Chapter 12: Incentive Design and Performance Evaluation of Policy-Guided Green Finance

A green credit guarantee fund in a mid-sized Chinese prefecture lent RMB 1.2 billion against government guarantees in its first three years. By the standard reporting metrics, the program worked: 214 firms received subsidized credit, the non-performing loan ratio stayed below 1.5 percent, and the local green loan book grew 41 percent year on year. Then an audit traced the borrowing firms' environmental registrations. Forty-one percent had been reclassified as "green" within six months of the fund's launch, most of them after submitting a one-page self-assessment. Their emissions per unit of output had not moved. The label had.

That audit is the problem this chapter is about, and it is not primarily a problem of bad faith. It is measurement and incentive compatibility failing at the same time. The fund's managers were evaluated on loan volume and default rate. The firms were evaluated on whether they could produce a green label. Nobody was evaluated on whether emissions fell. Each agent optimized against the metric in front of it, and the metric was wrong.

Chapter 11 closed on whether public equity instruments for green technology produce genuine catalysis or a well-intentioned transfer to private investors who would have been funded anyway. The question generalizes. Every policy-guided green finance instrument — subsidized credit, guarantee funds, green bonds with preferential approval, government guiding funds, interest subsidies — faces the same structural test: does the incentive design cause the agents who control capital to do something they would not otherwise have done, and does the performance evaluation detect the difference? This chapter works through both halves of that test.

12.1 Incentive Compatibility and Mechanism Design

12.1.1 The Principal-Agent Problem in Green Finance Is a Multi-Task Problem

The standard principal-agent model with a single task has a clean solution: pay the agent a share of output, and the agent exerts effort up to the point where marginal cost equals marginal benefit. Green finance does not have a single task. A bank branch manager who receives a green credit quota is asked to do at least four things at once: expand the green loan book, screen borrowers for genuine environmental performance, monitor whether the funded project delivers the claimed reductions, and keep default rates low. These tasks compete for the manager's time and attention, and they are measured with very different precision.

Holmstrom and Milgrom's multi-task framework (1991) predicts what happens next. When one task is measurable and another is not, the agent shifts effort toward the measurable task. If the measurable task is loan volume and the unmeasurable task is environmental screening, the agent will lend more and screen less. This is not a moral failing. It is the equilibrium of the contract.

The implication for green finance design is direct: any performance metric that rewards volume without adjusting for environmental outcome quality will produce volume without environmental outcome quality. The audit finding above — 41 percent reclassification with no emissions change — is the multi-task prediction confirmed in the field.

Formally, let the agent allocate effort \(e = (e_v, e_s)\) between volume expansion \(e_v\) and screening intensity \(e_s\). Output on the two dimensions is

where \(\varepsilon_v\) and \(\varepsilon_s\) are noise terms. If the principal can only observe \(y_v\) reliably (because \(y_s\) is measured with a long lag or not at all), the optimal linear contract pays \(\beta_v y_v\) with \(\beta_v > 0\) and \(\beta_s = 0\). The agent's best response sets \(e_s\) to its minimum. Screening collapses.

The fix is not to pay more for volume. It is to make \(y_s\) observable and to make the agent's payoff depend on it. That is a measurement problem before it is a compensation problem, which is why this chapter's second section on performance metrics is not separable from its first section on incentives. You cannot write a contract on a variable you cannot measure.

12.1.2 Information Rent and the Green Label

A second incentive problem arises between the regulator and the firm. The firm knows its own environmental performance; the regulator does not. When the regulator offers a subsidy or a preferential loan rate to firms that qualify as green, the firm has an incentive to overstate its performance to capture the rent.

This is a screening problem with adverse selection. Let \(\theta \in [0,1]\) denote the firm's true environmental performance, privately known. The regulator offers a transfer \(t\) to firms that report \(\hat{\theta} \geq \theta^*\). The firm's payoff from reporting is \(t - c(\hat{\theta})\), where \(c(\hat{\theta})\) is the cost of producing evidence supporting the report. If \(c\) is low — a one-page self-assessment, as in the audit case — every firm reports \(\hat{\theta} = 1\) and the screening threshold \(\theta^*\) carries no information.

