as a percentile, and use the percentile rather than the raw intensity as the rating input. The percentile is invariant to the choice of denominator within the sector, provided the same denominator is used for all firms in the comparison set. It is not invariant across sectors, which is why the sector must be defined narrowly—at the four-digit level, not the two-digit level—and why the rating must be recalibrated when the sector definition changes.

The percentile transformation has a second advantage. It converts a variable with a skewed, sector-specific distribution into a variable with a uniform distribution on \([0,1]\), which is well behaved in a logistic model and does not require the analyst to choose bin boundaries. The cost is that it discards information about the distance between firms at the extremes. A firm at the 99th percentile and a firm at the 95th are both "bad," but the first is much worse, and the percentile scale compresses the difference. For default prediction this compression is usually acceptable, because the relationship between intensity and default is flat at the extremes. For pricing, where the marginal effect matters, the raw intensity should be retained alongside the percentile.

3.2.3 Transition capacity: vintage, patents, and the credibility of plans

The second family of variables measures the firm's capacity to reduce emissions over the horizon of the loan. The most informative single variable is capital stock vintage. A firm's emissions are largely determined by the equipment it has installed, and equipment has a lifetime. A boiler installed in 2015 will operate until 2035; a boiler installed in 2005 will operate until 2025. The emissions trajectory is therefore embedded in the age distribution of the capital stock, and that distribution is observable from the depreciation schedule in the financial statements.

The construction is straightforward. For each asset class, compute the average age of the gross capital stock as the ratio of accumulated depreciation to annual depreciation expense. A firm with an average asset age of four years has a young stock; a firm with an average age of twelve years has an old one. The environmental interpretation depends on the technology: a young coal-fired boiler is a long-lived liability, while a young electric boiler is an asset. The vintage variable must therefore be interacted with the technology type, which is observable from the environmental impact assessment or the energy audit.

Patents are the second informative variable, with important caveats. A count of green patents—classified by IPC codes in the CPC Y02 family, which covers climate change mitigation technologies—measures inventive output. It does not measure the economic value of the invention, and patent counts are notoriously skewed: most patents are never cited, and a small fraction account for most of the value. The rating should therefore use a quality-weighted count, weighting each patent by its forward citations or by whether it has been maintained past the first renewal. Even quality-weighted, patents measure capacity, not deployment. A firm can hold a patent for a low-carbon process and never install it.

Transition plans are the third variable and the least reliable. A plan is a statement of intent, and intent is cheap. The rating should not reward the existence of a plan; it should reward the plan's credibility, which can be assessed along three dimensions: whether it is costed (a plan that specifies the capital expenditure and the financing source is more credible than one that does not), whether it is dated (a plan with milestones is more credible than one with a target year), and whether it is monitored (a plan whose progress is reported to the board and to the lender is more credible than one that is filed and forgotten). These three dimensions can be scored from the plan document by a trained analyst in under an hour, which is within the cost budget of a relationship-based lender but not of a transaction-based one.

3.2.4 Governance and disclosure: the penalty record as a verifiable signal

The third family is the cheapest to observe and the most susceptible to gaming. Environmental management system certification, disclosure quality scores, and board-level responsibility assignments are all variables that a firm can adopt without changing its process. The rating that rewards them is rewarding the adoption, not the substance.

There is one governance variable that is not gameable in this way: the environmental penalty record. A penalty is issued by a regulator after an inspection or a complaint, and it is recorded in a public database. It cannot be self-reported, and it cannot be withdrawn. In China, the environmental penalty records maintained by the Ministry of Ecology and Environment and by provincial counterparts are increasingly linked to the credit reference system, and a lender can query them at the point of application. The variable is sparse—most firms have no penalty in any given year—but when it is present it is highly informative. A firm with a penalty in the past three years has demonstrated both that its process is non-compliant and that its management did not detect the non-compliance before the regulator did.

The penalty variable should enter the rating in a specific form: as a count of penalties weighted by severity, with a decay factor so that older penalties count less. The decay factor should be estimated, not assumed. A penalty for an exceeded discharge limit is a leading indicator of future penalties; a penalty for a paperwork violation is not. The rating should distinguish them, which requires the lender to code the penalty type, which in turn requires the penalty database to expose the type. In practice this means the lender builds a small internal taxonomy and maps the regulator's codes onto it. The mapping is tedious and it is the kind of work that gets deferred. It should not be, because the penalty variable is the single most cost-effective environmental input available to a lender that does not have verified emissions data.

3.2.5 The collinearity problem

The three families are correlated with each other and with the financial variables. Firms with young capital stock tend to have lower emissions intensity, higher profitability, and lower leverage. Firms with environmental penalties tend to have weaker governance, which shows up in the financial variables as well. If all three families are entered into a single logistic regression alongside the financial ratios, the coefficients will be unstable and the standard errors will be large. The model will fit the estimation sample and fail out of sample.

Orthogonalization is the first remedy: regress each environmental variable on the financial variables and use the residual as the rating input. The residual is the part of environmental performance that is not explained by financial condition, which is precisely the incremental information the rating needs. The approach is clean and it fails when the relationship between the environmental and financial variables is non-linear, which it usually is. A firm at the bottom of the emissions distribution has a different financial profile from a firm in the middle, and a linear projection will not capture the difference.

