Chapter 15: Policy Implications and the Research Agenda

A provincial loan officer we interviewed in 2022 kept two spreadsheets on her desk. One listed her technology borrowers by credit score. The other listed the same firms by environmental compliance status, a column she had built herself after a regional pilot required banks to report "green" exposure. The two disagreed more often than they agreed. A battery-materials firm with a strong score on the first sheet sat near the bottom of the second because its wastewater discharge permit had lapsed during an expansion. A small software company with no environmental obligations at all scored high on the second sheet and mediocre on the first. She had no rule for reconciling them, so she did what most loan officers do when confronted with two conflicting signals: she lent on the strength of the first and filed the second for reporting purposes.

That scene is the practical form of the problem this book has addressed. The three instruments we have examined—environmental performance rating, insurance buffering, and sci-tech finance catalysis—each work in isolation. The question for policy is how to make them work in sequence, and the question for research is how to know whether the sequence works. The recommendations below are organized around the integrated mechanism. They are followed by an account of what our evidence cannot yet establish and where the research agenda should go.

15.1 Policy Recommendations for the Integrated Mechanism

15.1.1 The coordination principle

"Integrated" does not mean merging three instruments into one institution. It means aligning the information each instrument produces so that a decision made at one stage is legible at the next. A rating that cannot be read by an underwriter is wasted. An insurance contract whose payouts are invisible to a lender cannot reduce default probability in the lender's model. A catalysis fund that does not know which firms have been rated and which have been insured is allocating capital blind.

The policy principle that follows is coordination on data standards, not coordination on organizational charts. Governments should not create a single "green finance agency" that rates, insures, and invests. They should require that the outputs of each function share a common vocabulary. Three things must be standardized: the definition of an environmental performance score and its components; the reporting format for insurance coverage attached to a loan; and the disclosure of environmental performance by firms receiving public catalysis funds. Each is a modest regulatory act. Together they make the sequence possible.

Figure 15-1 shows the sequence and the policy lever at each link.

flowchart TD
    A[Firm applies for credit] --> B[Environmental performance rating]
    B --> C{Rating threshold met?}
    C -->|No| D[Technical assistance / rating improvement plan]
    D --> B
    C -->|Yes| E[Insurance underwriting]
    E --> F[Coverage attached to loan]
    F --> G[Lender prices default probability net of coverage]
    G --> H{Innovation-intensive?}
    H -->|Yes| I[Catalysis fund co-investment or guarantee]
    H -->|No| J[Standard green loan]
    I --> K[Portfolio monitoring]
    J --> K
    K --> L[Performance data feeds back to rating]
    L --> B

The loop at the bottom is the part most policy designs omit. Rating, insurance, and catalysis each generate data during the life of a loan. If that data does not return to the rating model, the system learns nothing and the second cohort of borrowers is rated with the same stale parameters as the first. A reporting requirement that closes this loop is worth more than a subsidy.

15.1.2 Rating: make the score public, the methodology auditable

Environmental performance rating should be treated as a public good in its basic form and a private good in its application. The basic score—whether a firm meets, exceeds, or falls short of environmental obligations—should be publicly available, because the cost of producing it is borne largely by public data systems (satellite monitoring, discharge registries, energy statistics) and because its diffusion reduces duplicative due diligence across lenders. The applied score—how a particular lender weights environmental performance against cash flow and collateral—should remain private, because that weighting is a commercial judgment.

This split has a precedent in credit reporting. Public credit registries in most countries distribute basic payment history; private bureaus and internal models combine that history with proprietary data to produce scores. The environmental analogue would be a public environmental registry that any lender can query, plus a requirement that lenders using the registry disclose the weight they assign to it. Disclosure of weights is the audit mechanism. If a bank claims to price environmental risk but assigns it a weight of 2 percent, the claim is testable.

Two design details matter. First, the rating must be firm-level and facility-level. A firm with three plants can have a clean headquarters and a non-compliant production line. Aggregation to the firm hides exactly the variation that predicts transition risk. Second, the rating must carry a timestamp and a confidence interval. Environmental performance changes with the compliance cycle; a rating from eighteen months ago is a different object from a rating from last quarter. Our interviews with loan officers found that the single most common complaint about existing green taxonomies was not their stringency but their staleness.

The methodology should be auditable in the narrow sense that a third party can reproduce the score from the underlying data. This rules out black-box machine learning models for the public score, however accurate they may be, and argues for a transparent index with published component weights. Private lenders can layer their own models on top. The public score's job is to be a common denominator, not a competitive advantage.