The standard remedy is to make \(c(\hat{\theta})\) increasing in \(\hat{\theta}\): require third-party verification, audited data, or measurement at the facility level. The cost of producing credible evidence must rise with the claimed performance, or the claim is cheap talk. This is why the EU Taxonomy Regulation's technical screening criteria matter more than its political signaling. A firm that claims alignment with the taxonomy must demonstrate it against specific numeric thresholds, and the demonstration is auditable. The cost of a false claim is not zero.

There is a trade-off here that is easy to miss. Raising verification costs reduces false positives but also deters genuine small firms that cannot afford the compliance burden. A green finance program that requires full lifecycle carbon accounting from a 40-person manufacturer will screen out most of the firms it was designed to help. The design problem is to set verification intensity proportional to the size of the rent being claimed, so that the cost of verification is small relative to the subsidy for small claims and large for large claims. Tiered verification — self-assessment plus random audit below a threshold, full third-party verification above it — is the practical compromise. It is also, in most programs, absent.

The trade-off can be made to yield a comparative static that the prose alone does not. Suppose the regulator must choose a single verification threshold \(\bar{t}\) (a claim size above which full verification is required), firms are distributed by claim size \(t\) with density \(f(t)\), verification costs \(k\) per verified claim, and the social loss from an unverified false claim is proportional to \(t\) with factor \(\lambda\). The regulator's problem is to choose \(\bar{t}\) to minimize

The first-order condition gives

The optimal threshold rises with the unit cost of verification and falls with the social cost of a false claim. Two consequences follow that are not obvious from the prose. First, as verification technology cheapens (remote sensing, satellite methane detection, continuous emissions monitoring), \(k\) falls and \(\bar{t}^*\) falls with it: more claims should be verified, and the tier boundary should move down over time. Programs that set the tier boundary once and never revise it are leaving false claims unverified as the technology improves. Second, in sectors where a false claim is especially damaging — a claim of zero methane leakage from a gas facility, say — \(\lambda\) is high and \(\bar{t}^*\) is low, so verification should be near-universal even for small claims. A single threshold across sectors is therefore wrong; the threshold should be sector-specific, and it should be revised as \(k\) changes.

12.1.3 Common Agency and the Overlapping Mandates Problem

Green finance agents typically report to multiple principals. A state-owned bank branch reports to its head office, to the local branch of the central bank, to the provincial development and reform commission, and to the banking regulator. Each principal sets targets. The targets are not always consistent.

Bernheim and Whinston (1986) and Martimort (1996) analyze common agency: multiple principals contracting with a single agent whose effort is not fully observable. The equilibrium depends on whether the principals cooperate or compete. When they compete — each offering its own incentive contract — the agent can extract rents by playing one principal against another. When they cooperate, the agent faces a single aggregated contract and the incentive is coherent.

The practical failure mode in green finance is competitive common agency without coordination. A bank that hits its green loan target set by the central bank but misses its local industrial policy target faces conflicting rewards. The branch manager learns to satisfy the metric that is monitored most frequently and punishes most visibly. If the green loan target is reported monthly and the environmental outcome target is reported annually, monthly wins.

Coordination mechanisms exist. China's green finance evaluation system, introduced by the People's Bank of China in 2018 and revised since, aggregates green loan volume, share of green loans in the total book, and environmental risk management into a single score that feeds into the central bank's assessment of financial institutions. The aggregation is imperfect — it weights volume heavily — but it reduces the number of conflicting contracts. The lesson is not that centralization is better. It is that the number of principals setting targets must be matched by a mechanism that aggregates their demands into one score the agent can optimize against.

12.1.4 Why "Do No Harm" Bonuses Fail

A common design adds a penalty for environmental incidents to an otherwise volume-based contract. The bank gets a bonus for green loan growth and a deduction if a funded project causes a pollution event. On paper this aligns incentives. In practice it often fails for a structural reason: the penalty is triggered by a low-probability event, and the agent is risk-neutral over the relevant range because the penalty is capped or absorbed by the institution rather than the individual.

Consider a branch manager choosing between two loan portfolios. Portfolio A has high expected green volume and a 5 percent chance of an environmental incident that triggers a penalty of \(P\). Portfolio B has lower volume and no incident risk. The manager chooses A if

If \(P\) is set at the institutional level and does not flow to the manager's compensation, the manager's effective \(P\) is zero and A always wins. The penalty must be internalized at the level of the decision-maker to change behavior. Most programs set penalties at the institutional level, which means they change the institution's reported risk but not the loan officer's choice.

