Chapter 6: Scoring Models and Fundamental Analysis: Complementary Tools for Green Credit Identification

In 1932, a statistician at the University of Minnesota named Paul Fitzpatrick published a small study comparing the financial ratios of firms that had defaulted with those of firms that had survived. He found a pattern, and in doing so he gave the credit industry its first empirical hint that a company's accounting numbers might carry information about its likelihood of failure. Four years later, Ronald Fisher formalized the statistical machinery—discriminant analysis—that would let later researchers separate such groups more rigorously. The line from Fitzpatrick's ratios to today's machine-scored credit models is long but unbroken. It is also, for our purposes, a cautionary line. Every step along it involved a decision about what to measure and what to ignore. Environmental performance, for most of that history, was on the ignore list.

This chapter is about the two dominant families of tools that banks use to identify creditworthy borrowers, and about what happens when environmental performance is added to the identification problem. The first family is quantitative scoring: models that convert a set of observable inputs into a number, a grade, or a probability of default. The second is fundamental analysis: the analyst's structured reading of a firm's accounts, its industry, its management, and its prospects. They fail in different ways, and their failure modes are complementary.

A green credit identification system that relies on either alone will misclassify systematically: the score will reward firms that are brown but efficient today, and the analyst will be captured by narrative and by the most recent site visit. The hybrid model—score as the screen, fundamental analysis as the adjudicator, with explicit rules about when the analyst may override the score and how that override is recorded and reviewed—is the only structure that survives contact with a real loan book.

6.1 The Role of Scoring Models: Advantages and Limitations of Quantitative Scoring

6.1.1 What a Score Actually Is

A credit score is a mapping. It takes a vector of borrower characteristics and returns a number that is monotonically related to some estimated probability of default. The mapping can be as simple as an Altman-style discriminant function, or as complex as a gradient-boosted tree with several thousand splits. The statistical machinery matters less than three design choices: which variables enter, how the model is trained, and how the output is used.

Consider the classic form. A linear score \(S\) for firm \(i\) might be written

where the \(x_{ij}\) are financial ratios—leverage, coverage, liquidity, profitability, size—and the \(\beta_j\) are estimated so that the score separates defaulters from non-defaulters in a historical sample. The score is then mapped to a rating grade or a probability of default through a calibration table. Everything about the model's behavior is contained in the choice of \(x\), the estimation sample, and the calibration.

The appeal is obvious. Once estimated, the model runs in milliseconds. It applies the same weights to every borrower in a segment. It does not get tired, does not have a favorite industry, does not remember that the last loan to a solar installer went bad. For a bank processing ten thousand small business loans a year, this consistency is worth a great deal. A human analyst reviewing ten thousand files would drift; the model will not.

There is a second, less discussed advantage. A score makes disagreement visible. When two analysts look at the same file and reach different conclusions, the reasons are buried in their separate judgments. When an analyst overrides a score, the override is a discrete, recordable event. That record is the raw material for model governance, and model governance is what separates a scoring system from a black box.

6.1.2 The Variables That Dominate—and What They Miss

Empirical work on default prediction has been remarkably stable in its findings. The variables that carry most of the signal are boring: leverage, interest coverage, cash flow relative to debt, firm size, and a short history of profitability. Fitzpatrick's ratios, Altman's 1968 function, and the internal scorecards of most commercial banks today are variations on the same short list. The stability is reassuring. It also tells us something uncomfortable: the accounting variables that predict default describe the firm's current financial condition, not its future environmental trajectory.

This matters for green credit because the environmental transition is, by construction, a forward-looking phenomenon. A cement plant with strong margins and low leverage scores well today. If a carbon price of $80 per ton is legislated to take effect in three years, the plant's future cash flows are impaired, but its current accounts are not. The score sees the present. The transition risk lives in the future. A scoring model trained on historical defaults will systematically under-price transition risk, because in the training sample, transition risk had not yet crystallized into default.

There is a partial fix. If environmental variables are added to the score as additional inputs—say, an emissions intensity measure, a brown-revenue share, a taxonomy alignment flag—the model can learn from any historical cases where high-emission firms did default at higher rates. But the learning is limited by the sample. If the sample period predates the transition, the coefficient on the environmental variable will be estimated near zero, and the model will conclude, correctly given its data, that environmental performance does not predict default. The model is answering a question about the past.

6.1.3 The k-Nearest-Neighbor Problem, Generalized

There is a technical version of this problem that repays attention. Many scoring systems, particularly for small and medium enterprises, use some form of nearest-neighbor classification: a new borrower is assigned the default rate observed among the \(k\) most similar borrowers in the historical pool. The choice of \(k\) is a trade-off. If \(k\) is too small, the classification is unstable—one unusual neighbor can swing the result. If \(k\) is too large, the neighborhood blurs and valuable information is lost. The same trade-off, in a different register, applies to the environmental dimension. If the pool of comparable green firms is small—as it is in most emerging markets, and in most specialized green technologies even in developed ones—then the effective \(k\) for a green borrower is tiny, and the score is unstable. The model has no peers to learn from.

