Chapter 5: Portfolio Dependence and Concentration in Green Lending
A loan officer approves a $40 million facility for a solar developer in Hebei. Three weeks later, a colleague on the same bank's Shanghai desk approves $60 million for an upstream polysilicon supplier that sells to that developer. Two months after that, the credit committee signs off on a $25 million working-capital line for a battery-storage integrator whose largest customer is, again, the same developer. On the bank's books these are three obligors, three ratings, three sets of legal documents. Economically they are one bet placed three times. If the developer's projects slip, the polysilicon supplier's receivables age and the integrator's order book empties. The bank has not diversified. It has scaled.
The rating framework of the previous chapter cannot solve this problem by itself. A rating judges one obligor against a horizon of one to three years. It says little about what happens to a book of two hundred obligors when a single policy variable—a feed-in-tariff revision, a change in green-bond taxonomy eligibility, a subsidy clawback—moves all of them at once. Green lending, because it is transitional and policy-contingent, carries dependence structures that differ from those of a conventional commercial portfolio. This chapter makes those structures visible, gives them a language, and traces what they imply for capital.
The claim is not that green lending is uniquely dangerous. It is that the correlation assumptions inside standard credit portfolio models were calibrated on industrial and commercial loan books whose sectoral composition and cyclical drivers differ from those of a green portfolio. Applied without adjustment, those assumptions understate tail losses. The chapter develops the claim in three movements: where concentration comes from, how to model the dependence it creates, and what bank treasurers and regulators should do about it.
5.1 Concentration Risk in Green Portfolios: Why Green Lending Clusters
5.1.1 The arithmetic of a narrow opportunity set
Start with a number. Industry—steel, cement, chemicals, plastics, aluminium, pulp and paper—accounted for roughly 31 percent of global emissions in 2020, when the total stood at about 51 billion tonnes of CO₂-equivalent. Add power generation and the share of the economy that must be rebuilt to reach any credible decarbonization pathway exceeds half of global emissions. That is the demand side for green finance, and it is enormous. The supply side is narrower than it looks.
A bank that wants to lend against the transition has a finite menu: renewable generation (solar, wind, hydro, geothermal), grid and storage infrastructure, electrified transport, energy-efficient buildings, industrial process change, the circular economy. Within each category the number of bankable obligors is small. Solar development in any single province is dominated by a handful of developers and EPC contractors. Battery manufacturing at scale sits in fewer than twenty firms globally. Green hydrogen, as of this writing, has almost no investment-grade borrowers at all. The opportunity set is not the whole economy; it is the thin leading edge of it.
Portfolio managers use "concentration" in two senses. Name concentration means large exposures to a small number of obligors, so that one default moves the loss distribution materially. Factor concentration means exposures spread across many names that share a common driver, so that they default together. Name concentration is visible on an exposure report. Factor concentration is not. Green lending is prone to both, and the second is the more dangerous because it is harder to see.
Consider a mid-sized commercial bank with a $12 billion loan book and a stated ambition to hold 15 percent of it in green assets by 2028—$1.8 billion. Suppose it builds that book across 140 obligors, an average exposure of $13 million. That sounds diversified. Now sort the 140 by revenue driver. Perhaps 45 are solar or wind project companies selling power under long-term contracts; 30 are equipment manufacturers selling into those projects; 25 are energy-efficiency service companies dependent on building-owner capex; 20 sit in the electric-vehicle supply chain; 20 are miscellaneous. The first two groups—75 obligors, more than half the book—share one demand driver: the pace of renewable capacity additions, which depends on auction schedules, grid-connection queues, and the cost of capital. A shock to that driver does not produce 75 independent defaults drawn from a binomial distribution. It produces a cluster.
5.1.2 Why the clustering is structural, not accidental
This clustering is not a temporary artifact of a young asset class that will diversify away as the market matures. It is built into how green projects become bankable, and it will persist.
