Chapter 5: Rating-Based Identification: Incorporating Environmental Performance into Technology Firm Credit Assessment

A commercial bank in Chongqing was asked in 2019 to price a loan to a lithium-battery recycler and had no procedure for the question. The firm's balance sheet looked ordinary: 180 million yuan in revenue, a debt-to-asset ratio of 61 percent, three consecutive years of positive operating cash flow. Its default risk, on the bank's internal scorecard, came out at BB. The firm's process chemistry, however, recovered 92 percent of the cobalt and 88 percent of the nickel from spent cathodes, against a national industry average near 70 percent for cobalt and much lower for nickel. That recovery rate was the firm's entire competitive advantage, and it appeared nowhere in the scorecard. The loan officer priced the facility at the BB rate. The firm borrowed elsewhere. The bank's credit committee later recorded the case as a lost opportunity without ever knowing why.

This chapter is about closing that gap. The argument is narrow. Existing credit rating models can be extended to identify green technology firms without abandoning their statistical foundations, provided environmental performance is treated as a structured input rather than a narrative overlay. I will show where traditional ratings fail, how environmental indicators can be converted into rating-grade information, and what an implementable framework looks like. The framework is diagnostic, not promotional: it will misclassify some firms, and I will say where.

5.1 Traditional Credit Rating and Its Gaps

5.1.1 What a Credit Rating Actually Measures

A credit rating is a forecast of the probability that a borrower will fail to meet a contractual payment. Everything else is instrumentation. The expected loss on a loan can be written as

where \(PD\) is the probability of default over the horizon, \(LGD\) the loss given default, and \(EAD\) the exposure at default. A rating model's job is the first term. The other two are matters of collateral, seniority, and structure.

This decomposition matters because it disciplines what environmental performance can and cannot do in a rating. Environmental performance can affect \(PD\) through cash-flow volatility, regulatory exposure, input-cost risk, and litigation. It can affect \(LGD\) through asset specificity: a coal-fired plant is worth less on a distressed sale than a solar farm with a power purchase agreement. It does not affect \(EAD\), which is a function of facility design. A framework that claims environmental scoring improves all three is overreaching.

The major agencies use a hybrid of statistical models and analyst judgment. S&P Global Ratings' corporate criteria framework combines business risk (industry risk, competitive position) and financial risk (cash flow, leverage, liquidity) into an anchor rating, then applies modifiers for diversification, capital structure, financial policy, liquidity, and management. Fitch's approach is similar in structure, with a heavier weight on through-the-cycle analysis. Both agencies publish transition and default studies; Fitch's global corporate finance average one-year transition rates and five-year transition rates are the empirical backbone of the whole enterprise. The rating is only as good as the stability of the mapping from rating category to default frequency.

That stability is contested. Nickell, Perraudin and Varotto (2000), studying 6,534 issuers from December 1970 to December 1997, found that transition matrices are not stable over time when economic cycles are ignored. Bangia, Diebold, Kronimus, Schagen and Schuermann (2002) observed that the longer the horizon over which a transition matrix is estimated, the less likely it is to change, which is another way of saying short-horizon matrices are non-stationary. The agencies' response is the through-the-cycle rating: a rating that looks past the current point in the cycle to a long-run average creditworthiness. This is a deliberate smoothing choice, and it has a cost. A through-the-cycle rating is slow to reflect structural change. A firm whose business model is being dismantled by a technology shift will carry an investment-grade rating for quarters after the shift is visible to everyone in the industry.

Green transition is precisely a structural change of this kind. That is the first gap.

5.1.2 The Four Gaps

I will name four limitations of traditional rating models when applied to green technology firms. They are distinct, and they require distinct fixes.

Gap one: environmental risk enters as a cost, never as a capability. In the standard treatment, environmental factors appear as contingent liabilities: a pollution fine, a remediation obligation, a carbon price. The 1980s US Superfund legislation, which made firms liable for pollution they caused and prompted banks to guard against environmental credit risk, established this framing. It is a defensive framing. It asks how much a firm might have to pay because of its environmental footprint. It does not ask whether the firm's environmental performance is a source of margin. For a recycler with 92 percent cobalt recovery, the answer to the second question is yes, and the standard model cannot see it.

Gap two: the data are not commensurable. A rating model needs inputs that can be compared across firms and mapped to default frequency. Environmental data arrive in incompatible units: tons of CO2, cubic meters of water, kilograms of hazardous waste, hectares of land. Converting them into a single score requires an aggregation rule, and every aggregation rule embeds a value judgment. Traditional models avoid the problem by excluding the data. That avoidance is itself a value judgment, and a costly one.

Gap three: the through-the-cycle assumption breaks. A through-the-cycle rating assumes the firm's industry has a stable long-run default rate. Green technology sectors do not. Solar manufacturing had a brutal consolidation between 2010 and 2013, with module prices falling roughly 80 percent and a long list of bankruptcies across Europe and China. Lithium-ion battery manufacturing is going through a capacity cycle now. Electric vehicle supply chains are being redrawn by subsidy changes in three major markets. The long-run average default rate for these sectors, estimated over any window that includes a consolidation, is a poor forecast for the next three years.

Gap four: the rating scale is too coarse for the decision. A bank's internal scorecard typically has 12 to 20 grades. Two firms in the same grade are treated as equivalent for pricing. For green technology firms, the within-grade dispersion is large. Two battery recyclers with the same leverage and the same revenue can differ by a factor of two in gross margin because one recovers nickel and the other does not. The rating grade averages them together and prices them identically. The bank then loses the good one to a competitor who can see the difference, or wins the bad one because no one else will lend to it. This is adverse selection produced by rating granularity.

The four gaps are not equally tractable. The first and fourth can be addressed by adding structured inputs and increasing resolution. The second requires a defensible aggregation methodology. The third is the hardest: it is a problem of statistical estimation under structural change, and no framework fully solves it. I will return to it.

5.1.3 Why "Add an ESG Score" Does Not Work

The obvious fix is to bolt an ESG score onto the existing model. Many banks have tried. The results are usually weak, and it is worth being precise about why.

An ESG score is a composite of environmental, social, and governance indicators, often produced by a third-party data provider. It is designed for portfolio screening and investor reporting, not for default prediction. Three problems follow.

First, the score's construction is opaque and its components are weighted by judgment. A firm can improve its score by improving its governance disclosure without changing its environmental performance at all. When the score enters a rating model, it imports noise.

Second, the score is not calibrated to default. A one-point improvement in an ESG score does not correspond to any known reduction in \(PD\). Without calibration, the model cannot price the improvement. The bank knows the firm is "greener" but not how much less likely it is to default.

Third, the score is contaminated by size and disclosure capacity. Large firms disclose more and score better. In a sample of technology firms, where many issuers are small and privately held, the ESG score becomes a proxy for firm size, and firm size is already in the model. The environmental signal is collinear with an existing variable and adds nothing.

The alternative is not to abandon environmental information. It is to select environmental indicators that have a plausible causal link to cash flow, measure them in physical units where possible, and calibrate their relationship to default or distress using available data. This is more work than buying a score. It is also the only version that survives a credit committee's questions.