A worked example. Consider a mid-sized battery-materials plant in a coastal province, applying for a four-year working-capital facility. The public registry returns a facility-level score built from four components, each normalized to a 0–100 scale and weighted as follows:

The weighted score is \(0.35 \times 41 + 0.25 \times 78 + 0.20 \times 62 + 0.20 \times 95 = 14.35 + 19.50 + 12.40 + 19.00 = 65.25\). The registry publishes the score with a 90 percent confidence interval of [58, 72], reflecting the measurement error in the satellite emissions estimate and the 47-day compliance gap. The interval width is itself informative: a facility with a clean inspection history and continuous monitoring carries a narrower interval than one whose score depends on self-reported data.

The lender's internal model then applies its own weight to this score. Suppose the bank assigns environmental performance a weight of 12 percent in its default model, alongside cash flow (40 percent), collateral (25 percent), sector outlook (15 percent), and management quality (8 percent). The facility's score of 65.25, mapped to a default-probability adjustment, raises its estimated one-year default probability from 1.8 to 2.3 percent. That 0.5-point adjustment is what the disclosure requirement makes visible: the bank must report that it applied a 12 percent weight and that the weight moved the estimate by 0.5 points. A regulator reviewing the disclosure can ask whether 12 percent is consistent with the bank's stated green-lending strategy. The bank cannot answer with a press release.

The confidence interval matters for pricing. A facility scoring 65 with an interval of [58, 72] is a different credit from one scoring 65 with an interval of [63, 67]. The first is a firm whose environmental status is genuinely uncertain; the second is a firm whose status is measured well. Lenders that ignore interval width will systematically underprice the first type. Requiring the interval to be published forces the lender to confront the uncertainty rather than collapse it to a point estimate.

15.1.3 Insurance: subsidize the instrument, not the premium

The case for public involvement in weather index insurance and forest insurance rests on a market failure that is specific and well documented: the correlation between the insured event and the insurer's own exposure. A drought that triggers payouts across an entire province simultaneously is not diversifiable within that province. Private insurers facing such correlated risk either withdraw or charge a premium that low-margin agricultural borrowers cannot pay. The standard remedy is a public reinsurance layer or a sovereign risk pool that absorbs the tail.

The policy recommendation here is to subsidize the instrument's infrastructure rather than the policyholder's premium. Premium subsidies are politically attractive and administratively simple, but they distort the price signal that should tell farmers and foresters how much risk they actually face. They also tend to be captured by larger operators. Subsidizing the infrastructure—weather stations, index design, claims validation systems, reinsurance capacity—lowers the cost of the product without severing the link between risk and price.

There is a counterargument worth taking seriously. If the goal is rapid adoption, premium subsidies work faster. Our reading of the evidence is that speed of adoption is less important than persistence of adoption. A farmer who buys insurance because it is cheap and drops it when the subsidy ends has not been insured in any meaningful sense. A farmer who buys it because the index reflects his actual losses and the claim is paid without a loss adjuster visiting the field has formed a habit. The infrastructure subsidy is slower and more durable.

A concrete design. Take a rainfall index contract for rain-fed maize in a semi-arid district, the kind of product that has been piloted in several provinces since the mid-2010s. The contract specifies a weather station, a sowing window, and a three-phase trigger structure:

• Phase 1 (establishment): cumulative rainfall from sowing to 30 days after sowing. If cumulative rainfall falls below 40 mm, the contract pays 30 percent of the sum insured. This covers replanting costs.
• Phase 2 (vegetative): cumulative rainfall from day 31 to day 75. If rainfall falls below 60 percent of the ten-year district average for the same window, the contract pays a linear proportion of 50 percent of the sum insured, scaled from zero at 60 percent of average to the full 50 percent at 30 percent of average.
• Phase 3 (grain fill): cumulative rainfall from day 76 to harvest. Same linear structure, 20 percent of the sum insured, triggered below 70 percent of the ten-year average.

The payout is triggered by the station reading, not by a field visit. A farmer whose rainfall is 55 percent of average in Phase 2 receives \(0.5 \times 0.5 \times S = 0.25S\), where \(S\) is the sum insured, within 14 days of the phase close. The basis risk—the gap between the station reading and the farmer's actual loss—is the central design problem. A farmer with a well, or with a plot in a micro-valley that collects runoff, may harvest a normal crop in a year the station records a deficit. The contract pays nothing, and the farmer concludes the product is a lottery.