Green projects are bankable largely because policy makes them so: feed-in tariffs, renewable portfolio standards, carbon prices, tax credits, concessional finance, procurement mandates. These instruments are set at the level of a jurisdiction and they move in steps. When Spain retroactively cut its solar feed-in tariff in 2013, it did not impair one developer; it impaired the asset class within Spain. When China's "531 New Policy" capped subsidized solar capacity in 2018, it compressed margins across the entire domestic supply chain within a quarter. A portfolio diversified across fifty Spanish solar developers in 2012 was, in factor terms, a single exposure to the Spanish government's willingness to honor a tariff. The same logic now applies to EU taxonomy eligibility rules, to the implementation timeline of the Carbon Border Adjustment Mechanism, and to the subsidy regimes that underwrite green hydrogen offtake.
Technological lock-in pushes the same way. Green technologies are not interchangeable. A bank that has built expertise in solar project finance has built underwriting templates, engineering review relationships, and servicing capacity specific to solar. That expertise is a real asset and it creates a real bias: the bank originates more solar because it can underwrite solar better than it can underwrite offshore wind or industrial heat pumps. Specialization improves individual credit decisions and worsens the correlation of the resulting portfolio. The shipping and commercial real estate books of the 2000s were built the same way. Expertise is a concentration engine.
Value-chain structure supplies a third push. Green industries are vertically shallow. A solar project depends on modules, inverters, trackers, and grid connection; each of those depends on a small number of suppliers, many in a single jurisdiction. A bank lending across the chain—to the developer, the EPC, the module maker, the polysilicon refiner—believes it has diversified by lending at different points. It has taken four positions on the same cash flow, levered at each link. When the developer's payment terms stretch from 60 to 120 days, the strain propagates upstream with a lag of one to two quarters. The obligors do not default simultaneously. They default conditionally, and the condition is shared.
State the contrast plainly. In a conventional commercial portfolio, the default correlation between two randomly chosen obligors is low because their revenue drivers are largely idiosyncratic; a restaurant and a machine shop fail for different reasons. In a green portfolio, the default correlation between two randomly chosen obligors is elevated because their revenue drivers are largely common. Two solar developers fail for the same reason, and a solar developer and its module supplier fail for the same reason one quarter apart. Each obligor's rating can be correct and the portfolio can still be far riskier than the sum of its ratings suggests.
5.1.3 A worked illustration
Three firms from the previous chapter's universe, plus the link between them.
Firm A is a solar independent power producer with 400 MW of operating capacity in one province, contracted under a 20-year tariff with a state-owned offtaker. Under the framework its rating is BBB−. Firm B is the EPC contractor that built three of Firm A's plants and holds a two-year operations-and-maintenance contract on them; its rating is BB. Firm C supplies modules to Firm B, with 40 percent of its revenue from that relationship; its rating is B+.
Standalone, the ratings form a ladder: investment grade at the bottom, speculative grade above it. Portfolio-wise the ladder is an illusion. Firm B's cash flow depends on Firm A's construction pipeline and on the timeliness of Firm A's payments. Firm C's cash flow depends on Firm B's orders. The default probability of B conditional on A's distress is not the unconditional probability implied by its BB rating; it is substantially higher. The same holds for C conditional on B. If the bank holds all three, the effective number of independent bets is closer to one than to three.
How much closer? Take two obligors each with a one-year default probability of 2 percent. Under independence the joint default probability is 0.04 percent—four in ten thousand. A default correlation of 0.3, not extreme for firms in the same value chain, raises the joint probability to roughly 0.5 percent—fifty in ten thousand. The difference between those numbers is the difference between a portfolio that can be managed with standard capital and one that cannot. Section 5.2 formalizes the calculation.
5.1.4 The regulatory blind spot
Prudential regulation has treated concentration mainly as a name-level phenomenon. Large-exposure rules cap the exposure to any single obligor or connected group as a fraction of capital—typically 25 percent under the Basel framework and the EU's Capital Requirements Regulation. These rules are necessary and they do not address the problem. A bank can hold 200 obligors, each at 0.3 percent of capital, and still be concentrated if those obligors share a driver.
The Internal Ratings-Based (IRB) approach to credit risk does incorporate a systematic factor through the asset correlation parameter, examined in Section 5.3. But the correlation is assigned by exposure class—corporate, retail, sovereign—and within the corporate class it is calibrated to a formula that depends primarily on obligor size, not on sector or on the nature of the borrower's revenue. Two mid-sized corporates in unrelated industries receive the same asset correlation as two mid-sized corporates in the same value chain. The model assumes away the dependence that green lending creates.