5.1.4 The Chinese Context: Green Credit Policy Without Green Rating

China's green credit framework illustrates the gap in institutional form. The regulatory definition is explicit: green credit means financial institutions identify, measure, monitor, and control environmental and social risks in credit activities, build an environmental and social risk management system, improve credit policies and procedures, strictly prevent credit funds from flowing into polluting industries, and direct them toward the low-carbon economy, the circular economy, and the ecological economy. The definition is about process and direction. It says nothing about how a rating model should treat a firm whose environmental performance is its competitive advantage.

In practice, Chinese banks have implemented green credit through three mechanisms: a list-based exclusion (no lending to firms on the environmental protection blacklist), a catalog-based inclusion (preferential treatment for projects on the green industry catalog), and a quota-based target (a share of the loan book classified as green). None of these requires a rating model to change. A firm is either on the list or not. The loan officer's discretion fills the space between.

That discretion has a cost. As one industry assessment put it, commercial banks look at the development trends of green industries "through fog and at the moon in water"—they lack a scientific and clear understanding, and as a result their prevention of potential risks is weak. The metaphor is apt. A loan officer who cannot distinguish a good battery recycler from a bad one will either lend to both or lend to neither. The first outcome accumulates risk; the second starves the sector.

The rating framework in this chapter is an attempt to replace discretion with measurement, at least for the subset of technology firms where environmental performance is material to cash flow.

5.1.5 What Counts as a Green Technology Firm

Before proceeding, a definition. I will use "green technology firm" to mean a firm whose principal revenue comes from products or services that reduce environmental impact per unit of output relative to the incumbent technology, and whose environmental performance is a material driver of gross margin. The second clause is doing the work. A solar module manufacturer qualifies. A software company that has installed solar panels on its office roof does not, even though its own operations are low-carbon. The distinction is between environmental performance as a business and environmental performance as a footprint.

This definition excludes some firms that green finance catalogs include, and includes some that they do not. A firm that manufactures scrubbers for coal plants qualifies under the first clause (it reduces sulfur dioxide emissions per unit of output) but its fate is tied to the coal fleet, which is a declining market. The framework must be able to rate it low for reasons that have nothing to do with its environmental performance. That is fine. The rating is a forecast of default, not a sustainability award.

5.2 Integrating Environmental Performance: Methodology

5.2.1 The Causal Chain Requirement

An environmental indicator belongs in a credit rating only if there is a causal chain from the indicator to cash flow. The chain need not be short, but each link must be articulable. Consider four examples.

Energy intensity. A manufacturer's energy consumption per unit of output determines its exposure to electricity price changes. In a province where industrial electricity tariffs rose 12 percent in one year, a firm with energy intensity twice the industry median saw its gross margin fall by 4.1 percentage points; a firm at the median saw a 2.0 point fall. The causal chain is: energy intensity → cost sensitivity → margin volatility → \(PD\). The chain is short and measurable.

Material recovery rate. A recycler's recovery rate of a valuable metal determines the spread between its input cost (scrap) and its output value (refined metal). At a cobalt price of $35,000 per ton, a recovery rate of 92 percent versus 70 percent is worth roughly $7,700 per ton of contained cobalt, which for a 5,000-ton-per-year facility is about $38 million of annual gross margin. The chain is: recovery rate → unit margin → cash flow → \(PD\). It is short and measurable, and it is the indicator that the Chongqing bank did not have.

Carbon intensity of the product. A cement producer's emissions per ton of clinker determine its exposure to a carbon price. If the carbon price is $50 per ton and the firm emits 0.8 tons of CO2 per ton of cement against a benchmark of 0.6, the firm faces an annual compliance cost of $10 per ton of cement, which at a 10 percent EBITDA margin is the entire margin. The chain is: carbon intensity → regulatory cost → margin → \(PD\). The strength of the chain depends on the credibility of the carbon price path, which is a policy forecast, not a physical measurement. This is a weaker link.

Environmental litigation history. A firm with a record of environmental violations faces a higher probability of fines and shutdowns. The chain is: violation history → enforcement probability → cash-flow shock → \(PD\). The link is real but the magnitude is hard to estimate, and the indicator is backward-looking.

The four examples are ordered by the strength of the causal chain. Energy intensity and material recovery rate are physical, measurable, and tightly linked to margin. Carbon intensity is physical and measurable but its link to margin runs through a policy variable. Litigation history is measurable but its link to future cash flow is weak and noisy.

A workable methodology uses the strong indicators as core inputs and the weak indicators as adjustments. It does not weight them equally. Equal weighting is the most common error in ESG integration, and it is the error that makes the resulting score uninterpretable.

5.2.2 Physical Units First, Scores Second

The second methodological principle is to collect environmental data in physical units and convert to scores only at the last step. The reason is that physical units preserve the causal chain. A recovery rate of 92 percent is a fact about a process. A "resource efficiency score" of 78 out of 100 is a fact about a scoring rubric. When the credit committee asks why the firm's rating improved, the first answer is defensible and the second is not.

The conversion from physical units to a score requires a benchmark. There are three options, and they are not equivalent.

Peer benchmark. The firm's indicator is compared to the median of its industry peer group. This is the most common approach and the easiest to implement. Its weakness is that a peer group can be uniformly bad. If every battery recycler in a province recovers 70 percent of cobalt, a firm at 72 percent looks excellent and is, in absolute terms, mediocre. Peer benchmarking measures relative performance, and relative performance can be misleading when the whole industry is behind.

Technical benchmark. The firm's indicator is compared to the best available technology, or to a physically derived standard. A recovery rate of 92 percent is compared to a theoretical maximum of, say, 96 percent. This benchmark is absolute and does not drift with the industry. Its weakness is that best-available technology changes, and the benchmark must be updated. It also penalizes firms in industries where the gap between average and best is structurally large.

Regulatory benchmark. The firm's indicator is compared to a regulatory threshold, such as a carbon intensity benchmark under an emissions trading system. This benchmark has the advantage of being tied to a real cost: a firm above the benchmark pays, a firm below it receives an allocation. Its weakness is that the threshold is a political variable and can be revised.

My preference is a two-tier structure. Use the technical benchmark for the core physical indicators, because it is stable and absolute. Use the regulatory benchmark for indicators that are directly tied to a compliance cost, because the cost is what matters for cash flow. Use the peer benchmark only as a tie-breaker or for indicators where no technical standard exists. This is more demanding than a single peer benchmark, and it requires the bank to maintain a small library of technical standards. That library is the bank's intellectual capital, and it is what distinguishes a rating framework from a checklist.

5.2.3 The Aggregation Problem

Suppose a firm has five environmental indicators, each converted to a score between 0 and 100. How do they combine into a single environmental performance score?

The naive answer is a weighted average. The weights are usually set by expert judgment, and the resulting score is presented as if it were objective. It is not. A weighted average assumes the indicators are substitutes: a firm can compensate for poor performance on one with good performance on another. That assumption is often wrong. A battery recycler with excellent energy efficiency and a poor recovery rate is not a good credit; it is a firm that has optimized the wrong variable. The indicators are complements, not substitutes.