A second remedy combines the environmental variables into a single index before entering them, using principal components or a fixed-weight composite. The index absorbs the collinearity within the environmental block, and the single index coefficient is stable. The cost is interpretability: a composite index of emissions, vintage, patents, and penalties has no natural unit, and the lender cannot explain to the borrower why its score changed. For a rating that is used in pricing and in regulatory reporting, interpretability matters.

The third remedy, and the one I prefer, estimates the environmental effect in a separate model and uses the predicted value as a covariate in the default model. This is the two-stage structure introduced earlier. The first stage predicts the transition outcome—whether the firm completes a defined upgrade within a defined window—using the environmental variables and a set of controls. The second stage predicts default using the financial variables and the predicted transition probability. The collinearity between the environmental and financial variables is handled in the first stage, where it affects the precision of the transition model but not the stability of the default model. The default model sees a single covariate, the transition probability, whose coefficient is estimated cleanly.

The preference for the third approach rests on more than statistics. It forces the lender to specify what "transition" means for each sector, and to date it. A transition is not a vague improvement in environmental performance; it is a specific change in process or product, completed by a specific date, at a specific cost. Specifying it makes the rating auditable. It also makes the rating useful: a firm that has completed its transition is a different credit risk from a firm that has not, and the two-stage model says so explicitly.

3.2.6 Timing and the horizon mismatch

The final construction problem is timing. A default model has a twelve-month horizon. A transition has a five-to-fifteen-year horizon. The environmental variables that predict transition are measured at the observation date, and their effect on default is realized over a period longer than the model's horizon. If the environmental variables are entered directly, the model will estimate a coefficient that reflects only the part of the transition effect that materializes within twelve months, which is small. The model will conclude that environmental variables do not matter, and it will be wrong about the longer horizon that the lender actually cares about.

The two-stage structure resolves this, provided the first stage is estimated on the correct horizon. The transition model should be estimated on a five-year window, using the environmental variables at the start of the window to predict completion at the end. The predicted transition probability is then a summary of the firm's five-year trajectory, and its coefficient in the twelve-month default model captures the part of that trajectory that affects near-term default—through compliance costs, through access to financing, through customer demand—as well as the part that affects the firm's long-run viability.

The horizon mismatch has a second implication. The default model's coefficients on the financial variables are estimated on a population that includes firms at different points in their transition. If the transition is accelerating, the historical relationship between leverage and default will understate the default risk of highly leveraged firms with poor transition prospects, because the historical sample contains too few such firms in the stressed state. The model should be re-estimated on a rolling window, and the window should be short enough that the transition is visible but long enough that the default count is adequate. Three to five years is the practical range.

3.3 A Hybrid Rating Framework

3.3.1 The latent variable

The framework I propose treats the rating as a forecast of a single latent variable: the probability that the firm services its debt through the loan's tenor without defaulting. Call this \(p_i^{\text{surv}}\). The forecast is built from two blocks of information, financial and environmental, and the two blocks are combined through the transition probability rather than averaged.

The structure is

where \(\Lambda(z) = 1/(1+e^{-z})\), \(\mathbf{f}_i\) is the vector of financial and behavioral variables that would appear in a traditional scorecard, \(\mathbf{e}_i\) is the vector of environmental variables, and \(\hat{\tau}_i(\mathbf{e}_i)\) is the predicted probability that firm \(i\) completes its sector-defined transition within the transition window. The coefficient \(\delta\) measures the effect of transition prospects on near-term default, conditional on financial condition. The financial coefficients \(\boldsymbol{\gamma}\) are estimated jointly with \(\delta\), so the model does not impose the traditional scorecard's weights; it re-estimates them in the presence of the transition covariate.

The first-stage transition model is

where \(\mathbf{z}_i\) contains sector and regional controls. The model is estimated on a sample that includes both borrowers and non-borrowers, because the transition outcome is observable for both—a firm's completion of an upgrade does not depend on whether it has a loan. This is the key advantage of the two-stage structure: the environmental data can be sparse in the lending portfolio while the transition model is estimated on a broader population.

There is a selection issue. The firms that borrow are not a random sample of the firms in the transition model's estimation population. If lenders select on financial condition, and financial condition is correlated with transition capacity, then the transition probabilities that enter the default model are estimated on a population that differs from the lending portfolio. The standard correction is to include the inverse Mills ratio from a selection equation, but the correction requires an exclusion restriction—a variable that affects the borrowing decision but not the transition outcome—which is hard to justify in this context. A simpler approach is to estimate the transition model on the sector population and to check, by comparing the distribution of \(\hat{\tau}\) in the portfolio to its distribution in the population, whether selection is severe. If it is, the model should be re-estimated on the portfolio with the missing-data problem handled explicitly.

3.3.2 Defining the transition outcome

The framework's credibility depends on the transition outcome being well defined. "Transition" is not a single event. For a cement producer, it might be the installation of a waste-heat recovery system, the substitution of clinker with fly ash beyond a threshold, or the commissioning of a carbon capture pilot. For a textile dyer, it might be the connection to a centralized wastewater treatment plant, the replacement of coal boilers with natural gas, or the adoption of waterless dyeing. For a components manufacturer, it might be the commissioning of an on-site renewable generation system or the certification of a product as compatible with a downstream decarbonization standard.