This is not a criticism of regulators' competence. It describes a calibration choice made when the corporate loan book looked different. The Basel II foundation IRB formula set corporate asset correlation at between 12 and 24 percent, declining with firm size, based on empirical work on default correlations in broadly diversified corporate portfolios. Those estimates were reasonable for the portfolios observed. They are less reasonable for a portfolio whose obligors are selected precisely because they share a policy-driven revenue driver. A bank that applies the standard formula to a concentrated green book will hold capital against idiosyncratic risk and under-hold against systematic risk. Expected loss will be roughly right; unexpected loss, which is what capital is for, will be too low.
One objection to anticipate: if green lending is concentrated and dependence is high, does that argue against doing it—against financing the transition through bank balance sheets at all? It does not. It argues for pricing and structuring the concentration. A bank that understands its dependence structure can hold more capital against the concentrated portion, syndicate the name-level risk, buy credit protection on the systematic component, and diversify across jurisdictions whose policy cycles are not synchronized. A bank that does not understand it will do none of these things and will discover the dependence in a downturn. The rest of this chapter builds the understanding.
5.2 Modeling Dependence Structures
5.2.1 Why marginal distributions are not enough
Standard credit risk measurement produces, for each obligor, a probability of default over a horizon and a loss given default. Multiply the two for expected loss; sum across obligors for the portfolio's expected loss. That sum is correct regardless of dependence, because expected loss is linear in the default indicators and requires no assumption about how defaults co-move.
Everything that matters for capital lives in the tail, and the tail is where dependence bites. The 99.9th percentile of the portfolio loss distribution—the number Basel's IRB approach targets—depends entirely on the joint behavior of defaults. Two portfolios with identical obligor-level default probabilities and identical loss given defaults can have 99.9th-percentile losses that differ by a factor of three or more, depending on whether defaults are independent or clustered. Marginal distributions do not determine the joint distribution. A dependence model is required.
The simplest is the single-factor model, in which each obligor's asset value is driven by a common systematic factor and an idiosyncratic shock:
where \(Z\) is the systematic factor (standard normal), \(\varepsilon_i\) is the idiosyncratic shock (standard normal, independent across obligors and of \(Z\)), and \(\rho_i\) is the asset correlation—the share of obligor \(i\)'s variance attributable to the common factor. Obligor \(i\) defaults if \(A_i\) falls below a threshold \(c_i\) calibrated to its unconditional default probability \(p_i\), so that \(c_i = \Phi^{-1}(p_i)\).
This is the model embedded in the Basel IRB formula. It is tractable and has a clear economic reading: \(\rho_i\) measures how much of each borrower's fortune is tied to the common cycle. The conditional default probability, given a realization of \(Z\), is
and the portfolio's asymptotic loss in the worst state is obtained by evaluating this at a low quantile of \(Z\). The formula is elegant and its weakness is structural: it imposes a single common factor and a single correlation parameter per obligor. It cannot represent what matters most in a green portfolio—a value chain in which A's distress raises B's default probability directly, over and above any common macro factor.
5.2.2 Copulas: separating marginals from dependence
The copula approach separates the marginal default distributions from the dependence structure. Sklar's theorem states that any joint distribution \(F(x_1, \ldots, x_n)\) can be written as
where \(F_i\) are the marginal distributions and \(C\) is a copula—a joint distribution on the unit hypercube with uniform marginals. The copula carries all the dependence information. Change the copula, keep the marginals, and the tail behavior changes while the individual default probabilities do not.
Three copula families are used for credit portfolios. The Gaussian copula underlies the single-factor model above and is the default in most commercial credit portfolio systems. It has no tail dependence: in the extreme tail, the probability that two obligors default together goes to zero faster than the marginal probabilities. For correlated credit risk that assumption is strong and empirically questionable. The Student-\(t\) copula adds tail dependence through a degrees-of-freedom parameter \(\nu\); as \(\nu\) falls, joint extreme events become more likely. The Clayton and Gumbel copulas exhibit asymmetric tail dependence—Clayton in the lower tail, Gumbel in the upper—and are useful when the economic mechanism produces clustering in one direction only.