A better aggregation for complementary indicators is a minimum or a lexicographic rule. The environmental performance score is the minimum of the indicator scores, or the score on the indicator designated as binding. This is harsh, and it will produce low scores for firms that are good on four dimensions and poor on one. That harshness is appropriate when the poor dimension is the one that drives default. A recycler with a poor recovery rate is a firm whose margin will disappear when metal prices fall, regardless of how efficient its furnaces are.

A practical compromise is a "weakest link with partial compensation" rule: the composite score is a weighted average, but any indicator below a threshold sets a cap on the composite. If the recovery rate score is below 40, the composite cannot exceed 60 regardless of the other indicators. This preserves some substitutability while ensuring that a critical weakness is not averaged away.

The choice of aggregation rule is a credit judgment, not a mathematical one. It should be documented and reviewed, and it should be tested against default data. If the minimum rule produces a better separation of defaulters from non-defaulters than the weighted average, use the minimum rule. The data will tell you, provided you have enough defaults to test with. Most banks do not, at least not in the green technology segment. In that case, the rule should be conservative: when in doubt, let the weakest indicator bind.

5.2.4 From Environmental Score to Probability of Default

The environmental score is not a rating. It is an input to a rating. The step that converts it into a probability of default is the calibration step, and it is where most frameworks fail.

There are three calibration strategies, ordered by data intensity.

Direct calibration. If the bank has a historical sample of green technology loans with both environmental scores and default outcomes, it can estimate the relationship directly. This is the ideal. It is also rare. The segment is young, defaults are few, and the sample is small. A bank with 200 green technology loans and 6 defaults cannot estimate a reliable relationship between a five-indicator score and default. The standard errors will be larger than the effects.

Proxy calibration. If direct data are unavailable, the bank can use a proxy outcome that is observed more frequently than default. Persistent negative operating cash flow, covenant breach, rating downgrade, and payment delinquency are all more common than default and are correlated with it. The bank estimates the relationship between the environmental score and the proxy, then maps the proxy to default using the existing rating model. This is a two-step procedure and it compounds estimation error, but it is feasible with a few hundred observations.

Expert-parameterized calibration. If neither is available, the bank assigns the mapping by judgment, informed by the causal chain. A firm with a recovery rate in the top decile of its industry is assigned a one-notch uplift; a firm in the bottom decile a one-notch downgrade. The parameters are documented and revisited annually. This is the least rigorous option and the most common in practice. It is defensible only if the bank is explicit that the parameters are judgments and commits to replacing them with estimates as data accumulate.

I will describe a framework that uses the third strategy as a starting point and specifies the data collection needed to move to the second and eventually the first. The framework is a scaffold, not a finished building.

5.2.5 The Double Materiality Question

There is a prior question that the methodology must answer: whose environmental performance is being measured? The firm's own operations, or the environmental impact of its products?

The distinction is called double materiality. Financial materiality asks how environmental factors affect the firm's cash flow. Environmental materiality asks how the firm's activities affect the environment. A rating model is concerned with the first. A green finance policy is often concerned with the second. The two can point in opposite directions.

A solar module manufacturer has low operational emissions and a product that displaces coal-fired electricity. On financial materiality, its environmental performance is good. On environmental materiality, its manufacturing process uses hazardous chemicals and its supply chain includes polysilicon production that is energy-intensive. The firm scores well on one dimension and poorly on the other.

For credit rating purposes, financial materiality is the relevant lens. The question is whether the firm's environmental performance affects its ability to repay. A firm with a hazardous chemical problem faces regulatory and litigation risk that can impair repayment. A firm with an energy-intensive supply chain faces input-cost risk. Both are financial materiality, even though they are about environmental harm.

This does not mean environmental materiality is irrelevant. It enters through the regulatory channel: a firm whose products cause environmental harm is more likely to face future regulation, which is a cash-flow risk. The rating model should capture that risk through the policy-exposure indicator, not through a separate environmental harm score. Keeping the two lenses separate and explicit prevents the model from becoming a sustainability rating in disguise.

5.3 A Proposed Rating Framework

5.3.1 Architecture

The framework has four layers. Figure 5.1 shows the structure.

flowchart TD
    A[Financial inputs: leverage, coverage, liquidity, cash-flow volatility] --> D[Anchor rating]
    B[Business inputs: industry risk, competitive position, technology maturity] --> D
    C[Environmental inputs: E1 energy intensity, E2 material efficiency, E3 carbon intensity, E4 policy exposure, E5 environmental compliance] --> E[Environmental performance score]
    E --> F[Environmental adjustment]
    D --> G[Adjusted rating]
    F --> G
    G --> H[Overlay: management, governance, structure]
    H --> I[Final rating]
    I --> J[PD mapping and pricing]

The anchor rating is produced by the existing model. The environmental adjustment modifies it by a bounded number of notches. The overlay applies the standard qualitative adjustments. The final rating maps to a probability of default through the bank's existing calibration table.

The bounded adjustment is important. If the environmental adjustment can move a rating by an unlimited number of notches, the framework becomes an environmental rating with a financial appendix. If it can move by at most one notch, the framework is a financial rating with a small environmental correction. The right bound depends on how much of the firm's cash-flow risk is environmental. For a battery recycler, the recovery rate is the dominant driver of margin, and a one-notch bound is too tight. For a general manufacturer, a one-notch bound is appropriate.

I propose a two-notch maximum adjustment in either direction, with the actual adjustment determined by the environmental score and the materiality of environmental factors to the firm's cash flow. The materiality weight is the share of gross margin variance explained by environmental variables, estimated from the firm's cost structure. A firm where energy and materials are 70 percent of cost gets full weight; a firm where they are 15 percent gets a fraction.

5.3.2 The Five Core Indicators

The framework uses five environmental indicators. Each is chosen because it has a causal chain to cash flow. Each is measured in physical units. Each is converted to a score against a technical or regulatory benchmark.

E1: Energy intensity. Energy consumption per unit of physical output, in gigajoules per ton or kilowatt-hours per unit. The benchmark is the best available technology for the firm's product category. The causal chain runs through cost sensitivity to electricity and fuel prices. The indicator is available from utility bills and production records; no third-party data is required. This is the most reliable indicator in the set.

E2: Material efficiency. For manufacturing firms, yield (output over input). For recycling firms, recovery rate of the target material. The benchmark is the technical maximum or the best-performing peer. The causal chain runs through unit cost and, for recyclers, through the spread between scrap input and refined output. This is the indicator that the Chongqing case was missing, and it is the most important one for the recycling segment.

E3: Carbon intensity. Emissions per unit of output, in tons of CO2 equivalent per ton or per unit. The benchmark is the regulatory allocation benchmark where one exists, otherwise the sector's best available technology. The causal chain runs through carbon price exposure. Where no carbon price exists, the indicator is a policy forecast and should be weighted accordingly. In China, the national emissions trading system covers the power sector and is expanding to other sectors; for firms outside the covered sectors, the carbon intensity indicator is a leading indicator of future compliance cost, not a current one.