The outcome must be defined at the sector level, by the lender's analysts working with sector engineers, and it must satisfy three criteria. It must be observable: the completion of the upgrade must be verifiable from a document the lender can obtain, such as a completion certificate, an updated pollution discharge permit, or a grid connection agreement. It must be material: the upgrade must change the firm's emissions intensity or its product's compatibility by a threshold that the sector analysis identifies as significant. And it must be dated: the upgrade must be completed within a window, typically five years, that matches the lender's exposure.

The definition is a judgment, and different lenders will define it differently. That is acceptable, provided the definition is stated, applied consistently, and revisited as the technology and the policy change. What is not acceptable is a definition that is so broad that every firm qualifies—"has an environmental policy"—or so narrow that only the largest firms can meet it. The definition should be calibrated so that, in the lender's portfolio, between 15 and 40 percent of firms in the exposed sectors are classified as having completed a transition within the window. A lower rate means the definition is too narrow; a higher rate means it is too easy.

flowchart LR
    A[Sector definition] --> B[Transition outcome: observable, material, dated]
    B --> C[Transition model: environmental variables + controls]
    C --> D[Predicted transition probability]
    D --> E[Default model: financial variables + transition probability]
    E --> F[Hybrid rating: score, grade, price]
    F --> G[Monitoring: transition completion, default]
    G --> H[Recalibration]
    H --> C
    H --> E

Figure 3.2: The transition outcome is defined at the sector level and drives both the transition model and the default model.

3.3.3 Estimating the environmental block

The environmental block \(\mathbf{e}_i\) should contain no more than six variables, for the reasons given in Section 3.1.2. My preferred specification, for an industrial firm with available data, is:

1. Emissions intensity percentile within the four-digit sector, computed on a three-year rolling average.
2. Capital stock vintage, measured as average asset age, interacted with a technology indicator (fossil versus electric).
3. Quality-weighted green patent count per unit of revenue, logged.
4. Environmental penalty count in the past three years, weighted by severity and decayed.
5. Transition plan credibility score, on a three-point scale (costed, dated, monitored).
6. A disclosure indicator: whether the firm's emissions data are third-party verified.

The six variables cover the three families: emissions (1), transition capacity (2, 3, 5), and governance (4, 6). They are not equally available. Variable 1 requires emissions data; variable 2 requires only the financial statements and the energy audit; variable 3 requires a patent database; variable 4 requires the penalty database; variable 5 requires an analyst's judgment; variable 6 requires a verification certificate. A lender that has only variables 2 and 4 can still estimate a transition model, with lower explanatory power. A lender that has all six should use all six, and should report the incremental \(R^2\) of each so that the cost of the missing variables is visible.

The functional form should be logistic, with the variables entered in the transformed form described above. The model should be estimated with sector fixed effects, because the baseline transition rate differs across sectors for reasons that have nothing to do with the firm—a sector with a generous subsidy will have a higher completion rate than one without. The fixed effects absorb the baseline, and the coefficients on the environmental variables measure the within-sector effect.

3.3.4 The coefficient on transition probability

The coefficient \(\delta\) is the parameter that carries the framework's central claim: that transition prospects affect default, conditional on financial condition. If \(\delta\) is zero, the environmental block adds nothing to the default forecast, and the traditional scorecard is adequate. If \(\delta\) is negative and significant, transition prospects reduce default probability, and the hybrid rating is an improvement.

The sign is not obvious ex ante. A firm with strong transition prospects may be a firm that is investing heavily, which consumes cash and increases near-term default risk. The net effect depends on whether the investment is funded from internal cash flow or from new debt, and on whether the transition reduces compliance costs enough to offset the investment burden. The framework does not assume the sign; it estimates it, and the estimate is the empirical content of the chapter.

My prior, from the pilot work described below, is that \(\delta\) is negative but modest in the full sample and larger in the exposed sectors. The mechanism is not that green firms are inherently safer. It is that firms with a credible transition path face a smaller variance of future compliance costs, and smaller variance of future cash flows translates into a lower probability of hitting the default boundary. The effect should be strongest for firms with long-dated fixed-rate debt, where the lender'sexposure spans the transition window, and weakest for short-dated working capital facilities, where the exposure is repaid before the transition matters.

A second channel complicates interpretation. Firms with strong transition prospects may have better access to subsidized green credit, which lowers their funding cost and improves their survival probability for reasons unrelated to their environmental performance per se. If the model is estimated on a population that includes subsidized borrowers, \(\delta\) will conflate the effect of transition prospects with the effect of the subsidy. The clean approach is to include a subsidy indicator in the default model and to report \(\delta\) with and without it. The difference between the two estimates is the size of the subsidy channel, which is itself a quantity of policy interest.

3.3.5 From score to grade

The default model produces a probability, not a grade. Grades are useful for pricing, for limit-setting, and for regulatory reporting, and they must be constructed so that each grade corresponds to a bounded range of default probabilities. The standard approach is to sort the portfolio by predicted probability and to choose cut points that produce a monotonic relationship between grade and realized default rate, with enough defaults in each grade to estimate the rate precisely.

The hybrid rating adds a complication: the grade should reflect both the default probability and the transition status, because a lender may want to treat a firm with a high default probability but a strong transition path differently from a firm with the same default probability and no transition path. The clean way to handle this is to produce two dimensions—a default grade and a transition grade—and to combine them in the pricing and limit rules rather than in the rating itself. The default grade is the output of the default model. The transition grade is the output of the transition model, bucketed into, say, four categories. The pricing matrix maps the pair to a spread.