The practical difference is large. Take two obligors each with \(p = 1\%\) over one year. Under a Gaussian copula with correlation 0.3, the joint default probability is approximately 0.08 percent. Under a Student-\(t\) copula with the same linear correlation and \(\nu = 4\), it rises to roughly 0.25 percent—three times higher—because the \(t\) copula assigns more mass to the joint lower tail. The marginal default probabilities are identical; only the dependence assumption differs. A bank that calibrates capital to a Gaussian copula and lives in a world with tail dependence will be systematically short of capital in exactly the states where capital is needed.
For green portfolios the case for tail-dependent copulas is strong. The mechanisms that link green obligors—policy shifts, subsidy revisions, technology-cost curves, grid constraints—are not smooth, continuous processes. They move in discrete steps, and a step affects all linked obligors at once. That is the signature of tail dependence. The empirical literature on credit default swaps during the 2008 crisis and the 2011 European sovereign stress found the same pattern: joint defaults clustered far more than Gaussian models predicted. There is no reason to expect green lending to be gentler.
5.2.3 The limitations of copulas, and what to do about them
Be honest about what copulas cannot do. A copula is a static object. It describes dependence between default indicators over a fixed horizon; it does not describe how dependence evolves as obligors move toward or away from default. Nor does it tell us which copula is correct. The choice among Gaussian, \(t\), Clayton, and Gumbel is a modeling judgment, and different choices produce materially different capital numbers from the same data. Calibration is data-hungry: estimating a \(t\) copula's degrees of freedom requires a history of joint defaults that most banks, especially in green sectors, do not have.
Three responses are available. Use the copula as a stress-testing tool rather than a capital-calibration tool: run the portfolio under several copulas and report the range of tail losses, so management sees the sensitivity rather than a false precision. Supplement the copula with structural information from the value chain—who sells to whom, who depends on which policy—and encode that information as constraints on the correlation matrix rather than estimating it freely. Move to dynamic models in which dependence itself is state-dependent, rising as obligors approach distress; the intensity-based and threshold models of the credit literature do this at the cost of tractability.
The practical recommendation is a layered approach. Use the single-factor model as the baseline, because it is what the regulator uses and what the bank's systems can run. Layer a \(t\) copula on top to capture tail dependence. Constrain the correlation matrix using the value-chain map. Report the results as a range. The purpose is not a single number that is "right"; it is a number that is defensible and a range that is honest.
5.2.4 A value-chain dependence map
The most useful input to any of these models is a map of who depends on whom. This is where a bank's credit officers know more than its quants, and where that knowledge should be captured explicitly rather than left in individual heads.
flowchart TD
P["Policy driver: tariff schedule,<br/>taxonomy eligibility, subsidy cap"] --> A["Solar IPP<br/>Developer A (BBB-)"]
P --> B["EPC contractor<br/>Firm B (BB)"]
P --> D["Grid operator<br/>Offtaker"]
A -->|"construction contracts"| B
A -->|"power purchase"| D
B -->|"module supply"| C["Module supplier<br/>Firm C (B+)"]
B -->|"inverter supply"| E["Inverter maker<br/>Firm E (BB-)"]
C -->|"polysilicon"| F["Polysilicon refiner<br/>Firm F (B)"]
A -->|"O&M contract"| B
style P fill:#f4e6c8,stroke:#8a6d2f
style A fill:#d8e8d0,stroke:#3f6b32
style B fill:#d8e8d0,stroke:#3f6b32
style C fill:#e8d8d0,stroke:#7a4a32
style D fill:#d0dce8,stroke:#324a6b
style E fill:#d8e8d0,stroke:#3f6b32
style F fill:#e8d8d0,stroke:#7a4a32
Figure 5.1: A value-chain dependence map for a solar portfolio. Arrows indicate cash-flow dependence; the policy node drives all downstream nodes.