E4: Policy exposure. A measure of the firm's dependence on subsidies, tax preferences, or regulatory support. This is not a physical indicator, and it is the weakest of the five. It is included because green technology firms are unusually policy-dependent, and a rating that ignores subsidy dependence will systematically underprice the risk of subsidy removal. The indicator can be measured as the share of gross margin attributable to subsidies or favorable regulatory treatment. A firm with 40 percent of margin from a feed-in tariff is more exposed than one with 10 percent. The causal chain is direct: subsidy removal → margin loss → \(PD\).

E5: Environmental compliance. A count of violations, fines, and enforcement actions over the past five years, normalized by firm size. The benchmark is zero for a clean record. The causal chain runs through fines, shutdown risk, and management attention. The indicator is backward-looking and noisy, and it should carry a low weight. Its main value is as a red flag: a firm with repeated violations has a management problem that the financial statements may not reveal.

Table 5.1 summarizes the indicators, their units, benchmarks, and proposed weights for a manufacturing or recycling firm.

Table 5.1: Core environmental indicators for credit rating

The weights are a starting point, not a universal rule. For a firm where E2 is the dominant margin driver, its weight should rise and E3's should fall. The weights should be set by segment, not by firm, to prevent the loan officer from gaming the model by reweighting. The segment-level weights are reviewed annually.

5.3.3 Scoring Each Indicator

Each indicator is converted to a score between 0 and 100 using a piecewise linear function. The function has three anchor points: a floor (score 0), a benchmark (score 60), and a best-in-class (score 100). The benchmark anchor is set at the industry median or the regulatory threshold; the best-in-class anchor at the technical maximum or the top decile.

For E2, material efficiency, the scoring function for a cobalt recycler might be:

where \(r\) is the recovery rate. A firm at the industry median of 70 percent scores 60. A firm at 92 percent scores \(60 + 40 \times (0.22/0.25) = 95.2\). A firm at 55 percent scores 15. The function is steep between 50 and 70 percent, where most of the industry sits, and flatter above 70, where the marginal credit benefit of further improvement is smaller. This shape reflects the credit reality: the difference between 55 and 70 percent recovery is the difference between a marginal and a viable business; the difference between 85 and 92 percent is a difference in profitability, not viability.

The anchor points must be set per product and per region. A cobalt recycler in Jiangxi faces different scrap prices and different technology than one in Guangdong. Using a national benchmark would misclassify both. The bank should maintain a benchmark library with regional and product-level granularity, updated annually.

5.3.4 The Composite Score and the Adjustment

The composite environmental score is a weighted average with a weakest-link cap:

where \(S_{E,\min}\) is the minimum indicator score and \(\kappa\) is a compensation allowance, set at 25 points. The cap means that a firm with one indicator at 20 cannot score above 45 regardless of its other indicators. The allowance means that a single weak indicator does not automatically set the composite to the minimum; the firm gets some credit for strength elsewhere.

The composite score maps to a notch adjustment through a table:

Table 5.2: Environmental score to notch adjustment

The adjustment is then multiplied by the materiality weight \(\mu\), which is the share of gross margin variance explained by environmental variables, and rounded to the nearest integer:

For a battery recycler with \(\mu = 0.8\) and \(S_E = 95\), the adjustment is \(\text{round}(0.8 \times 2) = +2\) notches. For a software firm with \(\mu = 0.1\) and the same score, the adjustment is \(\text{round}(0.1 \times 2) = 0\) notches. The same environmental performance produces different rating effects because it has different credit relevance. This design distinguishes the framework from an ESG overlay.

5.3.5 Worked Example: The Chongqing Recycler

Return to the firm from the opening. Revenue 180 million yuan, debt-to-asset ratio 61 percent, operating cash flow positive for three years. The bank's anchor rating is BB.

Environmental indicators:

• E1 energy intensity: 3.2 GJ per ton of processed cathode, against a BAT benchmark of 2.8 and a floor of 6.0. Score: \(60 + 40 \times (3.2 - 2.8)/(6.0 - 2.8)\) inverted—wait, lower is better for energy intensity. The scoring function must be inverted: a firm at the floor (6.0) scores 0, at the benchmark (2.8) scores 60, at best-in-class (2.0) scores 100. At 3.2, the score is \(60 \times (6.0 - 3.2)/(6.0 - 2.8) = 52.5\). This is below the benchmark, appropriately.

• E2 recovery rate: 92 percent cobalt, 88 percent nickel. Using the cobalt function, score 95.2. The nickel score, against a benchmark of 80 percent and a best-in-class of 95 percent, is \(60 + 40 \times (88 - 80)/(95 - 80) = 81.3\). The indicator score is the lower of the two, 81.3, because the binding constraint is the weaker metal. This is the weakest-link logic applied within the indicator.

• E3 carbon intensity: 1.8 tCO2e per ton of processed material, against a sector benchmark of 2.5 and best-in-class of 1.2. Score: \(60 + 40 \times (2.5 - 1.8)/(2.5 - 1.2) = 81.5\).

• E4 policy exposure: 15 percent of margin from a local recycling subsidy. Score: \(100 - 15 \times 4 = 40\). The function assumes a linear penalty of 4 points per percentage point of subsidy dependence, reaching 0 at 25 percent. This is a judgment parameter and should be calibrated.

• E5 compliance: one minor violation in five years, a wastewater discharge exceedance that was remediated. Score: 70.

Composite: \(0.20 \times 52.5 + 0.30 \times 81.3 + 0.20 \times 81.5 + 0.20 \times 40 + 0.10 \times 70 = 10.5 + 24.4 + 16.3 + 8.0 + 7.0 = 66.2\). The weakest-link cap: minimum score is 40 (E4), cap is \(40 + 25 = 65\). The composite is capped at 65. This is a meaningful result: the firm's subsidy dependence is the binding constraint, and it pulls the composite down from 66 to 65.

The composite of 65 maps to a notch adjustment of 0. The materiality weight is 0.8. The adjustment is \(\text{round}(0.8 \times 0) = 0\). The final rating is BB, unchanged.

Is this the right answer? The firm's recovery rate is excellent, but its subsidy dependence is a real risk. A zero adjustment says the two roughly cancel. The bank's original BB was not wrong; it was uninformed. The framework confirms the rating but explains it. That explanation is the value: the credit committee now knows that the rating is sensitive to subsidy policy and insensitive to metal prices, which is the opposite of what it would have assumed.

If the firm reduced its subsidy dependence to 5 percent, E4 would score 80, the cap would lift, and the composite would be \(66.2 + 0.20 \times 40 = 74.2\), mapping to +1 notch. The materiality-weighted adjustment would be \(\text{round}(0.8 \times 1) = +1\), and the rating would rise to BB+. The framework gives the firm a path to a better rating through a specific operational change. That is what a rating framework should do: it should tell the borrower what would improve its creditworthiness.

5.3.6 Implementation

Implementation proceeds through six stages, beginning with segment definition and ending with governance. Figure 5.2 shows the sequence.

flowchart LR
    S1[1. Segment definition] --> S2[2. Indicator selection]
    S2 --> S3[3. Benchmark library]
    S3 --> S4[4. Scoring functions]
    S4 --> S5[5. Calibration]
    S5 --> S6[6. Governance and review]
    S6 -.->|annual update| S3

The first stage fixes the unit of analysis. A segment is a group of firms with similar production processes and similar environmental cost drivers. Battery recycling is a segment. Solar module manufacturing is a segment. Energy-efficient lighting is a segment. The segment definition determines which indicators apply and which benchmarks are used. A bank with a large green technology portfolio may have 15 to 20 segments; a bank with a small one may have 4 or 5.