This is not an argument against penalties. It is an argument that the unit of incentive design must match the unit of decision. If loan officers decide, loan officers must face the consequence. If credit committees decide, the committee's composition and voting rules must build in the environmental constraint. A penalty that lands on the institution is a cost of doing business; a penalty that lands on the decision-maker is a constraint.

12.1.5 The Time-Inconsistency Problem

Green transition outcomes arrive slowly. A loan that funds a factory retrofit produces measurable emissions reductions two to four years after disbursement. Political and managerial evaluation cycles are annual. The mismatch creates a time-inconsistency problem: the principal promises to reward long-run outcomes, but the agent knows the principal will be evaluated on short-run metrics before the long-run outcome arrives, and will therefore renege on the long-run reward.

The result is that agents discount long-run environmental outcomes heavily, even when the contract formally rewards them. A bank that is told its green loan performance will be judged on three-year emissions reductions will still optimize for this year's reported green loan volume if this year's evaluation determines promotions and budget allocations.

The design response is to make the long-run reward credible through commitment devices: escrow accounts that release on verified outcomes, multi-year evaluation cycles that are legally fixed, or third-party verification that the principal cannot override. Without a commitment device, the long-run clause is cheap talk and the agent knows it.

The five failure modes above share a shape but not a symptom, and the symptoms are what a program manager can actually observe. Table 12-0 sets them out.

Table 12-0: Failure modes in green finance incentive design and their observable symptoms

Each row corresponds to a branch in the diagnostic logic of this section. The multi-task row is the opening audit. The cheap-talk row is the same audit seen from the firm side. The common-agency row is the branch manager with four masters. The uninternalized-penalty row is the pollution-event bonus that changes nothing. The time-inconsistency row is the retrofit loan judged on this year's volume.

12.2 Performance Metrics and Evaluation Methods

12.2.1 The Three Dimensions and Why They Conflict

Policy-guided green finance is evaluated on three dimensions: environmental performance, social performance, and financial performance. The dimensions are not independent. A program that maximizes emissions reductions per dollar may concentrate lending in a few large firms, harming the small and medium enterprise access objective. A program that maximizes financial return may avoid the early-stage, higher-risk projects where the transition need is greatest. A program that maximizes social inclusion may fund projects with weak environmental additionality.

The evaluation framework must therefore be explicit about the trade-off, not pretend it away. The standard approach in development finance is a weighted scorecard, but weights are political choices disguised as technical ones. A more honest approach reports the three dimensions separately and makes the trade-off visible, so that the reader can see what is being sacrificed for what.

Table 12-1 lays out the dimensions and their typical indicators.

Table 12-1: Performance dimensions and indicators for policy-guided green finance

The measurement lag column is the most important. Environmental outcomes are measured years after the lending decision, which is precisely the time-inconsistency problem from Section 12.1.5. Any evaluation that reports environmental performance on an annual cycle without acknowledging the lag is measuring claims, not outcomes.

12.2.2 Environmental Additionality: The Core Test

The central question for any policy-guided instrument is additionality: did the intervention cause environmental outcomes that would not have occurred otherwise? This is harder to measure than it sounds, because the counterfactual is unobservable.

Three approaches are in use.

The first is the before-after comparison. Measure emissions at the funded facility before and after the loan. This is the most common method and the weakest. It attributes all change to the intervention, ignoring secular trends, unrelated investments, and the possibility that the facility would have reduced emissions anyway. Its virtue is cost: a single facility's meter readings before and after, with no comparison group, can be assembled for a few thousand RMB.

The second is the comparison with a control group. Match funded firms with similar unfunded firms and compare outcomes. This is better, but matching on observables does not eliminate selection on unobservables. Firms that apply for green finance differ from firms that do not in ways that are correlated with their environmental trajectory. The data cost rises accordingly — you need firm-level records for both groups, which means access to the lender's loan book and to a comparison pool of non-applicants.

The third is the regression discontinuity design, exploiting a threshold in the program's eligibility rule. If firms below a size cutoff are eligible and firms above are not, and firms cannot precisely manipulate their size, the discontinuity identifies the causal effect. This is the strongest design and the least used, because most programs do not have a clean threshold or do not release the data needed to exploit it. Where a threshold exists, the data cost is modest — you need only firms in a narrow band around the cutoff — but the institutional cost is high, because the program must be willing to let the discontinuity be studied.