This two-dimensional presentation has a governance advantage. It makes explicit that the rating is not a single number, and it forces the lender to state how it trades off the two dimensions. A lender that wants to prioritize transition finance can set the pricing matrix so that a strong transition grade offsets a weak default grade; a lender that wants to prioritize asset quality can set it the other way. The choice is a business decision, not a statistical one, and it should be made by the credit committee with the matrix in front of it.

3.3.6 Validation

A rating that is not validated is an opinion. The hybrid rating must be validated on three dimensions, and the validation must be designed before the model is estimated, not after.

Discrimination is tested on a holdout sample: does the model rank-order defaults better than the traditional scorecard? The standard measure is the area under the receiver operating characteristic curve, or AUC. The comparison that matters is not the hybrid model against a random ranking; it is the hybrid model against the traditional scorecard estimated on the same sample. The incremental AUC is the quantity of interest, and it should be reported with a confidence interval. In the pilot described below, the incremental AUC was 0.021 in the full sample and 0.047 in the exposed sectors, with the latter significant at conventional levels and the former marginal.

Calibration asks a narrower question. Does the predicted default rate match the realized default rate within each grade? The test is a Hosmer-Lemeshow statistic or, for small samples, a binomial test within each grade. A model can discriminate well and calibrate poorly, and a poorly calibrated model will misprice. The hybrid model is more exposed to calibration failure than the traditional scorecard because the transition probability is estimated on a different population, and any selection between the two populations will show up as a calibration error. The validation should therefore include a comparison of the predicted and realized default rates in the portfolio, by transition grade.

Stability is the dimension that takes longest to establish and is easiest to skip. Do the coefficients and the discriminatory power hold up out of time? The model should be estimated on a training window and tested on a subsequent window, with the windows chosen so that the test window includes a policy change or a macroeconomic shock. A model that survives a policy change is more credible than one that has only been tested in a stable environment. The relevant policy change for a Chinese lender is the tightening of environmental enforcement that began in 2016 and accelerated through the subsequent five-year plans. A model estimated on data before 2016 and tested after 2016 is a genuine out-of-time test. The pilot's stability results, reported in Section 3.5.3, are weaker than its discrimination results, and I say so there rather than here.

There is a fourth dimension that is specific to this framework and that is usually omitted: the transition model must be validated on its own terms. Does the predicted transition probability correlate with realized transition completion? The test is a simple one—sort firms by predicted probability, and compare the completion rate in the top decile to the rate in the bottom decile. If the model has no power to predict completion, then the default model's \(\delta\) is estimating the effect of noise, and the framework collapses. The transition model's AUC should be reported alongside the default model's, and a transition AUC below 0.65 should be treated as a failure of the environmental block, not as a failure of the framework.

3.4 The Data Problem, Sector by Sector

3.4.1 Why sector matters more than firm

The environmental variables that predict transition are sector-specific in a way that financial variables are not. A leverage ratio means the same thing in cement and in software. An emissions intensity ratio does not: the cement producer's ratio is measured against a distribution of cement producers, and the software firm has no material emissions to measure. A rating framework that applies the same environmental variables across sectors will either misclassify the software firm—assigning it a transition probability based on variables that do not apply—or exclude it, which is the same as saying the framework does not cover the service economy.

The resolution is to define the environmental block at the sector level, with a common core and sector-specific extensions. The common core is the governance variables—penalties and disclosure—which apply to every firm. The extensions are the emissions and transition-capacity variables, which apply only to sectors with material environmental exposure. The framework should classify sectors into three tiers: exposed, transitional, and neutral.

Exposed sectors are those in which the transition policy binds directly: power generation, cement, steel, chemicals, non-ferrous metals, pulp and paper, and textiles dyeing. These sectors account for the majority of industrial emissions and are the target of the emissions trading scheme, the discharge standards, and the energy-intensity targets. The environmental block should be fully specified for these sectors, and the transition outcome should be defined narrowly.

Transitional sectors are those in which the transition affects the firm through its customers or its supply chain rather than directly: electronics, machinery, automotive components, and construction materials. The firm's own emissions may be modest, but its products must comply with downstream standards, and its customers are increasingly asking for embodied-carbon data. The environmental block for these sectors should emphasize product compatibility and disclosure rather than process emissions.

Neutral sectors are those in which environmental performance is not material to credit risk: software, business services, and most of retail. The environmental block should contain only the governance variables, and the transition outcome should be defined in terms of the firm's own operations—energy sourcing, for example—rather than its products. For many neutral-sector firms, the honest answer is that the environmental block adds nothing, and the rating should be the traditional scorecard.

The tiering is a judgment, and it will be contested at the boundaries. A logistics firm is transitional if its fleet is diesel and neutral if it is electric. A data center is neutral in its own operations and exposed through its electricity consumption. The boundaries should be set by the lender's sector analysts, reviewed annually, and documented so that the classification can be audited.

3.4.2 What is observable in the Chinese context

The data environment for Chinese lenders has improved substantially since 2015, but it remains uneven. Table 3.1 summarizes the availability of the six environmental variables across the three tiers, based on the sources a mid-sized commercial bank can access without building a proprietary data platform. The availability thresholds are drawn from a survey of 14 city and rural commercial banks conducted by the author in 2022; each bank reported, for a randomly selected sample of 100 small-business obligors, whether the variable could be obtained from documents already in the loan file or from a public registry within two working days. "Available" means obtainable for more than 80 percent of firms in the tier at that cost; "partial" means 30 to 80 percent; "not material" means the variable does not predict the outcome.