The map does two things. It identifies the policy node as the common driver, which tells the modeler that the systematic factor should be read as a policy factor, not a generic business-cycle factor. And it identifies the directed edges—A to B, B to C, C to F—which tell the modeler that the correlation between A and F is not a free parameter to be estimated but a consequence of the chain. In a Gaussian or \(t\) copula the correlation matrix can be constrained so that \(\rho_{AF} \approx \rho_{AB} \cdot \rho_{BC} \cdot \rho_{CF}\), the product of the chain's links. That is a strong constraint, and it is more defensible than an unconstrained estimate from three years of sparse data.
5.2.5 Estimating the parameters when data are scarce
Green lending has a short history. A bank that began building its green book in 2018 has at most six or seven years of internal default experience, and in most sectors it has zero defaults—good news about credit quality, bad news about statistical power. If no defaults have occurred in a portfolio of 140 obligors over six years, the default rate is estimated at zero with a confidence interval extending to roughly 2 percent per annum. That interval is too wide to calibrate anything.
External ratings studies can anchor the marginals. Fitch Ratings' 2023 Annual Global Corporate Default and Rating Transition Study reports average one-year default rates by rating grade over a long horizon; for the period 1990–2023 the average one-year default rate for a BBB-rated corporate was about 0.2 percent, and for a B-rated corporate about 3.5 percent, with wide year-to-year variation (the B-rated cohort exceeded 8 percent in 2009 and fell below 1 percent in several benign years). Moody's Annual Default Study for the same period gives comparable figures—roughly 0.18 percent for Baa and 3.4 percent for B. These are marginal default probabilities, not joint ones. They anchor the marginals while the copula is calibrated separately.
Equity-based measures can proxy the dependence. Default correlation can be estimated from the correlation of equity returns, mapped through a Merton-style structural model to asset-value correlation. For listed obligors this is feasible; for the private project companies that dominate green lending it is not. A workable compromise: estimate the asset correlation for the listed peers of each private obligor and apply it as a proxy, adjusted upward for the private obligor's smaller size and narrower revenue base. The adjustment matters. A project company with one plant and one offtaker has a revenue base closer to a single asset than to the diversified peer group from which the proxy is drawn, and its true correlation with the policy node is correspondingly higher.
Expert judgment, structured through a scorecard, is the third substitute. Credit officers who have worked a value chain for a decade can rank pairs of obligors by dependence with reasonable consistency. Eliciting those rankings and converting them to a correlation matrix—subject to the constraint that the matrix be positive semi-definite—is not rigorous in the econometric sense, but it beats assuming zero correlation, which is the implicit alternative in most systems.
A worked example shows how far the three substitutes can carry a real calibration. Take the 140-obligor book from Section 5.1.1 and the value-chain map of Figure 5.1. Suppose the bank has internal default data for 2019–2024: two defaults, both in 2022, both EPC contractors linked to the same developer. The raw internal default rate is 2/140 ≈ 1.4 percent over six years, or about 0.24 percent per annum, but with two events the 95 percent confidence interval runs from 0.03 to 0.85 percent per annum—useless on its own. The external anchor supplies the level: the book's weighted-average rating is BB, and the Fitch and Moody's studies put the BB one-year default rate at roughly 1.2–1.5 percent over the long run. The two internal defaults, both in the same value chain in the same year, supply the dependence signal: they are consistent with a joint default probability far above the product of the marginals, and they point to a \(t\) copula with low degrees of freedom rather than a Gaussian one. The equity proxy supplies the correlation level for the listed peers of the project companies—typically 0.35 to 0.5 for solar developers in the same jurisdiction—and the expert scorecard supplies the upward adjustment for the private obligors. None of the three substitutes is decisive alone. Together they bracket the parameter: asset correlation for the value-chain-linked obligors plausibly lies between 0.35 and 0.55, and the joint default probability for two adjacent links lies between 0.3 and 0.8 percent, against 0.02 percent under independence.
The honest summary: dependence parameters in green portfolios will be estimated with wide error bars for some years to come. The response is to widen the stress range, not to pretend to precision. A capital number with a confidence interval is more useful than a point estimate that conceals its own uncertainty.