The instability has a practical consequence that shows up in loan committees. A green technology firm with a thin track record will receive a wide confidence interval around its score. If the bank treats the point estimate as the decision, it will either over-lend to the first few firms in a new sector or refuse the sector entirely. Neither is what a transition strategy requires.

6.1.4 Where Scores Genuinely Win

None of this is an argument against scoring. It is an argument for using scores where they are strong. Three domains stand out.

First, scores are excellent at detecting deterioration within an established relationship. A firm that has been borrowing for eight years has a time series of financials, and the score's movement over time carries real information. A two-notch downgrade in a score is a signal to look, whatever the environmental overlay says.

Second, scores are the only practical way to price and provision at portfolio scale. A bank with 40,000 SME exposures cannot have an analyst read each file quarterly. It can, however, run a score and flag the bottom decile for review. The score's role here is triage, not judgment.

Third, scores impose discipline on the analyst. A model that says "this firm is a B" when the analyst's gut says "A" forces the analyst to articulate why. Sometimes the analyst is right and the model is missing something the analyst can see. Sometimes the analyst is pattern-matching on a charming management team. The disagreement is productive either way, provided the bank records it.

6.1.5 The Limits, Stated Plainly

Three limits deserve to be named, because each has a green-specific version.

The first is sample bias. A model trained on a period when environmental risk was not priced will underweight it. The green-specific version: a model trained on a loan book dominated by brown firms will have learned the brown firms' survival patterns and will treat green firms as outliers. The score will be systematically less reliable for the segment the bank most wants to grow.

The second is variable construction. Adding an environmental variable to a score requires that variable to be measured consistently across the portfolio. Emissions intensity is straightforward for a cement plant and ambiguous for a software firm whose main environmental impact is the electricity consumed by its data centers, which it does not own. The variable that can be measured for everyone will be coarse; the variable that is precise will be available for only part of the book. This is the same trade-off that governs the choice of \(k\) in nearest-neighbor methods, and it does not have a clean solution.

The third limit is horizon. A score is a statement about the probability of default over a horizon, typically one year. Transition risk often operates over three to ten years. A model calibrated to a one-year horizon will not, and should not, incorporate a carbon price scheduled for 2030. The score is answering a short question. The green credit decision is answering a long one.

6.2 Fundamental Analysis for Green Projects: Qualitative and Forward-Looking Assessment

6.2.1 What Fundamental Analysis Does That a Score Cannot

Fundamental credit analysis is the structured examination of a borrower's financial statements, business model, industry position, management quality, and legal and contractual environment, undertaken to form a judgment about the borrower's capacity and willingness to repay. The phrase "structured examination" is doing a lot of work. Good fundamental analysis is not a conversation. It is a sequence of questions with evidence requirements attached.

The tradition has a long lineage in credit practice, and its canonical form is well documented in the standard credit risk literature: accounting basics first, then a reading of the balance sheet, income statement, and cash flow statement, then the key ratios that summarize financial health, then the qualitative overlay of operations and economic environment. The overlay is where green analysis enters, and it is where the analyst earns their keep.

Consider a concrete case. A provincial bank in southwest China is evaluating a loan to a tea cooperative in Mangjing Village, in Pu'er prefecture, Yunnan. The village has 608 households and 2,568 people, 97 percent of them Blang ethnic minority. Eighty percent of household income comes from tea. The cooperative wants to finance a transition to organic certification and to install rainwater harvesting and shade-tree management. The financial statements are thin: the cooperative's accounts are kept by a part-time bookkeeper, and the "collateral" is largely the tea bushes themselves, which are difficult to value and harder to seize.

A scoring model applied to this file would produce a low score, and it would be right on the numbers. The cooperative's leverage is high relative to cash flow, its revenue is concentrated in a single buyer, and its accounting is weak. The score would say no.

The fundamental analyst asks different questions. What is the buyer concentration, and is the buyer a state-owned tea company with a multi-year purchase agreement? (In this case, yes, and the agreement is on file.) What is the certification pathway, and who pays for the audit? (A provincial program covers the first three years.) What is the climate exposure? (The village sits at altitude; spring frost is the main risk, and shade trees reduce it.) What is the social structure through which the loan would be intermediated? (The cooperative is organized with a village branch committee, a structure that concentrates decision authority and speeds execution but, over time, can dampen member participation.)

None of these questions is answered by the score. Each of them changes the credit judgment. The analyst may conclude that the loan is sound despite the weak ratios, or that it is unsound despite a strong buyer relationship. The point is that the judgment requires information the score does not contain.

6.2.2 The Green Analyst's Question Set

What distinguishes green fundamental analysis from conventional fundamental analysis is not a different method but a different question set. The analyst is asking, in addition to the standard questions, a set of transition questions. I find it useful to organize them into four groups.

The first group concerns physical exposure. What does the borrower's production process emit, and where in the process are the emissions concentrated? A cement plant's emissions are dominated by the calcination of limestone, which is a chemical necessity, not a fuel choice. A textile dyeing operation's emissions are dominated by the energy source for process heat, which is a fuel choice. The distinction matters because it determines what a transition pathway can actually change. For the cement plant, the pathway runs through clinker substitution and carbon capture; for the dye works, it runs through fuel switching. The analyst who does not know the difference will accept a transition plan that cannot deliver.