Within each segment, the bank selects the indicators that have a causal chain to cash flow. The five core indicators are a default menu, not a mandatory list. A segment where water is the binding environmental constraint (textile dyeing, for example) should include a water intensity indicator. A segment where hazardous materials are the main risk should include a hazardous waste indicator. The selection should be documented with the causal chain for each indicator.

The benchmark library is the framework's knowledge base. It holds technical maxima, regulatory thresholds, and industry medians, by segment and by region. It can be built from industry association data, engineering literature, and the bank's own loan portfolio. It must be updated annually. A benchmark that is five years old is a benchmark for a technology that no longer exists.

For each indicator, the bank defines the piecewise linear function that converts the physical measure to a score, with anchor points at the floor, the benchmark, and best-in-class. The anchor points come from the benchmark library. The functions should be tested against a sample of known firms to ensure they produce sensible scores.

Calibration maps the composite score to a notch adjustment. The initial calibration is by expert judgment, using the causal chain and the materiality weight. As data accumulate, the calibration is replaced by estimated relationships between the composite score and default or distress outcomes. The bank should set a target date for the first empirical calibration, say three years after implementation, and begin collecting the data needed for it from day one. The data collection is not a byproduct of the framework; it is the point of the framework. A bank that implements the scoring but never records the outcomes has built a reporting tool, not a rating model.

The final stage assigns ownership. Inmost banks this is the risk management department, with input from the green finance desk and the credit committee. The owner is responsible for the benchmark library, the annual recalibration, and the review of exceptions. Exceptions matter: a loan officer who overrides the environmental adjustment should document why, and the pattern of overrides should be reviewed. A cluster of overrides in one segment is a signal that the benchmarks are wrong or the loan officers are gaming the model.

The six stages take 9 to 15 months for a bank starting from scratch. The binding constraint is usually the benchmark library, not the scoring functions. Engineering data on best-available technology are scattered across industry associations, equipment vendors, and academic papers, and assembling them into a usable library is slow work. A bank that tries to shortcut this stage will end up with benchmarks that are either too generous (everyone scores well) or too harsh (no one qualifies), and the framework will be abandoned within a year.

5.3.7 The Technology Maturity Modifier

A green technology firm's default risk is driven by two things: its environmental performance relative to its current competitors, and the maturity of its technology. The five indicators capture the first. They do not capture the second.

Technology maturity matters because a firm with a proven process and a firm with a promising process have different failure modes. The proven firm fails if it is outcompeted on cost. The promising firm fails if the process does not scale. The second failure mode is more common and more sudden.

I propose a technology maturity modifier applied after the environmental adjustment. It has three levels.

Mature. The technology has been deployed at commercial scale for more than five years, with multiple independent operators and a stable cost curve. Example: crystalline silicon solar modules. Modifier: 0.

Scaling. The technology has been deployed at commercial scale for two to five years, with a small number of operators and a cost curve that is still falling. Example: lithium-ion battery recycling via hydrometallurgy. Modifier: −1 notch.

Emerging. The technology has been deployed at pilot or demonstration scale, with no commercial-scale operator that has run for two years. Example: direct lithium extraction. Modifier: −2 notches.

The modifier is a judgment, and it should be reviewed by someone with engineering competence, not only credit competence. A bank that does not have an engineer on staff should retain one for the annual review. The cost is trivial relative to the loss from misclassifying an emerging technology as scaling.

The modifier interacts with the environmental adjustment in a specific way. A firm with excellent environmental performance and emerging technology can end up with a net negative adjustment. That is the correct outcome. The firm's process may be the best in the world and still fail, and the rating should reflect the failure probability, not the process quality.

5.3.8 What the Framework Does Not Do

I will be explicit about the limits, because a framework that is oversold will be discarded when it fails to deliver.

It does not predict which technology will win. It rates the firm in front of it, not the industry trajectory. A bank that lends to the best battery recycler in a province is still exposed to the risk that the province's battery recycling capacity is three times what the waste stream can support. That is a concentration risk, and it belongs in the portfolio model, not the firm rating. Chapter 7 takes it up.

It does not replace the loan officer's judgment on management quality, fraud risk, or related-party transactions. The environmental adjustment is bounded at two notches precisely so that it cannot overwhelm the qualitative overlay. A firm with a 95 environmental score and a management team under investigation for embezzlement should be declined, and the framework will not stop the credit committee from declining it.

It does not solve the through-the-cycle problem. The technology maturity modifier is a partial fix, but it is a judgment, not an estimate. The deeper problem—that the long-run default rate for a young segment is unknown—remains. The framework's response is to be conservative: when the segment's default history is thin, the anchor rating should be set at the conservative end of the plausible range, and the environmental adjustment should be applied to that conservative anchor.

It does not work without data. A bank that cannot obtain utility bills, production records, and material recovery data cannot score the indicators. For listed firms, some of this is in annual reports and sustainability disclosures. For private firms, it requires a site visit and a data request. The data request is a cost, and it is a screening device: a firm that cannot or will not provide recovery-rate data is a firm whose recovery rate is probably not a selling point.

5.4 Testing the Framework Against Cases

A framework is a hypothesis. It should be tested against cases where the outcome is known. I will use three, drawn from the public record and from the kind of firm a Chinese commercial bank would encounter. The names are changed; the parameters are not.

5.4.1 Case One: A Solar Module Manufacturer That Survived the Consolidation

Firm A manufactured multicrystalline silicon modules in Jiangsu. In 2010 it had 1.2 GW of capacity, revenue of 4.1 billion yuan, and a debt-to-asset ratio of 68 percent. Its energy intensity was 0.9 kWh per watt of module capacity, against a 2010 industry median of 1.3 and a best-in-class of 0.7. Its material yield was 96 percent, against a median of 92 percent. Its carbon intensity was 1.1 kg CO2e per watt, against a sector benchmark of 1.5. It had no subsidy dependence beyond the national feed-in tariff that supported its customers, and no compliance violations.

Under the framework: E1 score \(60 + 40 \times (1.3 - 0.9)/(1.3 - 0.7) = 86.7\). E2 score \(60 + 40 \times (96 - 92)/(98 - 92) = 86.7\). E3 score \(60 + 40 \times (1.5 - 1.1)/(1.5 - 0.8) = 82.9\). E4 score 90 (10 percent exposure, mostly indirect). E5 score 100. Composite: \(0.20 \times 86.7 + 0.30 \times 86.7 + 0.20 \times 82.9 + 0.20 \times 90 + 0.10 \times 100 = 17.3 + 26.0 + 16.6 + 18.0 + 10.0 = 87.9\). No cap binds. Notch adjustment from the table: +2. Materiality weight for a module manufacturer: energy and materials are roughly 55 percent of cost, so \(\mu \approx 0.55\). Adjustment: \(\text{round}(0.55 \times 2) = +1\). Technology maturity: mature, modifier 0. Net adjustment: +1 notch.