Table 3.1: Availability of environmental variables by sector tier. Source: author's survey of 14 city and rural commercial banks, 2022, n = 1,400 obligors.

The table's message is that the exposed sectors are the best served, which is fortunate because they are where the environmental block matters most. The transitional sectors are the weak point: the variables that matter—product compatibility, embodied carbon—are the least available, because they require data from the firm's customers, which the lender does not have. The neutral sectors are well served by the governance variables, which are cheap and universal.

The capital stock vintage variable deserves a note. It is available for every firm that files financial statements, which in China means every firm with a bank loan. The computation requires the depreciation schedule, which is in the notes to the accounts, and the technology indicator, which is in the energy audit or the environmental impact assessment. For a firm that has neither, the technology can be inferred from the sector and the firm's age, with a loss of precision that the model should acknowledge by widening the confidence interval on the vintage coefficient.

3.4.3 The cost of data collection

The cost constraint from Section 3.1.4 is not a fixed barrier; it is a budget that the lender allocates. A relationship manager can complete a six-variable environmental questionnaire in about forty minutes, provided the questionnaire is well designed and the firm has the documents at hand. The documents—the energy audit, the pollution discharge permit, the environmental impact assessment—are usually in the firm's files because they were required for the firm's own compliance. The marginal cost to the lender is the relationship manager's time, which is already being spent on the credit application.

The cost that is not marginal is the cost of building the databases: the penalty register query, the patent database query, and the ETS registry query. These require system integration, which requires IT investment. Table 3.2 gives a line-item breakdown for a mid-sized bank with roughly 40,000 small-business obligors, based on quotations obtained from two system integrators in 2022 and on the bank's own estimates of internal staff time.

Table 3.2: Line-item cost of building the environmental data infrastructure. Source: quotations from two system integrators and internal bank estimates, 2022.

The total is a real cost, and it must be justified by the improvement in the portfolio's risk-adjusted return. The pilot evidence suggests that the improvement is concentrated in the exposed sectors, which are also the sectors where the loan book is largest. The business case is therefore stronger than it looks at first. A bank with RMB 8 billion in exposed-sector lending and a 2 percent expected loss would need to reduce expected loss by roughly RMB 8 million a year to recover the annual cost, which corresponds to a 10 basis point reduction in the average loss rate—well within the range the pilot's incremental AUC implies.

There is a cheaper path, which is to buy the data from a third-party provider. Several Chinese credit rating agencies and data vendors now offer environmental profiles of firms, assembled from public sources. The profiles are not verified, and their coverage of small firms is incomplete, but they are adequate for the transitional and neutral tiers. The lender that buys the data and builds the integrations only for the exposed sectors will spend less and get most of the benefit.

3.4.4 The missing-data problem

Even with the integrations, the environmental block will be missing for a substantial fraction of the portfolio. The missingness is not random: it is concentrated in small firms, in firms in regions with weak regulatory capacity, and in firms that have never been inspected. A model that drops the missing observations will be estimated on a sample that over-represents large, well-regulated firms, and its predictions for small firms will be unreliable.

The standard remedies are three. Complete-case analysis is the simplest and the most biased. Single imputation—replacing the missing value with the sector median—is less biased but understates the variance. Multiple imputation, which draws the missing values from a conditional distribution and repeats the estimation across draws, is the most defensible and the most computationally demanding. For a rating model estimated annually, the computational cost is trivial, and I recommend multiple imputation with at least twenty draws.

The imputation model should include the financial variables and the observed environmental variables, and it should be estimated separately by sector tier. The imputation should be flagged: the rating output should carry an indicator of whether the environmental block was complete, partially imputed, or fully imputed, and the pricing matrix should treat the three cases differently. A firm whose transition probability is based on imputed data should not receive the same green pricing as a firm whose data are complete, because the imputation's uncertainty is not reflected in the point estimate.

3.5 The Pilot: Two Firms, One Model

3.5.1 Design

The framework was tested on a pilot sample drawn from a commercial bank's small and medium enterprise portfolio in Zhejiang and Jiangsu, covering the period from 2016 to 2021. The sample contains 4,312 firms, of which 612 defaulted or entered a workout during the observation window. The sector composition is 38 percent textile and apparel, 22 percent machinery and equipment, 14 percent chemicals, 11 percent electronics, and 15 percent other. The transition outcome was defined by the bank's sector analysts as the completion, within five years of the observation date, of a documented process or product upgrade that reduced the firm's emissions intensity by at least 20 percent or certified its product as compatible with a downstream low-carbon standard.

The environmental block contains five of the six variables; third-party verification was dropped because it was available for fewer than 10 percent of the sample. The default model contains the bank's standard financial variables—leverage, current ratio, interest coverage, return on assets, revenue growth, firm age, and collateral type—plus the predicted transition probability. The models were estimated on the 2016–2018 cohort (n = 2,706, 361 defaults) and tested on the 2019–2021 cohort (n = 1,606, 251 defaults).

3.5.2 The two firms

The aggregate results are less instructive than the individual cases, so consider the two firms from the opening scene, now with their environmental profiles.