The second group concerns policy exposure. What carbon price, efficiency standard, or disclosure requirement is scheduled or plausible in the borrower's jurisdiction over the loan's tenor? The analyst does not need a forecast. They need a schedule of announced policies and a judgment about which are likely to bind. A firm whose products are exported to the EU faces the Carbon Border Adjustment Mechanism on a known timetable. A firm selling domestically faces whatever the domestic carbon market does, which is less certain but not unknowable.

The third group concerns transition capability. Does the borrower have the capital, the technical staff, and the management attention to execute a transition? This is where the analysis becomes genuinely qualitative. A firm with a credible chief engineer and a board that has discussed the transition is in a different position from a firm whose transition plan consists of a paragraph in the annual report. The analyst's job is to distinguish the two, and the evidence is usually documentary: board minutes, capital expenditure plans, hiring records, the existence of a named responsible executive.

The fourth group concerns the counterparty's own incentive structure. This is the oldest insight in fundamental credit analysis and it applies with full force to green lending. A borrower's decisions are shaped by the incentives of its managers and shareholders. If a state-owned enterprise's leadership is evaluated on output and employment, a transition plan that reduces both will not be executed, whatever the loan covenant says. If a private firm's controlling shareholder has pledged shares against a loan, the shareholder's tolerance for transition-related earnings volatility is lower than the firm's long-term interest would suggest. The analyst who ignores incentives will be surprised by execution failure that was visible in the ownership structure.

6.2.3 The Failure Modes of Fundamental Analysis

Fundamental analysis fails in ways that are the mirror image of scoring's failures. Where the score is consistent but blind, the analyst is sighted but inconsistent.

The first failure mode is narrative capture. A compelling management team with a good story can talk an analyst into a favorable judgment that the numbers do not support. This is not corruption; it is ordinary human cognition. The remedy is procedural: the analyst's written assessment must state, for each material claim, what evidence supports it and what evidence would falsify it. A claim that cannot be falsified should not enter the credit decision.

The second failure mode is recency. The analyst who has just visited a site where the transition is visibly underway will overweight that visit relative to the documentary record. A single site visit is a sample of one. The analyst should treat it as one input among many, and should record what was seen separately from what was inferred.

The third failure mode is scale. Fundamental analysis of the depth described above takes an experienced analyst two to five days per borrower. A bank with 40,000 SME exposures cannot do this for all of them. It can do it for the subset that matters—large exposures, new sectors, and cases flagged by the score—but not for the book.

The fourth failure mode is the green-specific version of recency: the analyst who has recently read a report on a technology or a policy will overweight it. A carbon capture technology that is technically feasible but commercially unproven at scale will be treated as available. A policy that has been announced but not legislated will be treated as certain. The discipline here is to distinguish, in writing, between what is operating, what is demonstrated, and what is announced.

6.2.4 A Note on the Boundary Between the Two

It is tempting to draw a clean line: scores handle the quantitative, analysts handle the qualitative. The line is not that clean, and the reason is worth stating.

Much of what the analyst assesses can, in principle, be quantified. Policy exposure can be modeled. Transition capital expenditure can be read from the capital plan. Management quality can be proxied by tenure, turnover, and the record of plan execution. The reason these are not in the score is not that they are unquantifiable; it is that they are not consistently available across the portfolio, or not available in a form that can be validated against outcomes.

This suggests a useful way to think about the boundary. The score handles what is measured consistently for everyone. The analyst handles what is measured inconsistently, or not at all, but matters for this borrower. As data infrastructure improves—as emissions reporting becomes mandatory, as transition plans become standardized, as policy schedules become public—the boundary moves. Variables migrate from the analyst's domain into the score.

The migration has a cost, though, and it should be acknowledged. When a variable enters the score, it acquires a weight and a direction that are fixed until the model is re-estimated. The analyst's treatment of the same variable can vary with context. A bank that moves too much into the score too quickly will lose the context sensitivity that green lending, in its current state of data scarcity, still requires.

6.3 Combining Both Approaches: A Hybrid Model for Robust Identification

6.3.1 The Architecture

The hybrid model has three layers, and the logic of each layer is determined by what the layer below it cannot do.

The first layer is the score. It runs on every borrower in the segment. Its job is to sort the book into three bands: a pass band where the score is high and no environmental red flag is present, a review band where the score is middling or an environmental flag is present, and a decline band where the score is low and no compensating factor is evident. The score's job is triage, and the bands should be set so that the review band captures the cases where judgment can change the outcome.

The second layer is the environmental flag. This is a set of binary or ordinal indicators that fire independently of the score: the borrower is in a sector on the transition watch list; the borrower's emissions intensity is in the top quartile of its segment; the borrower has a material policy exposure within the loan tenor; the borrower's transition plan is absent or unverifiable. A flag does not determine the decision. It routes the file to the analyst.