The firm's anchor rating in 2010, on a standard scorecard, would have been BBB−. The framework gives BBB. The firm survived the 2011–2013 consolidation, though its margins compressed severely; it did not default. The one-notch uplift was directionally correct. It was not sufficient to protect a lender from a 40 percent inventory write-down, but the rating model is not a price-forecast model. It said the firm was more likely to survive than its peers, and it was.

5.4.2 Case Two: A Thin-Film Startup That Did Not

Firm B developed cadmium telluride thin-film modules in Hebei. In 2012 it had 80 MW of pilot capacity, revenue of 210 million yuan, and a debt-to-asset ratio of 44 percent. Its energy intensity was 0.6 kWh per watt, better than Firm A. Its material yield was 88 percent, below the multicrystalline median. Its carbon intensity was 0.8 kg CO2e per watt, better than Firm A. Its subsidy dependence was high: 35 percent of margin came from a provincial technology demonstration grant. It had no violations.

Framework scores: E1 100 (best-in-class). E2 \(60 \times (88 - 80)/(92 - 80) = 40\). E3 100. E4 \(100 - 35 \times 4 = 0\) (the function floors at 0). E5 100. Composite: \(0.20 \times 100 + 0.30 \times 40 + 0.20 \times 100 + 0.20 \times 0 + 0.10 \times 100 = 20 + 12 + 20 + 0 + 10 = 62\). Weakest-link cap: minimum is 0 (E4), cap is 25. Composite capped at 25. Notch adjustment: −2. Materiality weight: 0.7. Adjustment: \(\text{round}(0.7 \times -2) = -1\). Technology maturity: emerging (pilot scale, no commercial operator with two years of runtime), modifier −2. Net adjustment: −3 notches.

The firm's anchor rating was BB. The framework gives B−. The firm ceased operations in 2014 when the provincial grant was not renewed. The framework's verdict was correct, and the two mechanisms that produced it—the subsidy-dependence cap and the technology-maturity modifier—were both doing work. A framework with only the five indicators and no maturity modifier would have given −1, which is too generous. A framework with only the maturity modifier and no subsidy indicator would have given −2, which is closer but still misses the specific failure mode (grant withdrawal, not technology failure). The two together got it right.

5.4.3 Case Three: A Recycling Firm With a Governance Problem

Firm C recycled electronic waste in Guangdong. In 2018 it had revenue of 340 million yuan, a debt-to-asset ratio of 57 percent, and a gold recovery rate of 94 percent from printed circuit boards, against an industry median of 85 percent. Its energy intensity was at the benchmark. Its carbon intensity was at the benchmark. Its subsidy dependence was 20 percent. It had two environmental violations in five years, both for improper storage of hazardous waste, both fined and remediated.

Framework scores: E1 60. E2 \(60 + 40 \times (94 - 85)/(97 - 85) = 90\). E3 60. E4 20. E5 40 (two violations, normalized). Composite: \(0.20 \times 60 + 0.30 \times 90 + 0.20 \times 60 + 0.20 \times 20 + 0.10 \times 40 = 12 + 27 + 12 + 4 + 4 = 59\). Weakest-link cap: minimum is 20 (E4), cap is 45. Composite capped at 45. Notch adjustment: 0. Materiality weight: 0.75. Adjustment: 0. Technology maturity: scaling, −1. Net adjustment: −1 notch.

The firm's anchor rating was BBB−. The framework gives BB+. In 2020 the firm's general manager was detained on suspicion of illegally importing hazardous waste from abroad, a related-party transaction that the financial statements did not disclose. The firm's credit facility was frozen. The framework did not predict the detention. It did downgrade the firm by one notch, and the E5 score of 40 was a visible red flag on the rating sheet. A credit committee that read the score would have asked about the violations. Whether it would have uncovered the import scheme is another matter. The framework's contribution is to make the question askable, not to answer it.

The three cases illustrate the framework's behavior. It rewards genuine process advantages (Case 1). It penalizes subsidy dependence and technology immaturity with enough force to change the rating (Case 2). It surfaces compliance problems as a visible score rather than burying them in a qualitative note (Case 3). None of these outcomes required data that a bank could not obtain. All three required the bank to have a benchmark library and a scoring function, which is the work.

5.4.4 A Fourth Case: When the Framework Overcorrects

The three cases above show the framework working. A framework that is only tested on cases it gets right is not tested. I want to present one where the framework's output was worse than the analyst's judgment, because the failure mode is instructive.

Firm D manufactured industrial heat pumps in Zhejiang. In 2019 it had revenue of 620 million yuan, a debt-to-asset ratio of 52 percent, and an order book that was 70 percent from a single provincial clean-heating program. Its energy intensity was at the benchmark. Its material efficiency was at the benchmark. Its carbon intensity was at the benchmark. Its subsidy dependence was 45 percent of margin, and its compliance record was clean.

Framework scores: E1 60, E2 60, E3 60, E4 \(100 - 45 \times 4 = 0\) (floored), E5 100. Composite: \(0.20 \times 60 + 0.30 \times 60 + 0.20 \times 60 + 0.20 \times 0 + 0.10 \times 100 = 12 + 18 + 12 + 0 + 10 = 52\). Weakest-link cap: minimum is 0 (E4), cap is 25. Composite capped at 25. Notch adjustment: −2. Materiality weight: 0.6. Adjustment: \(\text{round}(0.6 \times -2) = -1\). Technology maturity: mature, modifier 0. Net adjustment: −1 notch.

The firm's anchor rating was A−. The framework gives BBB+. The analyst who covered the firm argued against the downgrade. Her case was that the provincial program was written into a five-year plan with budget appropriations through 2023, that the firm had a 60 percent share of the program's procurement, and that the program's political sponsor had made clean heating a signature initiative. The subsidy dependence was real, but the probability of removal within the rating horizon was low.

The firm did not default. It grew revenue by 30 percent in 2020 and 2021, and it refinanced at a rate close to the A− level. The framework's downgrade was, in this case, a false positive. The E4 scoring function treated all subsidy dependence as equally fragile, and it could not distinguish a statutory entitlement from a discretionary grant.

The lesson is that E4 needs a second dimension: the durability of the policy instrument. A feed-in tariff written into a national law is not the same as a provincial demonstration grant renewed annually. The scoring function should multiply the exposure share by a durability factor—say, 1.0 for statutory multi-year programs, 0.6 for programs with a published multi-year budget, and 0.3 for annual discretionary grants. Under that revision, Firm D's effective exposure would be \(45 \times 0.6 = 27\) percent, its E4 score would be \(100 - 27 \times 4 = 0\) still floored, but the cap would lift only to 25 and the composite would still be capped. The revision would not have saved Firm D. A more aggressive durability factor of 0.3 for a program with a published budget would give an effective exposure of 13.5 percent, an E4 score of 46, a cap of 71, and a composite of 52 uncapped—still a notch adjustment of −1 at \(\mu = 0.6\).