The dyeing firm (call it Firm A) has an emissions intensity at the 78th percentile of its four-digit sector, an average asset age of 11.4 years, a coal-fired boiler installed in 2009, no green patents, one environmental penalty in 2018 for exceeding a chemical oxygen demand limit, and a transition plan that is dated but not costed. Its transition probability from the first-stage model is 0.19. The components maker (Firm B) has an emissions intensity at the 31st percentile, an average asset age of 3.8 years, an electric process, two green patents in low-temperature plating, no penalties, and a transition plan that is costed, dated, and monitored. Its transition probability is 0.71.

In the traditional scorecard, Firm A scores 612 and Firm B scores 588, as in the opening scene. In the hybrid model, the financial variables contribute a log-odds of \(-1.84\) for A and \(-2.21\) for B, and the transition probability contributes \(-0.19 \times 0.19 = -0.036\) for A and \(-0.19 \times 0.71 = -0.135\) for B. The total log-odds are \(-1.88\) for A and \(-2.35\) for B, corresponding to one-year default probabilities of 13.2 percent and 8.7 percent. The ranking reverses.

The reversal is the framework's point, and it is worth being precise about what it does and does not show. It does not show that Firm B is a better credit in every respect; its receivables concentration is a real risk that the financial block captures and that the transition block does not offset. It shows that the transition information, which the traditional scorecard omits, is large enough to change the ranking for a pair of firms whose financial scores are close. For firms whose financial scores are far apart, the transition information will not reverse the ranking, and it should not: a firm in severe financial distress will default before its transition prospects can matter.

3.5.3 Aggregate results

Across the pilot sample, the hybrid model's AUC on the test cohort was 0.741 (standard error 0.014), against 0.720 (standard error 0.015) for the traditional scorecard estimated on the same sample. The incremental AUC of 0.021 carries a bootstrap standard error of 0.011 and a 95 percent confidence interval of \([-0.001, 0.043]\); it is marginal, not significant at the 5 percent level. The decomposition by sector is more informative, and it is also where the sample-size caveat bites. In the exposed sectors—textiles dyeing and chemicals, n = 612 in the test cohort with 118 defaults—the incremental AUC was 0.047 (bootstrap standard error 0.019, 95 percent CI \([0.010, 0.084]\)). In the transitional sectors—machinery and electronics, n = 531 with 74 defaults—it was 0.018 (standard error 0.022), indistinguishable from zero. In the neutral sectors, n = 463 with 59 defaults, it was 0.004 (standard error 0.025). The exposed-sector estimate rests on 118 defaults, which is enough to reject zero but not enough to pin the magnitude within a factor of two; a second institution's data would be needed to do that.

The coefficient on transition probability was \(-0.19\) with a standard error of 0.07 in the full sample, and \(-0.31\) with a standard error of 0.11 in the exposed sectors. Both are significant at the 5 percent level. The coefficient on the subsidy indicator, when included, was \(-0.22\) (standard error 0.09), and its inclusion reduced the transition coefficient to \(-0.14\) (standard error 0.08), which suggests that roughly a quarter of the estimated transition effect operates through access to subsidized credit. That decomposition is rough, because the subsidy indicator is endogenous, but it is the right order of magnitude and it should be reported.

The calibration was acceptable in the exposed and transitional sectors and poor in the neutral sectors, where the model over-predicted default by about 1.5 percentage points in the lowest transition grade (Hosmer-Lemeshow \(p = 0.03\)). The over-prediction is consistent with the transition model having no power in the neutral sectors: the predicted transition probabilities there are essentially noise, and the default model's coefficient on noise is estimated with error. The remedy is to restrict the hybrid model to the exposed and transitional tiers and to use the traditional scorecard for the neutral tier. That is a less elegant framework, but it is honest about where the environmental block adds information.

Stability was tested by re-estimating the model on the 2016–2018 cohort and applying it to 2019–2021, which spans the COVID-19 shock. The exposed-sector incremental AUC fell from 0.047 in the pooled test to 0.031 when the 2020 vintage was excluded, which suggests that part of the apparent gain in the exposed sectors came from the model's ability to identify firms vulnerable to the shock for reasons correlated with their environmental profile. The stability result is weaker than the discrimination result, and it is the main reason I describe the pilot as a proof of concept rather than a validated model.

3.5.4 What the pilot does not show

The pilot is small, single-institution, and concentrated in two provinces. Its results should be read as a proof of concept, not as a calibration that another lender can adopt without re-estimation. Three limitations deserve emphasis.

First, the transition outcome was defined by the bank's analysts, and a different definition would produce different transition probabilities and a different \(\delta\). The sensitivity of the results to the definition is not known, because the pilot did not estimate the model under alternative definitions. A proper validation would.

Second, the pilot's test window, 2019–2021, includes the COVID-19 shock, which affected the exposed sectors disproportionately. The incremental AUC in the exposed sectors may partly reflect the model's ability to identify firms that were vulnerable to the shock for reasons correlated with their environmental profile—for example, firms with old capital stock and weak balance sheets—rather than a genuine transition effect. The distinction matters for the framework's interpretation, and it cannot be resolved with a single test window.

Third, the pilot did not include a cost-benefit analysis. The data collection cost was borne by the research project, not by the bank, and the bank's marginal cost of the environmental block in production would be different. The business case remains to be made.

3.6 Governance: Who Owns the Rating

3.6.1 The separation of model and judgment

A rating that combines financial and environmental information raises a governance question that the traditional scorecard does not: who decides the weights? In a traditional scorecard, the weights are estimated from data and approved by the model risk committee. In a hybrid rating, the environmental block contains analyst-assessed variables—the transition plan credibility score, the sector tier assignment, the transition outcome definition—that are not estimated from data in the same way. The boundary between the estimated and the judged must be drawn explicitly, and the two must be governed differently.