The third layer is the analyst's assessment, conducted under a structured template with evidence requirements. The analyst produces a recommendation, a written rationale, and—crucially—a record of any override of the score, with the reason coded against a fixed list.

flowchart TD
    A[Borrower file] --> B[Score S and grade]
    B --> C{Score band?}
    C -->|High, no flag| D[Pass: standard monitoring]
    C -->|Low, no compensating factor| E[Decline]
    C -->|Review band or flag| F[Environmental flag check]
    F --> G[Analyst assessment]
    G --> H{Override score?}
    H -->|No| I[Decision follows score]
    H -->|Yes| J[Override recorded and coded]
    J --> K[Credit committee review]
    I --> K
    K --> L[Decision and covenant package]
    L --> M[Post-decision monitoring]
    M --> B

The loop back from monitoring to the score is the part most banks omit, and it is the part that makes the system learn. Every override, every flag that did or did not predict deterioration, every analyst assessment that proved right or wrong is data. Without the loop, the hybrid model is two separate systems bolted together. With it, the model improves and the analysts' judgment is disciplined by outcomes.

6.3.2 The Override Protocol

The override is where the hybrid model lives or dies. If analysts can override freely, the score is decorative and the bank has reverted to pure judgment with extra steps. If analysts cannot override, the score's blind spots become the bank's blind spots, and green lending stalls.

A workable protocol has four elements.

First, overrides are permitted in both directions. An analyst may upgrade a file the score would decline, and may downgrade a file the score would pass. The second direction is more important than it looks. In practice, banks that allow only upgrades find that their green portfolio accumulates risk that the score tried to flag.

Second, every override requires a written rationale structured against the flag list. The analyst must state which specific factor the score missed and what evidence supports it. "Management quality" is not a rationale. "The borrower has a signed 10-year purchase agreement with a state-owned utility that covers 70 percent of output, which the score does not capture because the contract is off-balance-sheet" is a rationale.

Third, overrides above a threshold—by exposure size, or by number within a portfolio segment—require credit committee approval. The threshold should be set so that routine overrides are handled at the desk and material ones are escalated.

Fourth, overrides are reviewed against outcomes at regular intervals. If a cohort of upgrades performs worse than the score predicted, the override reasons in that cohort are examined. If they cluster around a particular factor, that factor is a candidate for inclusion in the score. If they cluster around a particular analyst or a particular sector, that is a governance signal.

6.3.3 Worked Example: A Solar Module Manufacturer

Consider a mid-sized solar module manufacturer seeking a five-year working capital facility. The firm has grown rapidly; revenue has tripled in three years, but margins have compressed and receivables days have lengthened. The score, dominated by leverage and coverage, places the firm in the review band.

The environmental flags fire. The firm is in a green sector, which is not itself a negative, but the flag system also notes that the firm's customers are concentrated among a small number of project developers whose own financing depends on subsidy regimes that are being reformed. The policy exposure flag is set.

The analyst's assessment proceeds through the question set. Physical exposure: the firm's own operations are relatively clean; the emissions are upstream, in polysilicon production, and downstream, in installation. Policy exposure: the subsidy reform is real and scheduled, and the firm's order book for the next eighteen months depends on projects that have already been awarded, so the near-term exposure is limited but the medium-term exposure is material. Transition capability: the firm has a credible plan to move into the commercial and industrial segment, where demand is less subsidy-dependent, and has hired a sales team for it. Incentives: the founding family owns 45 percent and has not pledged shares; the family's stake aligns with a multi-year transition.

The analyst recommends approval with covenants: a minimum interest coverage ratio, a cap on customer concentration, and a requirement to report the share of revenue from non-subsidy-dependent customers semi-annually. The score would have declined the file. The override is recorded, coded to the policy-exposure factor, and reviewed at the eighteen-month mark against the covenant reports.

This example illustrates the general pattern. The score sees the firm's current financial condition and is right to be cautious. The analyst sees the firm's position in a transition and is right to be constructive. The hybrid decision—approve with covenants that monitor the specific risk the analyst identified—is better than either alone.

6.3.4 Worked Example: A Cement Plant

Now consider the opposite case. A cement plant with strong margins, low leverage, and a long operating history scores well. The score places it in the pass band.

The environmental flags fire anyway. The plant is in a sector on the transition watch list, and its emissions intensity is in the top quartile of its segment. The flag routes the file to the analyst despite the pass score.

The analyst's assessment is less favorable than the score. Physical exposure: the plant's emissions are dominated by calcination, which cannot be eliminated by fuel switching; the credible pathways are clinker substitution and carbon capture, both of which require capital the plant has not committed. Policy exposure: the provincial carbon market is expanding, and the plant's free allocation is scheduled to decline. Transition capability: the plant has no named transition executive and its capital plan is devoted to capacity expansion, not decarbonization. Incentives: the plant is owned by a local government, and the leadership's evaluation criteria emphasize output and employment.

The analyst recommends decline, or approval only with a short tenor and a covenant requiring a transition plan within twelve months. The score would have passed the file. The override is recorded, coded to the transition-capability factor, and reviewed.