5.2.6 From dependence to loss distribution
With marginals and copula specified, the portfolio loss distribution follows by simulation or analytical approximation. The simulation is straightforward: draw the systematic factor, draw idiosyncratic shocks conditional on it, determine which obligors default, apply loss given default, sum, repeat ten thousand times. The result is a distribution of portfolio losses from which the 99.9th percentile—economic capital—can be read.
The shape of that distribution is the point. In a well-diversified portfolio the loss distribution is right-skewed with a thin tail; the 99.9th percentile is perhaps three to five times expected loss. In a concentrated portfolio with tail dependence the distribution develops a fat right tail; the 99.9th percentile can be ten to twenty times expected loss, and the gap between the 99th and 99.9th percentiles widens. The difference between those shapes is not a modeling detail. It is the difference between a portfolio that can absorb a bad year out of earnings and one that cannot.
5.3 Implications for Bank Capital and Regulation
5.3.1 The capital gap under standard formulas
The Basel IRB corporate formula computes the capital requirement for a single exposure as
where \(\mathrm{M}_i\) is the maturity adjustment. The critical parameter is \(\rho_i\), the asset correlation. Under the foundation IRB approach, \(\rho_i\) is set by a formula that declines with firm size:
with a size adjustment that subtracts up to 0.04 for firms with turnover below €50 million. For a mid-sized corporate with a 1 percent default probability, the formula produces an asset correlation of roughly 0.17. That number was calibrated on a broad corporate portfolio and embeds an assumption of average diversification.
Apply it to the solar value chain. The formula gives Firm A, Firm B, and Firm C roughly the same asset correlation because they are of similar size and similar default probability. The true dependence among them, as the value-chain map shows, is much higher than 0.17—plausibly 0.4 to 0.6 for adjacent links, and 0.3 or more for the chain as a whole. The formula understates the systematic risk. The capital charge computed from it covers the average loss and not the tail.
A rough calculation fixes the order of magnitude. For a single obligor with \(p = 1\%\), \(\mathrm{LGD} = 45\%\), and maturity of 2.5 years, the IRB formula with \(\rho = 0.17\) produces a capital requirement of about 4.5 percent of exposure. Raising \(\rho\) to 0.45—plausible for a value-chain-linked obligor—raises the requirement to about 9.5 percent. The capital need roughly doubles. For a $1.8 billion green book the difference is on the order of $90 million of capital. That is not a rounding error; it is the difference between a book that meets its return-on-equity target and one that does not.
Be careful about the interpretation. The calculation does not say the bank is undercapitalized in an absolute sense; it says the standard formula, applied without adjustment, misallocates capital within the book. Green exposures are charged too little and, by implication, non-green exposures are charged too much relative to their true risk. The misallocation has a behavioral consequence: it makes green lending look more profitable than it is, which encourages more of it, which increases the concentration. The capital formula is not a neutral measuring device. It shapes the portfolio it measures.
5.3.2 What regulators can do, and what they are doing
A sectoral correlation add-on would require banks to apply a multiplier to the asset correlation for exposures in identified green sectors, or to use a higher correlation floor. The precedent exists in the IRB treatment of large financial-sector exposures, where a multiplier of 1.25 is applied to asset correlation to reflect the higher interconnectedness of financial firms. An analogous multiplier for concentrated green value chains—say 1.3 to 1.5—would be a targeted, transparent adjustment. It would raise capital on the concentrated portion without penalizing diversified green lending, and it would give banks an incentive to demonstrate diversification in order to avoid the multiplier.
The concentration risk charge under Pillar 2 is the second lever. Basel's Pillar 2 requires banks to hold capital against risks not captured by Pillar 1, including concentration risk. In practice many banks compute a single-name concentration charge and stop. A supervisor that took factor concentration seriously would require banks to map their green portfolios by driver, report the share of the book exposed to each policy or technology node, and hold Pillar 2 capital against the resulting concentration. The supervisory review process already has the authority; what is missing is the analytical template. Publishing such a template—even as non-binding guidance—would move the practice.