The estimated components—the coefficients on the financial variables, the transition probability, and the environmental variables in the first stage—should be governed by the model risk process: independent validation, annual recalibration, and backtesting. The judged components—the sector tier, the transition outcome definition, the plan credibility score—should be governed by a separate process: a sector review committee that includes credit analysts and, where possible, external technical experts, with documented decisions and an annual review. The two processes should meet at the rating committee, which approves the final framework and the pricing matrix.

The separation matters because the judged components are where the greenwashing risk enters. An analyst who is sympathetic to a borrower can score its transition plan as credible when it is not. A sector committee that is captured by an industry can define the transition outcome so loosely that every firm qualifies. The governance must make these judgments visible and contestable: the plan credibility scores should be distributed and compared across analysts, and the sector definitions should be published so that they can be challenged.

3.6.2 The audit trail

Every rating should carry an audit trail that records, for each environmental variable, the source document, the date of collection, the analyst who collected it, and the transformation applied. The trail serves three purposes. It allows the validation team to reproduce the rating. It allows the regulator to inspect the basis of a green classification. And it allows the lender to defend a decision if a borrower disputes it.

The trail is not free. A rating with a complete trail takes perhaps 25 percent longer to produce than one without. That cost is justified for the exposed sectors, where the environmental block is material and where the green classification carries pricing consequences. For the neutral sectors, a lighter trail—source and date only—is adequate.

The trail should also record the version of the sector definition and the version of the penalty taxonomy in force at the rating date. Without those version tags, a rating produced in 2024 cannot be compared to one produced in 2026 after the definitions have been revised, and the validation sample becomes a mixture of incompatible regimes. The version tags are cheap to add and expensive to retrofit.

3.6.3 The regulatory interface

The hybrid rating sits at the intersection of two regulatory regimes that were designed separately. The prudential regime governs credit risk models: the internal ratings-based approach under Basel, and the Chinese implementation through the China Banking and Insurance Regulatory Commission's capital rules. The green finance regime governs the classification of green assets: the People's Bank of China's green bond endorsed project catalogue, the green credit statistics, and the EU taxonomy for lenders with cross-border exposure.

The two regimes have different purposes, and a rating that serves both will be pulled in two directions. The prudential regime cares about default prediction and wants the model to be validated on default data. The green regime cares about the environmental classification and wants the model to be transparent about the environmental criteria. A single model can serve both if the environmental block is reported separately from the default model, so that the green classification can be audited without reopening the default calibration. That is the architecture I have proposed: two models, two outputs, one rating.

The regulatory interface also constrains the data the lender can use. The prudential regime requires that model inputs be verifiable and that the model be stable over time. The green regime is evolving, and its criteria will change as the taxonomy is revised. A lender that ties its environmental block too tightly to the current taxonomy will have to rebuild the block when the taxonomy changes. The block should therefore be built on physical variables—emissions, vintage, patents, penalties—that are stable across taxonomy revisions, and the mapping from the block to the taxonomy should be a separate layer that can be updated without re-estimating the model.

3.7 Objections

3.7.1 "Environmental performance is already priced"

The first objection is that the market has already incorporated environmental information into credit spreads, so the rating adds nothing. The evidence against this is that green bond spreads are small. In my own tabulation of 214 Chinese green bonds issued between 2019 and 2022 and matched to 428 conventional bonds from the same issuers, the average yield spread of green over conventional at issuance was 2.1 basis points, with a standard error of 0.8 and a 95 percent confidence interval of \([0.5, 3.7]\) basis points; after controlling for tenor, rating, and issuer fixed effects, the differential fell to 1.4 basis points and was no longer significant. If the market were pricing transition risk, the spread differential between exposed and neutral sectors would be larger than it is. The market is pricing the label, not the risk. A rating that measures the risk is therefore adding information, not duplicating it.

3.7.2 "The data are too poor"

The second objection is that environmental data are too unreliable to be used in a rating that affects pricing. The objection has force for the emissions variable, which is self-reported for most firms and verified for few. It has less force for the vintage, patent, and penalty variables, which come from administrative sources that the firm cannot manipulate. The framework should therefore weight the variables by their verifiability, and it should report the rating's sensitivity to the emissions variable. A rating that changes materially when the emissions variable is dropped is a rating that depends on unverified data, and the lender should know that.

3.7.3 "This is a subsidy in disguise"

The third objection is that the hybrid rating is a mechanism for directing credit to favored firms under the cover of risk measurement. The objection is serious, and it cannot be fully answered by statistics. The answer is procedural: the framework's weights are estimated from data, the transition outcome is defined by a committee that includes external experts, and the pricing matrix is approved by the credit committee with the trade-off made explicit. A lender that wants to subsidize green firms should say so and use a subsidy instrument; a lender that wants to measure risk should use the rating. The framework can serve both purposes only if the two are kept separate, which is why the subsidy indicator should be reported alongside the transition coefficient rather than folded into it.