The design response is to reduce basis risk through station density and index calibration, not through loss adjustment. A district with one station covering 400 square kilometers carries more basis risk than one with four stations covering 100 square kilometers each. The infrastructure subsidy should pay for the stations and the calibration work, because those are the costs that determine whether the product is trusted. A premium subsidy that leaves the station density unchanged produces cheap insurance that farmers buy once.

The link to credit is the part policy has most neglected. Insurance that reduces default probability should be recognized in loan pricing. In practice, lenders rarely adjust the interest rate when a borrower attaches an insurance policy, because the lender's internal model has no parameter for coverage. A regulatory nudge—requiring that lenders document how insurance coverage enters the default probability estimate, or offering a lower risk weight for insured green exposures—would make the link operational. Without such a nudge, insurance remains a standalone product that protects the borrower but does not lower the cost of capital.

15.1.4 Catalysis: accept the market failure, design for exit

Sci-tech finance catalysis—government-guided funds, public venture capital, guarantee schemes for early-stage green technology—exists because private capital underinvests in early-stage innovation with long payback periods and diffuse spillovers. The market failure is real. The policy question is not whether to intervene but how to intervene without crowding out the private capital the intervention is meant to mobilize.

The evidence on crowding out is mixed and depends on design. The three studies most often cited in this debate differ in setting, sample, and finding, and the differences are instructive.

Cumming and MacIntosh (2006) examined Canadian public venture capital programs using a sample of several thousand Canadian venture-capital investments from the 1990s. They found that public funds tended to invest alongside private funds in the same deals and at the same stages, and that the presence of public money was associated with lower subsequent private investment in the same portfolio companies—a crowding-out pattern, though the authors were careful to note that the effect varied by program design and was strongest where public funds invested on identical terms to private ones.

Leleux and Surlemont (2003) studied a pan-European sample of public and private venture-capital investments and reached a different conclusion. In their data, public investment in early-stage deals was associated with subsequent private investment in the same firms and sectors—a seeding effect rather than crowding out. The difference from the Canadian result appears to lie in the stage at which public money entered: European public funds in their sample were more likely to take the earliest positions, leaving later rounds to private investors.

Lerner (1999) evaluated the U.S. Small Business Innovation Research program, which provides grants rather than equity. Comparing SBIR awardees to matched non-awardees, he found that awardees grew faster and attracted more private investment, with no evidence that the grants displaced private capital. The grant structure—non-dilutive, modest in size, awarded on technical merit—appears to have avoided the crowding-out problem that equity programs face.

The pattern across the three studies is not the presence of public money but the terms on which it enters. Public funds that invest on the same terms as private funds, in the same rounds, at the same valuations, crowd out. Public funds that invest earlier, take subordinate positions, or provide guarantees that lower private investors' downside without capping their upside seed the market. The Canadian evidence is the strongest caution; the European and U.S. evidence shows that design can avoid the outcome the Canadians documented.

The design recommendation is therefore to make the public position structurally subordinate. A guarantee that covers the first 30 percent of losses on a green technology portfolio costs the public less than an equity stake of the same nominal size and leaves the upside with private investors, who then have a reason to select and monitor. A first-loss guarantee also solves the information problem that plagues public equity investment: the public fund does not need to pick winners if it is only absorbing a slice of losses on a portfolio that private managers pick.

Exit discipline is the second design requirement. Public catalysis funds that never exit become permanent subsidies. The fund's mandate should specify a holding period and a divestment path, and performance evaluation should judge the fund on the returns it realizes after exit, not on the size of its portfolio. The next subsection takes up evaluation directly.

15.1.5 Evaluation: measure the sequence, not the components

The evaluation of policy-guided green instruments has a structural flaw. Each instrument is evaluated on its own metrics: the rating pilot on the number of firms rated, the insurance program on the number of policies sold, the catalysis fund on the amount of capital deployed. None of these metrics tells us whether the sequence works. A rating that no lender uses, insurance that no borrower renews, and a fund that invests in firms that would have been funded anyway can all post impressive headline numbers.