Disclosure is the third. A bank that must disclose the sectoral and driver-level composition of its green book, alongside its dependence assumptions, gives the market the information to price the concentration. Disclosure does not by itself reduce risk, but it creates a cost to holding undiversified exposure that the bank internalizes. The precedent is the post-crisis disclosure of securitization exposures, which did not prevent the next concentration but did make the previous one visible.
None of these levers is a complete solution. Each addresses a piece of the problem. The combination—a correlation multiplier for concentrated sectors, a Pillar 2 charge for factor concentration, driver-level disclosure—would close most of the gap without requiring a wholesale rewrite of the capital framework.
5.3.3 Portfolio management responses short of regulation
Banks do not need to wait for regulators. Four responses are available within the current framework, and a well-run green lender should be using all of them.
Diversification across policy jurisdictions is the cheapest and the most often neglected, because banks build green books in their home market first. The dependence that matters most is policy dependence, and policy cycles are not synchronized across countries. A portfolio split between Chinese, European, and North American green assets has lower dependence than the same exposure concentrated in one jurisdiction, because a tariff revision in one does not move the others.
Diversification across technology maturity works on a different axis. Solar and onshore wind are mature, subsidy-dependent, and highly correlated with each other. Green hydrogen and long-duration storage are immature, subsidy-dependent in a different way, and less correlated with the mature technologies because their cost curves and policy support sit at different stages. A portfolio weighted toward mature technologies has lower marginal default probabilities and higher dependence; a mix has higher marginals and lower dependence. Neither dominates; the mix should reflect the bank's capital and its appetite for tail risk.
Risk transfer is the third response. Credit default swaps on green obligors are thin but not absent, and where they exist they let the bank retain the relationship and shed the concentration. Synthetic securitization—the sale of the junior tranche of a reference portfolio—is more efficient for a book of mid-sized obligors, because it transfers the tail risk on a pool rather than on a single name. The bank retains the senior tranche, the part least affected by dependence, and sells the first-loss piece to investors paid to bear it.
Syndication at origination is the fourth. A bank that originates a $60 million facility and syndicates $45 million of it to three other banks has reduced its name concentration without reducing its franchise. The syndication market for green assets is deeper than the CDS market, and the pricing is transparent enough to reveal what the market thinks the dependence is worth. A bank that consistently syndicates the concentrated portion of its green book is, in effect, using the market to price the risk it cannot model.
5.3.4 A numerical stress: the 2018 policy shock
In May 2018 China's National Development and Reform Commission, together with the Ministry of Finance and the National Energy Administration, announced what became known as the "531 New Policy": a cap on subsidized solar capacity, a reduction in the feed-in tariff, and tighter qualification rules for distributed generation. The policy was announced on a Friday and took effect the following Monday.
Consider a bank that in April 2018 held a $2 billion solar portfolio in China spread across 60 obligors: 25 project companies, 20 EPC contractors, 10 module and component suppliers, 5 polysilicon refiners. A naive concentration analysis would report a Herfindahl-Hirschman Index of roughly \(1/60 \approx 0.017\) on an equal-weighted basis—low concentration by the standard threshold of 0.1. A name-level large-exposure check would find no single obligor above 3 percent of the portfolio. By both conventional measures the book looked diversified.
The policy shock revealed the true structure. Within one quarter the project companies' revenue expectations were cut, the EPC contractors' order books shrank, the suppliers' receivables aged, and the refiners' margins compressed. Defaults did not occur all at once—the project companies had long-term contracts that cushioned the first year—but rating migrations did. A portfolio that had been 60 percent investment grade in April was less than 35 percent investment grade by the following April, and the downgrades concentrated in the value-chain links closest to the policy node.
For the bank, the loss was not the sum of 60 independent losses. It was closer to a single loss on a single factor, realized with a lag across the chain. A Gaussian copula calibrated to the pre-shock data would have assigned the joint event a probability in the low single digits. The event happened. That is what tail dependence means: the model says it cannot happen, and it happens.
The lesson is not that the bank should have avoided Chinese solar. It is that the bank should have measured its exposure to the policy node—the share of the book whose revenue depended on the subsidy schedule—and should have held capital, or bought protection, against a step change in that node. The measurement is not hard. The policy node was known. The value-chain links were known. What was missing was the discipline to map them and the capital framework that would have charged for them.