3.7.4 "The transition outcome is unknowable"

The fourth objection is that the transition is too uncertain for a five-year outcome to be forecast. The objection is correct in the aggregate and overstated at the firm level. The path of policy is uncertain; the direction is not. A firm's capacity to reduce its emissions intensity is a function of its capital stock, its technology, and its management, all of which are observable. The transition model does not need to forecast the policy; it needs to forecast whether the firm will complete an upgrade that is already economically attractive under a range of policy scenarios. That is a more tractable problem, and the pilot's transition AUC of 0.69 (standard error 0.03) suggests it is tractable in practice.

3.8 From Rating to Pricing

3.8.1 The pricing matrix

The rating's output is a pair of grades: a default grade and a transition grade. The pricing matrix maps the pair to a spread over the funding cost. The matrix should be set so that the spread covers the expected loss, the unexpected loss capital charge, and the operating cost, with the transition grade shifting the spread within a band that reflects the lender's strategic preference.

The expected loss for a facility is

where \(\mathrm{PD}_i\) is the one-year default probability from the hybrid model, \(\mathrm{LGD}_i\) is the loss given default, and \(\mathrm{EAD}_i\) is the exposure at default. The transition grade affects \(\mathrm{PD}_i\) through the model and affects \(\mathrm{LGD}_i\) through the collateral: a firm that has completed a transition has newer, more valuable collateral, which lowers \(\mathrm{LGD}\). The matrix should reflect both channels.

The strategic band is the lender's choice. A lender that wants to grow its green portfolio can set the band so that a strong transition grade reduces the spread by up to 30 basis points; a lender that wants to protect margin can set it at 10. The band should be disclosed to the regulator and reviewed annually, because it is a policy variable, not a market variable.

3.8.2 The limit and covenant interface

The rating also drives limits and covenants. A firm with a weak transition grade but a strong default grade should receive a limit that is sized to its default risk but a covenant that requires progress on the transition—for example, a commitment to complete a specified upgrade by a specified date, with the limit reduced if the commitment is missed. The covenant is the instrument through which the rating's forward-looking information is converted into a contractual obligation.

The covenant design should follow the same logic as the transition outcome definition: observable, material, dated. A covenant that requires the firm to "improve environmental performance" is unenforceable. A covenant that requires the firm to connect to the centralized wastewater treatment plant by the end of the second year, evidenced by a connection agreement, is enforceable. The lender's legal team should draft the covenant templates by sector, and the credit committee should approve them alongside the pricing matrix.

3.8.3 The feedback loop

The final element of the framework is the feedback loop. The rating is not a one-time assessment; it is updated as the firm's transition progresses and as the portfolio's default experience accumulates. The update should be triggered by three events: the annual financial review, the completion or abandonment of a transition milestone, and the accumulation of enough new defaults to warrant recalibration.

The feedback loop is where the framework earns its keep. A rating that is estimated once and never updated will drift as the transition accelerates, and its discriminatory power will decline silently. A rating that is updated annually, with the transition model re-estimated on the expanded sample of completed transitions, will improve as the data accumulate. The improvement is the return on the data collection cost, and it should be tracked against a specific threshold: the transition model's AUC should rise by at least 0.02 within three recalibration cycles; if it does not, the environmental block is not learning and should be re-specified.

sequenceDiagram
    participant RM as Relationship manager
    participant DM as Data platform
    participant TM as Transition model
    participant CM as Default model
    participant CC as Credit committee
    RM->>DM: Submit application and environmental documents
    DM->>DM: Query penalty, patent, ETS registries
    DM->>TM: Environmental block (with imputation flags)
    TM->>CM: Transition probability and grade
    DM->>CM: Financial block
    CM->>CM: Default model with transition covariate
    CM->>RM: Rating pair, price, limit, covenants
    RM->>DM: Monitoring: milestone completion, default
    DM->>TM: Recalibration data
    DM->>CM: Recalibration data

Figure 3.3: The rating workflow. The transition model and the default model are estimated separately and meet at the credit committee.

3.9 The Unresolved Question

The pilot's largest weakness is not its sample size or its single-institution design. It is that the transition outcome was defined once, by one group of analysts, and the model was never re-estimated under an alternative definition. If the bank's sector committee had set the emissions-reduction threshold at 30 percent instead of 20 percent, or had counted a fuel switch as a transition only when it was accompanied by a change in the pollution discharge permit, the set of firms classified as having completed a transition would have changed, the first-stage coefficients would have changed, and \(\delta\) would have changed. How much it would have changed is unknown. That sensitivity is the first thing a lender adopting this framework should measure, because a rating whose central coefficient moves with an arbitrary definitional choice is not a rating a credit committee can defend to a supervisor.

The second open item is the stability result. The exposed-sector incremental AUC fell from 0.047 to 0.031 when the 2020 vintage was excluded, which means part of the apparent gain came from the model's ability to sort firms by COVID-19 vulnerability rather than by transition prospects. A test window that contains no macroeconomic shock would settle the question, and no such window exists in the pilot data. A lender with a longer history, or a lender in a jurisdiction where the transition policy tightened without a simultaneous demand shock, could run the test the pilot could not.

Both items point to the same next step: the framework has to be applied to a sector where the transition runs through product innovation rather than process upgrade, because that is where the definitional problem is hardest and where the financial profile of the borrower is least like the small manufacturer the pilot was built on. Technology firms fit that description. Their environmental exposure is indirect, their collateral is intangible, their cash flow is negative for years, and their transition path is a sequence of product generations rather than a dated capital investment. Whether the two-stage structure survives that translation is the question Chapter 4 takes up.