The two examples are mirror images, and they make the same point. The score and the analyst see different things. The hybrid model's value is that it forces both views into the same decision, with a record of which view prevailed and why.

6.3.5 The Governance That Makes It Work

A hybrid model without governance degrades into whichever component has more organizational power. In most banks, that is the score, because the score is what the risk function owns and the risk function has veto power. The result is a system that is consistent and blind.

The governance needed has four components, and they are not exotic. They are the standard apparatus of model risk management, applied to the green dimension.

The first is a documented model inventory. Every score used in green credit decisions is listed, with its variables, its estimation sample, its performance metrics, and its known limitations. The inventory is the basis for the override protocol, because an analyst cannot know what the score misses without knowing what it contains.

The second is a flag taxonomy. The environmental flags are defined in a fixed list, with clear triggering rules. The taxonomy is reviewed annually and updated as policy and data change. Without a fixed taxonomy, override reasons drift and cannot be aggregated.

The third is an override log with outcome tracking. Every override is recorded with its reason code, the analyst, the exposure, and the decision. The log is matched against subsequent performance at twelve and twenty-four months. The matching is the learning mechanism.

The fourth is a periodic review of the boundary. Annually, the bank asks which override reasons have proven predictive. Reasons that are consistently predictive are candidates for inclusion in the score. Reasons that are not predictive are candidates for removal from the override list. The boundary moves, deliberately and on evidence.

6.3.6 What the Hybrid Model Cannot Do

It is worth being explicit about the limits, because green credit is a field in which overclaiming is common.

The hybrid model cannot make transition risk visible if the underlying data does not exist. If a borrower does not measure its emissions, the analyst cannot assess them and the score cannot include them. The model's accuracy is bounded by the data infrastructure, which is the subject of a later chapter.

The hybrid model cannot resolve the horizon mismatch. The score is calibrated to a one-year horizon; the transition is a ten-year phenomenon. The hybrid decision can incorporate a longer view through covenants and tenor, but the score itself will continue to answer the short question. Banks that need a long-horizon measure need a separate tool, not a modified score.

The hybrid model cannot substitute for pricing. A bank that identifies a green borrower correctly but prices the loan as if the transition risk were zero will still misallocate capital. Identification is necessary but not sufficient; the green premium, discussed in an earlier chapter, is the pricing counterpart.

And the hybrid model cannot fix incentives inside the bank. If loan officers are compensated on volume and not on portfolio performance, they will use the override protocol to approve files the score would decline, and the override log will fill with rationalizations. The governance described above detects this, but detection is not correction. Correction requires that the compensation structure reward the analyst for being right over the life of the loan, not for getting the deal done.

6.3.7 The Evidence Base, and Its Gaps

The empirical case for hybrid models in green credit is thinner than the conceptual case, and honesty requires saying so.

What is well established is that combined quantitative and qualitative approaches outperform either alone in conventional credit. Altman's original 1968 discriminant model, built on five financial ratios, achieved 95 percent accuracy in classifying bankrupt and non-bankrupt manufacturers one year before failure, but subsequent work by Grunert, Norden, and Weber (2005) on German bank data showed that the introduction of rating systems under Basel II improved banks' screening and monitoring, and that the internal ratings that incorporated soft information—management quality, relationship history—outperformed models built on hard financial data alone. The finding is robust across markets and periods, though the incremental contribution of soft information shrinks as the quality of hard data improves.

What is less established is how much the green dimension adds. Studies of ESG scores and credit risk have produced mixed results. Some find that firms with better environmental performance have lower credit spreads and lower default rates; others find the relationship is driven by size and profitability and disappears once those are controlled. The mixed results are not surprising. ESG scores are noisy, the samples are short, and the mechanisms are heterogeneous across sectors. A cement plant's environmental performance and a software firm's environmental performance are not the same variable.

The honest conclusion is that the green dimension's incremental predictive power is plausible but not yet demonstrated at the level that would justify replacing analyst judgment with a model. This is precisely the condition under which a hybrid model is appropriate. When the data is thin and the relationships are uncertain, the decision should combine a consistent quantitative screen with structured human judgment, and the combination should be designed to generate the data that will eventually settle the question.

6.3.8 A Note on the Chinese Context

The institutional setting in which most of this book's cases sit has features that shape the hybrid model's design.

The first is the role of policy banks and government-guided funds. A large share of green credit in China is intermediated through institutions whose mandates include policy objectives alongside commercial ones. For these institutions, the score's job is different. It must predict default, and it must also screen for policy alignment. A hybrid model in this setting needs a policy-alignment flag in addition to the environmental flags, and the override protocol needs to accommodate policy-driven approvals that the score would decline.

The second is the data environment. Corporate environmental disclosure in China has improved rapidly but remains uneven across firm sizes and sectors. The national carbon market covers the power sector and is expanding; provincial pilot markets cover others. For firms outside the covered sectors, emissions data is often estimated rather than measured. The analyst's assessment must therefore state the data quality explicitly, and the score's environmental variables must be flagged as estimated where they are.