What would have saved Firm D is a different treatment of the cap itself. The weakest-link cap assumes that a single weak indicator is fatal. For E4, that assumption holds when the subsidy is discretionary and fails when it is statutory. The cap should be conditional on the indicator: a binding E2 or E5 score caps the composite, but a binding E4 score should cap it only if the durability factor is below 0.5. This is a refinement that a bank would discover only by testing the framework against its own portfolio, which is why the annual review matters more than the initial design.

I include Firm D because it shows the framework's error rate is not zero and because it shows where the error comes from: a scoring function that collapses a multi-dimensional risk into a single number. Every indicator in the framework does this, and every one of them will produce false positives and false negatives at the margin. The framework's defense is not to eliminate the errors but to make them visible, so that an analyst can override with a documented reason and the override can be reviewed.

5.5 Objections and Responses

I will address four objections that a skeptical credit officer would raise. They are the objections I would raise.

5.5.1 "The environmental score is just another judgment dressed up as a number"

This is the strongest objection. The scoring functions have anchor points that are set by judgment. The weights are set by judgment. The materiality weight is estimated, but crudely. The framework produces a number, and the number has an aura of precision that the underlying inputs do not support.

The response is not to deny the judgment. It is to make the judgment visible and contestable. Every anchor point in the benchmark library should carry a source: an industry standard, an engineering paper, a regulatory threshold, or a documented expert estimate. Every weight should be reviewed annually and changed only with a written rationale. The scoring functions should be published to the credit committee, not hidden in a spreadsheet. A credit officer who disagrees with the E2 benchmark should be able to see where it came from and propose a different one. The framework's precision is not in the number; it is in the audit trail.

There is a deeper point. The alternative to a judgmental score is not the absence of judgment. It is a judgment that is made implicitly, by a loan officer, without documentation, and without review. The Chongqing case at the opening of this chapter is an example: the loan officer made a judgment that the firm was a BB, based on the financials, and the judgment was incomplete because the environmental information was not in front of him. The framework does not eliminate judgment. It moves it to a place where it can be examined.

5.5.2 "We do not have enough defaults to calibrate the model"

True, and the framework's response is the three-tier calibration strategy described earlier. Start with expert parameters, move to proxy outcomes, and move to direct calibration when the data permit. The important thing is to start collecting the data now. A bank that waits until it has enough defaults to calibrate will wait forever, because the segment is young and the defaults accumulate slowly. A bank that implements the framework with expert parameters and records the outcomes will have a calibration dataset in three to five years. The choice is between a model with acknowledged uncertainty and no model at all.

There is a second response. The bank does not need to calibrate the full model to get value from it. The environmental score is useful even before it is mapped to a notch adjustment, because it forces the loan officer to collect the data and ask the questions. A credit committee that sees a recovery rate of 55 percent against a benchmark of 70 will ask why, and the answer may be more valuable than the notch adjustment. The framework's diagnostic value precedes its statistical validation.

5.5.3 "The framework will penalize firms in provinces with lenient environmental enforcement"

This is a real problem. The E5 compliance indicator is based on violations, and violations are a function of enforcement intensity as well as firm behavior. A firm in a province with a lax environmental regulator will have fewer recorded violations than an identical firm in a province with a strict regulator. The indicator will reward the firm in the lax province, which is the opposite of what it should do.

The fix is to normalize E5 by the province's enforcement intensity. The bank can construct an enforcement index from public data: the number of environmental enforcement actions per 10,000 industrial firms, or the ratio of fines to inspections. The firm's violation count is divided by the index. A firm with two violations in a province with an index of 1.0 scores the same as a firm with one violation in a province with an index of 0.5. This is an imperfect correction, and it requires data that the bank may not have. But it is better than using raw violation counts, and it signals to the loan officer that the indicator is about behavior, not about the regulator's budget.

The same logic applies to E3, carbon intensity, where the regulatory benchmark varies by province. The benchmark library should be regionalized, and the scoring function should use the province-specific benchmark. A firm at 1.8 tCO2e per ton in a province with a 2.5 benchmark is performing well; the same firm in a province with a 1.5 benchmark is performing poorly. The score should reflect the difference.

5.5.4 "The framework is too complex for a loan officer to use"

The complexity is in the construction, not the use. Once the benchmark library and the scoring functions are built, the loan officer enters five physical measurements and reads a score. The measurements come from utility bills, production records, and a site visit. The score is computed by the system. The loan officer's job is to collect the data and to explain any anomaly, not to compute the score.

The construction is the hard part, and it is a one-time cost per segment. A bank that builds the battery recycling segment in year one, the solar manufacturing segment in year two, and the energy-efficient equipment segment in year three will have a working framework for its main green technology exposures within three years. The segments are built once and updated annually. The marginal cost of adding a firm to an existing segment is small.

There is a real risk that the framework becomes a compliance exercise: the loan officer fills in the fields, the system produces a score, and no one looks at it. The governance step is the defense. The credit committee should review the environmental scores of the largest exposures each quarter and ask about the outliers. If the scores are never questioned, the framework is dead, and the bank has spent 12 months building a form.

5.5.5 "The framework will be gamed"

Any scoring system that affects pricing will be gamed. The question is which gaming strategies the design closes and which it leaves open.

The most obvious strategy is to select the benchmark. A loan officer who wants a higher score can argue that the firm's product category is different from the one in the library, and that a different benchmark applies. The defense is segment governance: the segment assignment is made by the risk department, not the loan officer, and it is recorded. A firm cannot be reassigned to a friendlier segment without a documented reason.

A second strategy is to time the measurement. Energy intensity varies with utilization: a firm running at 50 percent capacity has high energy per unit of output because fixed energy loads are spread over fewer units. A firm can be measured in a good quarter and scored on it. The defense is to use a trailing twelve-month average and to record the measurement date. The data model should reject a single-quarter measurement for E1.

A third strategy is to redefine the output. Material efficiency is output over input, and a firm can change what it counts as output. A recycler that counts only the primary metal and excludes the slag can report a higher recovery rate than a recycler that counts both. The defense is a standardized definition per segment, published in the benchmark library, with a required unit and a required boundary. The definition should be specific enough that two firms in the same segment cannot report on different bases.

None of these defenses is airtight. A determined firm with a cooperative loan officer can still game the score. The framework raises the cost of gaming and makes the attempt visible in the data record, which is an improvement over a system where the environmental information is not collected at all. It does not eliminate the problem, and a bank that expects it to will be disappointed.

5.6 The Rating as an Identification Device

The framework's purpose is not only to price loans. It is to identify which firms are genuinely engaged in the green transition and which are not. This is the "rating-based identification" in the chapter title, and it deserves separate treatment.

A green credit policy that identifies firms by catalog or by list is a policy of self-declaration. A firm is green because it says it is, or because its product category is on a list. The identification is coarse and gameable. A firm can be on the list and have a poor environmental performance; a firm can be off the list and have an excellent one. The rating framework replaces self-declaration with measurement. A firm is identified as green technology because its environmental indicators are above the benchmark and because those indicators drive its cash flow. The identification is continuous, not binary, and it is tied to the firm's actual operations.