The evaluation should be built around the transitions in the sequence. Three metrics capture most of what matters:

• The conversion rate from rating to priced loan, disaggregated by rating band. If firms with strong environmental ratings do not receive better loan terms, the rating is decorative.
• The default rate on insured versus uninsured green loans, matched on observable characteristics. This is the empirical test of the buffering hypothesis.
• The private-capital multiplier on catalysis: dollars of private investment mobilized per dollar of public capital at risk, measured at exit.

These metrics are harder to collect than headcounts. They require loan-level data linking the three stages, which in turn requires the data standard discussed at the outset. That is the point. Evaluation and coordination are the same problem viewed from different ends.

Table 15-1 sets out the policy actions, the responsible level of government, the failure mode each action is meant to address, and a concrete cost anchor.

The pattern in the table is that most of the high-leverage actions are cheap. Disclosure requirements, supervisory recognition of insurance, and evaluation mandates cost little and change behavior. The expensive items—registries and guarantees—are expensive because they produce public goods that no private actor will produce alone. The policy error to avoid is spending on the expensive items while neglecting the cheap ones, which produces infrastructure that no one uses.

15.1.6 Sequencing across levels of government

A practical objection: many of these actions sit at different levels of government, and coordination across levels is slow. The registry is national because environmental data is national. The insurance infrastructure is provincial because weather and forest conditions vary. The catalysis funds are often regional because industrial policy is regional. How should a country sequence the reforms?

Our recommendation is to start with the link that requires the least new institution: the disclosure requirement that makes insurance coverage visible in loan files. This can be done by supervisory guidance within a year and creates immediate pressure for the other links. The registry is the second step because it has the longest build time and the widest benefit. Catalysis reform is third because it is the most politically contested; it is easier to argue for first-loss guarantees once the rating and insurance links are in place and the portfolio data shows which firms are being screened out. The sequence is not a hierarchy of importance. It is a hierarchy of feasibility.

There is a risk in this sequencing. If the disclosure requirement arrives before the registry, lenders will disclose weights based on whatever environmental data they already have, which is often self-reported by borrowers. Self-reported data is better than nothing and worse than satellite data. The transitional period will produce ratings that flatter firms with good disclosure departments. This is acceptable if the registry arrives on schedule and the transitional ratings are labeled as such. It is not acceptable if the registry slips and the self-reported ratings become the permanent standard.

15.2 Limitations and Future Research Directions

15.2.1 What our evidence cannot establish

The argument of this book rests on three empirical claims: that environmental performance can be measured at the firm level with acceptable error; that insurance reduces the default probability of green borrowers; and that catalysis mobilizes private capital when structured as a subordinate position. Each claim is supported by evidence we have assembled, and each is weaker than the argument requires. We should say where.

On measurement, the rating work in Chapter 7 draws on technology firms in three provinces, with a sample of 412 firms observed over four years. The samples are not representative of the full firm-size distribution—small firms are underweighted because they are less likely to appear in the administrative data—and the environmental performance data is drawn from a period when reporting requirements were changing. The measurement error is likely non-classical: firms with better compliance departments are both more likely to report and more likely to score well. This biases the rating's predictive power upward in our sample relative to what a national rollout would achieve. The default-probability separation we report in Chapter 9—roughly 0.6 percentage points between the top and bottom environmental quartiles—is probably an upper bound.

On insurance, the default-probability reduction estimated in Chapter 11 comes from a setting where the insured peril is rainfall and the insured crop is a single annual. The estimate—a 0.4-point reduction in one-year default probability for insured borrowers relative to matched uninsured borrowers—does not transfer automatically to forest insurance, where the peril is fire and the insured asset is multi-year, or to the credit insurance products that attach to receivables. The mechanism—reduction of cash-flow variance at the moment of loan repayment—is general. The magnitude is not.

On catalysis, the crowding-out estimates in Chapter 13 are the least secure. The counterfactual—what private investors would have done absent the public fund—is unobservable, and every identification strategy we have tried relies on an assumption about selection that is plausible but not testable with our data. The honest statement is that we know more about how catalysis funds fail than about how they succeed.

15.2.2 The measurement agenda

The first research priority is a proper measurement model for environmental performance at the facility level. This is a statistics problem before it is a finance problem. We need to know the error structure of satellite-based emissions estimates, the lag between a compliance violation and its appearance in administrative data, and the correlation between self-reported and independently measured performance. Without these parameters, every rating model is calibrated on noise.