The third is the structure of rural and cooperative lending, which is the subject of later chapters. The Mangjing tea cooperative is not an outlier; it is representative of a large class of borrowers for whom financial statements are weak, collateral is ill-defined, and the credit judgment depends on social structure and production relationships. For this class, the score's role is minimal and the analyst's role is central. A hybrid model that assumes a minimum data standard will exclude this class entirely, which is the opposite of what a transition strategy needs.

6.3.9 The Boundary as a Design Variable

I want to close by returning to the boundary between the score and the analyst, because it is the chapter's central design variable and it is usually treated as fixed when it should be treated as a choice.

The choice has a cost on each side. Moving a factor into the score buys consistency and scale, and costs context sensitivity and the ability to learn from a small number of cases. Leaving a factor with the analyst buys context and forward-looking judgment, and costs consistency and scale. The right position depends on the factor's data quality, the size of the segment, and the stability of the relationship between the factor and default.

For most banks in most markets today, the right position for the green dimension is mostly on the analyst's side, with a small number of well-measured variables—sector classification, emissions intensity where measured, policy exposure where scheduled—in the score. This is not a permanent arrangement. It is the arrangement that the current data supports. As measurement improves and the historical record lengthens, the boundary will move, and the hybrid model should be designed to move with it.

The design that accommodates movement has three features. The override log is structured so that override reasons can be aggregated and tested. The score is re-estimated on a schedule that allows new variables to enter. The flag taxonomy is reviewed annually against the override log, so that flags which consistently predict outcomes become score variables and flags which do not are retired. None of these features is novel. What is novel is applying them to the environmental dimension, which banks have tended to treat as a separate reporting exercise rather than as an input to the credit decision.

6.3.10 Summary of the Argument

Quantitative scoring is cheap, consistent, and scalable, but it is trained on the past and blind to transition risk that has not yet crystallized. Fundamental analysis can see forward and can assess what the score cannot measure, but it is expensive, inconsistent, and vulnerable to narrative. Neither alone can identify green credit reliably. A hybrid model—score as screen, flags as router, analyst as adjudicator, override log as learning mechanism—combines their strengths and contains their failures. The boundary between the two is a design choice that should move as data improves, and the governance that makes the choice deliberate is the difference between a system that learns and one that does not.

The next chapter turns to a risk that the hybrid model makes visible but cannot resolve on its own: concentration. A bank that identifies green borrowers well may still fail if its green portfolio is concentrated in a single technology, a single policy regime, or a single buyer. Identification and diversification are separate problems, and the second one is where the portfolio effects discussed next begin.

Before leaving the hybrid model, though, I want to make the case more concrete with a third example, because the two worked examples above are both clean cases where the analyst and the score disagree in a singledirection. Real files are messier, and the mess is instructive.

6.3.11 Worked Example: A Regional Bus Operator

A city bus operator seeks financing to replace 200 diesel buses with electric models. The operator is municipally owned, has a stable fare revenue base, and carries debt that is guaranteed by the municipal government. The score places the operator in the pass band with room to spare. The guarantee does most of the work; the operator's own coverage ratios are thin, but the municipal backstop makes the file look strong.

The environmental flags fire on the positive side. The fleet replacement is a genuine emissions reduction, and the file is a candidate for green bond classification under the local taxonomy. The analyst could simply confirm the score and move on.

The analyst's assessment complicates the picture. Physical exposure: the emissions reduction is real, but the electricity the buses draw is grid-supplied, and the grid's carbon intensity in the province is above the national average. The net reduction is smaller than the headline figure suggests. Policy exposure: the municipal guarantee is a contingent liability, and the municipality's own fiscal position has deteriorated as land-sale revenue has fallen. The guarantee is only as good as the guarantor. Transition capability: the operator has no experience with electric fleet maintenance, and the depot charging infrastructure is not yet built. The capital plan includes the buses but not the charging equipment, which is budgeted separately and not yet approved.

The analyst recommends approval with conditions: confirmation of the charging infrastructure budget, a covenant requiring disclosure of the grid emission factor used in the green bond reporting, and a review of the municipal guarantee against the municipality's updated fiscal position. The score would have approved the file without conditions. The override is recorded, coded to the guarantee-quality factor, and reviewed.

This case is messier than the first two because the analyst agrees with the score's direction but disagrees with its confidence. The hybrid model needs to accommodate that. The override protocol described above handles upgrades and downgrades, but the most common override in practice is neither: it is an approval with conditions that the score did not require. The override log should record conditions as a separate category, because conditions are where the analyst's forward-looking judgment is operationalized, and the performance of conditions is a direct test of whether the analyst's judgment was sound.

6.3.12 The Conditions as the Unit of Analysis

This observation deserves to be generalized. In green credit, the analyst's most useful output is often not a yes or a no but a set of conditions attached to a yes. The conditions are the mechanism through which forward-looking judgment enters the loan contract.

A well-designed condition has three properties. It is observable, meaning the borrower can report on it and the bank can verify it. It is material, meaning a breach would change the credit judgment. And it is proportionate, meaning the cost of compliance does not exceed the risk it addresses. A condition that requires a small borrower to commission an expensive third-party audit for a loan of modest size fails the proportionality test; the audit cost becomes a deadweight loss.