The identification can allocate subsidized credit, monitor transition progress, and price the green premium. Take the first use. A bank that wants to direct funds to green technology firms can use the environmental score as a screening criterion: firms above a threshold get the subsidized rate, firms below do not. The threshold is transparent and the score is auditable. This is more defensible than a catalog, because it does not depend on the firm's self-description. The second use follows from the first. A firm's environmental score can be tracked over time, and a firm whose score is rising is improving its process while a firm whose score is falling is losing ground. The trajectory is more informative than the level. A firm at 60 that was at 40 three years ago is a better credit than a firm at 65 that was at 80, even though the second firm has a higher score today. The framework should record the score history and use the trend as an input to the qualitative overlay. The third use connects the score to cost structure. The green premium, as Chapter 4 defined it, is the extra cost of doing the right thing. A firm with a high environmental score often has a lower cost of doing the right thing, because its process is efficient. A firm with a recovery rate of 92 percent has a negative green premium relative to a firm at 70 percent—it is cheaper to be green, not more expensive. That negative premium is the firm's competitive advantage, and the rating framework makes it visible by anchoring the score to a physical benchmark rather than a disclosure.

The identification device is not a substitute for the financial analysis. A firm with a high environmental score and a weak balance sheet is still a weak credit. The framework's contribution is to add a dimension that the financial analysis lacks, not to replace it.

5.7 Integration With the Existing Rating Infrastructure

A bank that adopts the framework does not discard its existing rating models. It adds a module. The integration points are four.

Data infrastructure. The environmental indicators are new data fields. They must be captured in the loan origination system, stored with the loan record, and made available to the rating engine. The data model should record the measurement, the unit, the benchmark used, the score, and the date. The benchmark used is important: a score of 60 against a 2019 benchmark is not the same as a score of 60 against a 2024 benchmark, and the record should allow a reader to reconstruct the calculation.

Rating engine. The environmental adjustment is a post-processing step. The anchor rating is computed as before. The environmental score is computed from the indicators. The adjustment is applied, the technology maturity modifier is applied, and the final rating is produced. The engine should record each step so that the final rating can be decomposed. A rating that cannot be decomposed cannot be defended.

Calibration tables. The mapping from final rating to probability of default should be recalibrated for the green technology segments. The existing table, estimated on a broad corporate portfolio, may not fit. A segment with a higher default rate than the portfolio average should have its own table, or a segment-specific multiplier. The multiplier should be estimated from the segment's default history, and until that history exists, it should be set conservatively.

Reporting. The framework produces a set of reports: the distribution of environmental scores by segment, the share of the portfolio above and below the threshold, the trend of scores over time, and the list of exceptions. These reports are the basis for the annual review and for the dialogue with the regulator. A bank that cannot report its green technology exposure by environmental score cannot demonstrate that its green credit policy is doing what it claims.

The integration is not trivial. It requires changes to the loan origination system, the rating engine, and the reporting stack. A bank that treats it as an IT project will fail, because the hard part is the benchmark library and the governance, not the software. A bank that treats it as a risk management project, with IT as a supporting function, has a chance.

There is a cost question that the integration raises and that I have not yet addressed. The benchmark library, the scoring functions, and the annual recalibration require staff time: roughly 0.5 to 1.0 full-time equivalent for a mid-sized commercial bank with 15 to 20 green technology segments, plus the cost of an engineering consultant for the technology maturity reviews. Against that, the framework's benefit is the avoidance of adverse selection in a segment where within-grade dispersion is large. If the bank's green technology book is 2 billion yuan and the framework improves the selection of borrowers by even 50 basis points of annual loss rate, the benefit is 10 million yuan a year, which is several times the cost. If the book is 200 million yuan, the benefit is 1 million yuan and the cost may exceed it. The framework's economics are scale-dependent, and a small bank should consider a simplified version—three indicators, a single benchmark per segment, and expert calibration—rather than the full design.

5.8 The Limits of Rating-Based Identification

I have argued that rating-based identification is superior to list-based identification. I should also say where it stops.

It stops at the boundary of the firm. A rating model assesses a borrower. It does not assess the borrower's supply chain, its customers, or the systemic risk its industry poses. A battery recycler with an excellent environmental score may source scrap from informal recyclers who dump acid into rivers. The firm's score is high; the supply chain's environmental performance is poor. The rating framework does not see this, and it should not pretend to. Supply chain environmental risk is a separate assessment, and it belongs in the counterparty risk or reputational risk function, not the credit rating.

It stops at the boundary of the measurable. Some of the most important environmental risks are not measurable at the firm level: the risk that a region's water supply is exhausted, the risk that a technology is rendered obsolete by a breakthrough, the risk that a policy regime changes abruptly. These are scenario risks, and they belong in stress testing, not in the point-in-time rating. The rating framework can flag them—the policy exposure indicator is a partial flag—but it cannot quantify them.

It stops at the boundary of the data. A firm that does not report its recovery rate cannot be scored on it. The framework's response is to treat missing data as a red flag, not as a neutral value. A firm that cannot provide the data is scored at the benchmark, not above it, and the missing data is recorded. This is a conservative choice, and it creates an incentive for firms to report. It also means that the framework's coverage is biased toward firms with better disclosure, which are usually larger firms. The bias is real and should be acknowledged.

None of these limits undermines the framework. They define its scope. A rating model is a tool for a specific job: forecasting default for a specific borrower over a specific horizon. It does that job better when it includes environmental information that is material to the borrower's cash flow. It does not do other jobs, and a bank that asks it to will be disappointed.

5.9 Summary of the Framework

For reference, the complete framework is summarized below.

Inputs. Five environmental indicators measured in physical units: energy intensity (E1), material efficiency (E2), carbon intensity (E3), policy exposure (E4), environmental compliance (E5). Financial and business inputs from the existing rating model.

Scoring. Each indicator is converted to a 0–100 score using a piecewise linear function with floor, benchmark, and best-in-class anchors drawn from a segment-specific benchmark library.

Aggregation. Weighted average with a weakest-link cap: the composite cannot exceed the minimum indicator score plus a 25-point compensation allowance.

Adjustment. The composite maps to a notch adjustment of −2 to +2, multiplied by a materiality weight \(\mu\) and rounded to the nearest integer.

Technology modifier. A separate adjustment of 0, −1, or −2 notches based on technology maturity (mature, scaling, emerging).

Final rating. Anchor rating plus environmental adjustment plus technology modifier plus qualitative overlay.

Calibration. Three-tier: expert parameters initially, proxy outcomes as data accumulate, direct default calibration when the sample permits.

Governance. Annual review of benchmarks, weights, and scoring functions; quarterly review of large exposures and exceptions; documented rationale for every parameter change.

The framework's value is not in any single component. It is in the combination: physical measurement, segment-specific benchmarks, bounded adjustment, and explicit governance. Remove any one and the framework either becomes an ESG overlay (remove the benchmarks), a sustainability rating (remove the bound), or a black box (remove the governance). The combination is what makes it a credit tool.

The framework's value lies in making environmental performance visible to credit committees. A bank that adopts it will not win every green technology loan, and it will still decline firms it should have financed. What changes is that the decline will be a decision rather than an oversight. The Chongqing recycler's recovery rate will sit on the rating sheet, next to the leverage ratio and the cash-flow coverage, and the committee will have to weigh it. That is a lower bar than predicting the transition, and it is the bar that a rating model can actually clear.