A concrete research design: take a sample of facilities for which both satellite estimates and inspection records exist, and estimate the confusion matrix between the two measurement systems. The off-diagonal entries tell us how often a facility that looks clean from space is dirty on inspection, and vice versa. That matrix is the input a rating model needs and currently lacks. Several remote-sensing groups have the data to do this. The finance literature has not asked for it.

The second measurement priority is the default model. The standard credit scoring model estimates default probability as a function of firm characteristics. Adding an environmental performance variable changes the model in two ways: it adds a regressor, and it changes the interpretation of the other regressors if environmental performance is correlated with them. A firm with poor environmental performance may also have poor management, and the environmental variable may be picking up management quality. Disentangling the two requires either a natural experiment in which environmental enforcement changes exogenously, or an instrument for environmental performance that is uncorrelated with management. We have neither. The research agenda should prioritize quasi-experimental settings—new enforcement regimes, changes in monitoring technology—where the environmental variable moves for reasons unrelated to firm quality.

15.2.3 The insurance-credit link agenda

The claim that insurance reduces default probability is intuitive and under-tested. The mechanism is that insurance truncates the lower tail of the borrower's cash-flow distribution at the moment loan payments are due. The size of the effect depends on the correlation between the insured peril and the borrower's income, the timing of payouts relative to payment dates, and the basis risk between the index and actual loss.

Each of these is a researchable parameter. Basis risk is the most important and the least studied in the credit context. An index insurance contract that pays out when rainfall falls below a threshold protects the lender only if the borrower's repayment capacity falls with the same rainfall measure. If the borrower has irrigation, or off-farm income, or a crop that is drought-tolerant, the correlation weakens and the insurance is worth less to the lender than to the borrower. A research program that estimates the correlation between index triggers and repayment outcomes, across crops and regions, would let lenders price insurance coverage properly. Our own estimates cover one crop in one region. That is a start, not a foundation.

The forest insurance case is more complex because the insured asset is also the collateral. A fire that destroys a forest destroys both the borrower's income and the lender's security. Insurance that pays the borrower restores the income but not the collateral unless the payout is directed to replanting. The design question—whether forest insurance payouts should be tied to replanting obligations—is a research question as much as a policy one. We do not know whether tying payouts to replanting increases or decreases take-up, and we do not know how lenders value a replanted forest as collateral relative to a mature one.

15.2.4 The catalysis agenda

The catalysis literature has a selection problem at its core. Public funds choose which firms to support, and the firms they choose are not random. The standard approach is to compare supported firms to rejected applicants, but rejected applicants are a selected sample—they are firms that applied and were turned down, which is different from firms that never applied. The ideal comparison is to firms that would have applied had the fund not existed, which is unobservable.

Two research strategies can make progress. The first is a discontinuity design around the fund's selection threshold. If the fund scores applicants and funds those above a cutoff, firms just below the cutoff are a plausible counterfactual for firms just above. This design requires that the fund's scoring be observable and that firms cannot manipulate their scores, both of which are strong assumptions but testable in some settings. The second is a randomized encouragement design: the fund identifies a pool of eligible firms and randomly offers some of them application assistance. The offer affects application but not eligibility, so the comparison identifies the effect of funding on the marginal applicant. Neither design has been used in the green catalysis context. Both should be.

A case to anchor the exit question. Consider a regional green-technology fund established in 2014 with $120 million in public capital, mandated to co-invest in early-stage clean-tech firms alongside private venture partners on a subordinate basis—the public side taking the first 25 percent of losses on each deal. By 2024, the fund had committed $98 million across 34 portfolio companies. Of those 34, six had reached IPO or trade sale, returning $41 million to the public account against $22 million invested in those six—a gross multiple of 1.86 on the exited positions. Eleven had been written off, representing $29 million in losses. Seventeen remained in portfolio, with holding periods now between four and ten years, and the fund's mandate specified a ten-year horizon with a two-year extension option.

The exit history is instructive in two ways. First, the exited positions returned more than the fund's overall cost of capital, but the write-offs consumed most of the gain: net of losses on the eleven failures, the fund had returned $41 million against $98 million committed, with $47 million still unrealized. Second, the six exits were concentrated in two sectors—battery recycling and grid software—where private co-investors had led the later rounds. The fund's subordinate position appears to have done what the design intended: it absorbed downside on deals that private investors would not have entered on senior terms, and the private partners selected and monitored the firms that eventually exited. Whether the model works at scale, and whether the seventeen remaining positions exit within the mandate, is the question the next five years will answer.