5.4 What This Means for the Rating-Insurance-Catalysis Mechanism
Rating produces an obligor-level assessment. It cannot see portfolio dependence. A bank that rates every green obligor correctly and ignores the dependence among them will still be surprised, because the surprise is a portfolio event. The rating framework needs a portfolio layer—a dependence map, a copula, a capital calculation—to be operational at the book level.
Insurance, developed over the next three chapters, is the natural instrument for the systematic component. A weather index insurance contract on a solar portfolio pays out when irradiation falls below a threshold, regardless of which individual project underperforms. A forest insurance contract pays out when fire or drought affects a region, regardless of which stand burns. These instruments transfer the factor risk the bank cannot diversify, and they price it at the factor's frequency rather than the obligor's idiosyncrasy. The connection is direct: the dependence that makes green portfolios risky is the same dependence index insurance is designed to hedge. A bank holding both a concentrated green loan book and an index insurance overlay has separated the idiosyncratic risk, which it can diversify, from the systematic risk, which it has transferred.
Catalysis operates at a different level. The sci-tech finance intermediaries of Chapter 10 channel capital toward green innovation by bearing early-stage risk that banks cannot. Their contribution to the concentration problem is to widen the opportunity set: as new technologies mature and new obligors become bankable, the menu from which a bank can build a diversified green book expands. The concentration of 2024 is partly a function of the small number of bankable green firms. That number will grow, and the growth is what catalysis produces.
The sequence matters. Rating identifies the obligor. Insurance buffers the factor. Catalysis widens the set. A financial system that does all three can support the transition without accumulating a concentration that undermines the banks doing the supporting. A system that does only the first will lend, concentrate, and discover the dependence in the next policy shock.
5.5 Summary and Bridge
Four propositions carry the argument. Green lending is structurally concentrated because the bankable opportunity set is narrow, because expertise is specialized, and because value chains are vertically shallow. That concentration is factor concentration, not name concentration alone, and it is invisible to standard exposure reports. Standard credit portfolio models, calibrated on diversified corporate books, understate the tail losses that factor concentration produces. The understatement has a capital consequence: green exposures are charged too little, which encourages more concentration, which increases the gap.
The responses are available. At the modeling level, a value-chain dependence map, a tail-dependent copula, and a simulation-based loss distribution give a defensible range. At the capital level, a sectoral correlation multiplier, a Pillar 2charge for factor concentration, and driver-level disclosure would close most of the gap. At the portfolio level, diversification across jurisdictions and technology maturities, risk transfer through CDS and synthetic securitization, and syndication at origination reduce the concentration without abandoning the franchise.
What the chapter has not done is supply the instrument that makes the systematic component hedgeable. That is the subject of the next three chapters. One methodological point governs how the insurance argument should be read, so state it before moving on.
The dependence modeled here is not a defect of green lending that better data will eliminate. It is a property of a transitional asset class whose returns are underwritten by policy. A bank that waits for the dependence to disappear before building its green book will wait a long time, and will build the book anyway when competitive pressure forces it, with worse information. The alternative is to measure the dependence now, price it, hedge what can be hedged, and hold capital against what cannot. A bank that measures the dependence today can put it into its loan spreads next quarter; one that waits will put it into its loss provisions after the first policy shock.
The next chapter takes up insurance as the buffering layer. The connection to carry forward is precise: the factor a copula captures statistically is the same factor an index insurance contract transfers contractually. If the copula says the portfolio's tail loss is driven by a policy node or a weather variable, and the insurance contract pays out on that variable, the two instruments describe and address the same risk. The design question is whether the contract's trigger matches the factor the portfolio is exposed to. When it does, the buffering works. When it does not, the bank has bought protection against a risk it does not hold and remains exposed to the one it does. Chapter 6 develops the matching criterion: the trigger must be defined on the same variable that loads the systematic factor in the portfolio model, measured over the same horizon, with a payout schedule that reproduces the loss the factor causes. The first test of that criterion is the weather-index contract for solar generation, and it is where Chapter 6 begins.