The conditions that work in practice tend to be reporting conditions rather than performance conditions. A performance condition—"the borrower shall reduce emissions intensity by 15 percent within three years"—is difficult to verify and harder to enforce, because the bank's remedy for breach is uncertain. A reporting condition—"the borrower shall report emissions intensity annually using the methodology in Schedule 4"—is verifiable, cheap, and gives the bank the information it needs to reprice or exit if the trend is adverse. The reporting condition also builds the data that the score will eventually use. This is the migration mechanism in miniature: the condition generates the measurement, the measurement enters the monitoring, the monitoring feeds the override log, and the override log informs the next re-estimation.

6.3.13 What the Score Should Be Asked to Do in Green Lending

There is a temptation, when a bank decides to build a green credit capability, to ask the score to do more than it can. The score is asked to identify green borrowers, to price transition risk, and to monitor environmental performance. It can do the first weakly, the second barely, and the third not at all.

A more defensible division of labor assigns the score three tasks and no more.

The first is to rank the book on conventional credit quality. This is what the score was built for, and it does it well.

The second is to flag segment membership. The score can carry a sector code and a green-taxonomy classification as inputs, not because these predict default but because they route the file. A sector flag is a routing device, not a risk weight.

The third is to provide a stable reference point against which overrides can be measured. The score's value in the hybrid model is partly that it does not move when the analyst's mood does. A bank that re-estimates its score every quarter in response to the latest policy news has lost the reference point.

Everything else—the forward-looking assessment, the transition capability judgment, the policy exposure analysis, the condition design—belongs to the analyst, supported by whatever data infrastructure the bank has built. Asking the score to do these things produces a model that is complex, opaque, and no more accurate than the analyst it was meant to replace.

6.3.14 The Cost Side of the Ledger

The hybrid model is not free, and the costs are worth stating because they determine whether it is worth building.

The first cost is analyst time. A structured green assessment of the kind described above takes two to five days per file for an experienced analyst. For a bank with a large SME book, this is prohibitive if applied universally. The routing logic exists precisely to concentrate the analyst's time on the files where judgment changes the outcome. A well-designed flag system should route perhaps 10 to 20 percent of the book to full assessment, with the rest handled by the score and a lighter-touch review.

The second cost is data. The conditions generate reporting obligations, and reporting obligations generate administrative cost for the borrower and verification cost for the bank. For small borrowers, the cost can be material relative to the loan. This is an argument for proportionality in condition design and for shared infrastructure—sector-level reporting templates, third-party verification services—that spreads the cost across many borrowers.

The third cost is governance. The override log, the outcome tracking, the annual boundary review, and the model inventory all require staff time and management attention. Banks that treat these as compliance overhead will do them badly. Banks that treat them as the learning mechanism will find that the hybrid model improves with use.

The fourth cost is the risk of false precision. A hybrid model produces a decision with a written rationale and a set of conditions, and the documentation can create an impression of rigor that the underlying data does not support. The antidote is to state the data quality explicitly in the assessment, to distinguish measured from estimated variables, and to record the confidence with which the analyst holds each judgment. A credit committee that reads "emissions intensity estimated from sector averages, confidence low" will treat the number differently from one that reads "emissions intensity measured and verified, confidence high." The distinction should be in the file.

6.3.15 The Comparative Question

A reader might reasonably ask why the hybrid model is the right structure, rather than a simpler alternative. Two alternatives deserve consideration.

The first alternative is pure judgment. Abandon the score for green lending and let experienced analysts decide. This is what many banks do in practice for new sectors, and it has the virtue of flexibility. It fails on scale and consistency, and it fails on learning, because unstructured judgment does not generate the data that would improve it. The override log is the minimal structure that makes judgment learnable.

The second alternative is pure model. Build a sufficiently rich score that incorporates environmental variables, transition indicators, and policy exposure, and let it decide. This is the direction that some fintech lenders have taken, and it can work where the data is dense and the outcomes are frequent—consumer lending, for example. In green SME lending, the data is sparse and the outcomes are slow. A pure model trained on sparse data will overfit to the sample and will be confident where it should be humble. The hybrid model is the appropriate structure for a domain where the data is thin and the stakes are long.

There is a third alternative that is sometimes proposed: outsource the assessment to an external ESG rating provider. This substitutes a third party's model for the bank's own, and it inherits the third party's blind spots without the bank's ability to correct them. External ratings are useful as a cross-check, particularly for large borrowers where multiple providers cover the same firm. They are not a substitute for the bank's own assessment, because the bank bears the credit risk and the rating agency does not.

6.3.16 The Learning Loop in Detail

The learning loop is the feature that distinguishes a hybrid model from a pair of disconnected tools, and it is worth describing its mechanics.

The loop has four stages. In the first stage, the analyst records an override with a reason code drawn from the flag taxonomy. The reason codes are fixed, which is what makes aggregation possible. A free-text rationale is also required, but the coded reason is the unit of analysis.