The exit question is equally open in the literature. Public funds that exit successfully—through IPO or trade sale—generate returns that can be recycled. Funds that do not exit hold assets indefinitely and become, in effect, permanent owners. We know little about what determines exit for public funds. The private venture literature emphasizes syndication, board composition, and the legal environment. Whether these factors operate the same way for public funds, which have political constraints on exit timing and valuation, is an open question. A dataset of public fund exits across countries would be a valuable research infrastructure.

15.2.5 The integration agenda

The most important gap is the one this chapter opened with: the interaction among the three instruments. Almost all existing research studies one instrument at a time. The integrated mechanism is a system, and systems have properties that their components do not.

Consider a specific interaction. A firm with a strong environmental rating receives a loan at a lower rate. The lower rate improves its cash flow, which improves its rating next period. This is a feedback loop that could either be virtuous—better performance lowers the cost of capital, which funds further improvement—or perverse, if the rating rewards firms that already have access and penalizes firms that need the cheap capital to improve. Which loop dominates is an empirical question about the slope of the rating-to-rate relationship and the cost of environmental improvement. We do not know the answer.

A second interaction: insurance coverage might reduce the lender's monitoring incentive. If the lender knows that a weather event will trigger a payout that covers the loan, the lender may monitor the borrower's environmental compliance less closely. This is a moral hazard at the institutional level, distinct from the borrower-level moral hazard that insurance usually addresses. Whether it operates in practice depends on how the insurance contract allocates monitoring responsibility. The contract design question is researchable; the empirical question of whether monitoring actually falls is unstudied.

A third interaction: catalysis funds might prefer firms that have already been rated and insured, because those firms are easier to diligence. This would concentrate public support among firms that are already well-served, which is the opposite of the intended targeting. The selection data to test this exists inside fund portfolios. It has not been analyzed.

These interactions share a structure. Each is a case where the output of one instrument becomes an input to another, and the sign of the effect depends on parameters we have not estimated. The research agenda for the next decade is to estimate them. A system that cannot estimate its own feedback loops cannot be managed.

15.2.6 Methodological note: the case for linked administrative data

Every item on this agenda requires data that links firms across the three stages. Rating data sits with lenders and regulators. Insurance data sits with insurers. Catalysis data sits with public funds. In most countries these datasets are held by different institutions under different confidentiality regimes, and linking them requires legal authority that does not exist.

The policy recommendation that follows is unglamorous but foundational: create a legal basis for linked, de-identified administrative data on green finance, accessible to accredited researchers under controlled conditions. Several countries have such arrangements for health and labor data. Finance has lagged because the data is commercially sensitive. The sensitivity is real, and the standard safeguards—de-identification, secure enclaves, publication review—are well developed. Without linked data, the integration agenda stalls at the level of theory.

15.3 Concluding Remarks: Toward a Green Financial System

15.3.1 The argument restated

This book began with a question: how can the financial system identify and support the green transition of industries? The answer we have developed is that identification and support are two functions that must be performed by different instruments in sequence. Rating identifies. Insurance buffers. Catalysis accelerates. No single instrument can do all three, and an attempt to make one instrument do all three produces the failures we documented in the early chapters—credit rationing that excludes the firms most in need of transition finance, insurance markets that fail to form because the correlated risk is unmanageable, and public investment that crowds out the private capital it was meant to mobilize.

The sequence matters because each stage solves a problem the next stage would otherwise face. Rating solves the information problem that makes lenders ration credit to green firms whose environmental performance they cannot verify. Insurance solves the risk problem that makes lenders reluctant to lend to firms whose cash flow depends on weather. Catalysis solves the scale problem that makes private investors reluctant to fund early-stage innovation whose returns are distant and diffuse. Remove any stage and the others degrade. Rating without insurance leaves the lender bearing a risk it cannot price. Insurance without rating insures the wrong firms. Catalysis without either invests in firms whose performance cannot be measured.

The evidence for each link is uneven, and we have been explicit about where it is thin. The rating link is the best established, resting on the 412-firm sample in Chapter 7 and the default-separation estimate in Chapter 9. The insurance link is plausible and supported by the estimates in Chapter 11, but rests on a narrow empirical base—one crop, one region. The catalysis link is the most uncertain, because the counterfactual is hardest to construct. A reader who wants a single sentence should take this: the integrated mechanism is a design hypothesis with partial empirical support, not a proven blueprint. The research agenda in the previous section is what would turn it into a blueprint.

15.3.2 What a green financial system would look like

If the recommendations in this chapter were implemented, what would an observer see? Not a new institution. The observer would see a loan file with three additional documents: a facility-level environmental score with a date and a confidence interval, an insurance schedule showing which perils are covered and how payouts are triggered, and a note showing how the first two documents entered the credit decision. The observer would see a public registry that any lender can query and a supervisory report that shows, by rating band, the conversion rate from application to priced loan. The observer would see a catalysis fund whose mandate specifies an exit date and whose annual report shows realized returns, not deployed capital.

None of this is dramatic. The drama in green finance has been in the announcements—the pledges, the taxonomies, the funds with large headline numbers. The work that makes those announcements mean something is clerical. It is the construction of a score that can be reproduced, an insurance contract that pays without a loss adjuster, a guarantee that absorbs the first loss so that a private investor will take the second. The clerical work is where the transition either happens or does not.

There is a temptation to end with a vision of a fully green financial system, in which every loan is priced for its environmental impact and every firm's transition path is financed at its social cost. We should resist the temptation, because the vision is not operational and the operational version is more useful. The operational version is a system in which environmental performance is measured with known error, insured against known perils, and financed with known public subsidies for a known period. That system would not be fully green. It would be honest about what it knows and what it does not.

15.3.3 The political economy of implementation

A final observation about implementation. The recommendations in this chapter are technically modest and politically difficult. The technical modesty is real: a registry, a disclosure requirement, a supervisory nudge, a guarantee structure. The political difficulty is also real, and it has a specific shape. The costs of these measures are concentrated and visible—a reporting burden on lenders, a contingent liability on the budget, a constraint on a public fund's investment discretion. The benefits are diffuse and delayed—better allocation of capital across an economy, lower transition risk in loan portfolios, a measurable reduction in the cost of green capital over a decade.

This is the classic political economy of public goods, and it does not solve itself. The measures that are cheap in budgetary terms are expensive in bureaucratic terms, because they require regulators to coordinate across agencies that have no habit of coordinating. The measures that are expensive in budgetary terms are cheap in bureaucratic terms, because they can be implemented by a single agency writing a check. The path of least resistance runs toward the check, which is why so many green finance programs consist of subsidies without infrastructure.

The counterweight is evidence. The reason to fund the registry before the subsidy is that the registry makes the subsidy's performance measurable. A government that spends on catalysis without a rating and insurance infrastructure has no way to know whether the spending worked. It will be evaluated on deployed capital, which is the metric that is easiest to hit and least informative. A government that builds the infrastructure first can be evaluated on conversion rates and default rates and realized returns, which are harder to hit and more informative. The research community's job is to make those metrics available and credible. The policy community's job is to demand them.

The loan officer with two spreadsheets is still at her desk. The policy question is whether the next version of her credit system has one sheet or two, and whether the environmental column is connected to the pricing column. A regulator that wants to answer that question can start by requiring, in the next supervisory reporting cycle, that every lender using the public environmental registry disclose the weight it assigns to the registry score in its default model, the resulting change in estimated default probability, and the confidence interval it attaches to the score. That disclosure is due within twelve months of the registry's first national release. It is a filing requirement, not a pledge, and it is the smallest step that makes the rest measurable.

15.3.4 Notes on sources and further reading

The crowding-out evidence discussed in 15.1.4 draws on Cumming and MacIntosh (2006), Leleux and Surlemont (2003), and Lerner (1999). The common-agency framework that underlies the coordination argument is developed in Dixit, Grossman, and Helpman (1997) and Martimort (1996). The incentive-design issues in public catalysis funds follow Holmström and Milgrom (1991) on multitask principal-agent problems. The credit-rationing logic that motivates the rating stage is from Jaffee and Modigliani (1969) and Bernanke (1983, 1988). The insurance-credit link builds on the credit insuranceunderwriting logic described in the receivables literature, particularly the treatment of discretionary credit limits. The EU ETS, launched in January 2005 and covering power, steel, cement, and glass across member states plus Norway and Iceland, provides the regulatory backdrop for the transition-risk discussion in 15.2.3. The linked-administrative-data proposal in 15.2.6 follows the model of secure research enclaves used for health and labor data in several jurisdictions; the fintech literature on alternative data (Frost et al. 2019; Hau et al. 2019) illustrates both the potential and the privacy constraints of such linkages.