In the second stage, the override is matched to outcomes. The outcome measures are the standard ones: default, delinquency, rating migration, covenant breach. For green lending, there is a fourth outcome measure that matters: whether the environmental condition was met. A borrower that met its emissions reporting condition and one that did not are different observations, even if both repaid.

In the third stage, the override reasons are tested against outcomes. The test is simple: for each reason code, compare the outcome rate of overrides coded to that reason with the outcome rate of the score's prediction. A reason code that consistently predicts better outcomes than the score is a candidate for inclusion in the score. A reason code that consistently predicts worse outcomes is a candidate for removal from the override list, or for retraining the analysts who use it.

In the fourth stage, the boundary is adjusted. Variables migrate into the score, flags are retired or added, the override protocol is revised. The adjustment is annual, not continuous, because continuous adjustment destroys the reference point.

The loop is not complicated. What makes it work is that it is actually run. Many banks have the components—the override log, the outcome data, the annual model review—and never connect them. The connection requires that someone own the loop and that the results are reported to the credit committee, not buried in a risk report.

6.3.17 What Would Change My Mind

The argument for the hybrid model rests on a claim about the current state of green credit data: that it is too sparse and too noisy to support a pure model, and that analyst judgment is therefore necessary. If that claim is wrong, the argument weakens.

What would falsify it? Three findings would. First, if environmental variables were shown to add substantial and stable predictive power to conventional credit scores across a range of sectors and markets, the case for leaving the green dimension to the analyst would shrink. Second, if the override log showed that analyst overrides in green lending had no predictive value—that overridden files performed no differently from what the score predicted—the analyst's role would be questionable. Third, if the cost of analyst assessment fell sharply, through standardization or automation, the routing logic would matter less and the hybrid model would converge toward a richer score.

None of these findings is currently in hand. The evidence on environmental variables and credit risk is mixed. The override logs of most banks are not public, and the ones I have seen suggest that analyst overrides in green lending do carry information, particularly in sectors where policy is changing. The cost of assessment is falling, but slowly. The hybrid model is the right structure for the evidence we have.

6.3.18 The Chapter's Place in the Chain

This chapter sits between the rating framework of the previous chapter and the concentration analysis of the next. Its contribution to the book's argument is the identification step.

The rating framework described earlier makes environmental performance visible to the credit committee. This chapter describes the machinery through which that visibility is operationalized: the score that sorts the book, the flags that route the files, the analyst who adjudicates, and the log that makes the adjudication learnable. Without this machinery, the rating framework is a reporting exercise. With it, the framework enters the decision.

The identification step is necessary but not sufficient for the book's larger argument. A bank that identifies green borrowers well but concentrates its green portfolio in a single technology or a single policy regime has not solved the transition problem; it has taken a concentrated bet. The next chapter takes up that problem. The hybrid model described here makes the concentration visible, because the flag taxonomy and the override log record what the book is exposed to. But visibility is not diversification, and the portfolio-level tools are separate.

There is one connection to the later insurance chapters that is worth flagging now. The conditions attached to green loans—reporting conditions, covenant packages—are a form of risk buffering at the loan level. They do not transfer risk the way an insurance contract does, but they create information and contractual triggers that reduce the bank's exposure to adverse transition outcomes. The insurance instruments discussed later operate at a different level, buffering the borrower's physical and weather risk. The two mechanisms are complementary: the loan conditions buffer the bank against the borrower's transition failure, and the insurance buffers the borrower against the shocks that would cause such failure. A green credit system that has identification but neither form of buffering is exposed on both sides.

6.3.19 A Closing Observation on Method

The chapter has argued for a hybrid model on analytical grounds, and it has been candid about the gaps in the evidence. I want to close with a methodological observation that applies to the book as a whole.

The green transition is a phenomenon in which the past is an unreliable guide to the future. Credit models are machines for extrapolating from the past. This is not a temporary condition that better data will fix; it is a structural feature of a transition, which is by definition a period in which the relationships that governed the past are being deliberately changed by policy and technology. A credit system that relies entirely on historical models will be systematically late to the transition, because it will only recognize transition risk after it has crystallized into default.

The hybrid model is a partial answer to this structural problem. It does not predict the transition; it creates a structured space in which human judgment about the transition can enter the credit decision, and it records that judgment so that it can be tested. The record is the important part. Over time, the judgments that prove right become the variables that the score can use. The system learns because the organization has built a mechanism for turning judgment into data.

That mechanism is the chapter's real subject. The score, the flags, the analyst, and the log are components. The learning loop is the design. A bank that builds the components without the loop has built a more elaborate version of the single-channel lending that this book opened by criticizing. A bank that builds the loop has built something that can adapt as the transition unfolds.

The adaptation will be tested soon enough. The policy schedules now legislated in major markets—carbon prices, disclosure requirements, border adjustments—set deadlines in the late 2020s and early 2030s. The loans being underwritten today will mature into those deadlines. Whether the identification machinery described in this chapter is adequate will be known, not by argument, but by the performance of the loan books that used it. That is the test the hybrid model is designed to pass, and the override log is where the evidence will accumulate.