The matched-sample evaluation referenced at the end of Chapter 11 belongs to the second category. It is a real improvement over before-after comparison. It is not a substitute for a design-based evaluation, and the difference matters when the policy question is whether to continue or scale the program.

A concrete number illustrates the stakes. Suppose a green credit program lends RMB 10 billion at a 1.5 percentage point interest subsidy, costing RMB 150 million per year. If the program's additionality is 20 percent — one in five funded projects would not have proceeded without the subsidy — then the subsidy per genuinely additional project is five times the headline figure. If additionality is 60 percent, the cost per additional project is less than twice the headline. The same budget produces very different environmental returns depending on a parameter that most evaluations do not estimate.

The design choice also carries a measurable cost. Before-after evaluation of a mid-sized guarantee fund — say 200 funded firms — can be assembled from facility meter readings and loan records for perhaps RMB 200,000 and three months. A matched-sample evaluation of the same fund, requiring a comparison pool of 400–600 non-applicant firms and two years of outcome data, costs roughly RMB 1.2–1.8 million and takes nine to twelve months. An RDD evaluation around a size cutoff, restricted to firms within a narrow band, costs less in data assembly — perhaps RMB 600,000–900,000 — but requires the program to release the eligibility rule and the loan book, which is the binding constraint in most Chinese programs. The cost per evaluation rises with design quality by a factor of three to nine; the credibility of the additionality estimate rises by more, because before-after estimates in green credit programs routinely overstate additionality by 30–50 percentage points when checked against a matched comparison. The evaluation budget should be set against that gap, not against the cheapest available method.

12.2.3 The Green Premium as an Evaluation Input

Chapter 4 introduced the green premium as a diagnostic metric: the cost differential between a green project and its conventional substitute. In evaluation, the green premium serves a second function. It measures the size of the market failure the policy is trying to correct.

If the green premium for a technology is negative — the green option is already cheaper on a lifecycle basis — then a subsidy for that technology is not correcting a market failure. It is paying for something the market would do anyway. The additionality is near zero by construction.

If the green premium is positive and large, the subsidy has a rationale, but the required subsidy per unit of adoption is also large. The evaluation should report the premium alongside the subsidy, so that the reader can see whether the instrument is calibrated to the gap it claims to close.

Let \(\pi_g\) and \(\pi_c\) denote the levelized cost of the green and conventional options. The green premium is \(\Delta = \pi_g - \pi_c\). A subsidy \(s\) per unit closes the gap when \(s \geq \Delta\). The efficiency of the instrument is the ratio of the environmental outcome delivered to the subsidy paid, and the maximum achievable efficiency is bounded by \(1/\Delta\). A program that pays \(s \gg \Delta\) is over-subsidizing and will attract inframarginal participants. A program that pays \(s \ll \Delta\) will not move adoption.

Many green finance programs set subsidy rates by budget availability rather than by the measured premium. The result is either under-subsidization of the technologies with the largest gap or over-subsidization of the technologies with the smallest. Both are allocation errors that a premium-based evaluation would catch.

12.2.4 Financial Performance and the Subsidy Dependence Ratio

Financial performance evaluation for policy-guided instruments must answer a question that conventional bank evaluation does not ask: how much of the reported return depends on the subsidy?

The subsidy dependence ratio (SDR) is the standard tool, adapted from microfinance evaluation. It is defined as

where \(S\) is the total value of subsidies received (interest subsidies, guarantees, tax exemptions, below-market funding) and \(R\) is the operating revenue of the institution or program. An SDR of 0.3 means that 30 percent of revenue depends on subsidy. An SDR above 1.0 means the program cannot cover its costs from operations at all.

The SDR is useful because it makes explicit what headline return figures hide. A green fund reporting a 6 percent return with an SDR of 0.4 is not earning 6 percent; it is earning roughly 3.6 percent on an unsubsidized basis and receiving the rest as transfer. Whether that is good or bad depends on the environmental outcome per unit of subsidy, but the evaluation cannot even ask the question without the SDR.

The ratio should be reported alongside the default rate, not instead of it. A program with a 1 percent default rate and an SDR of 0.8 is a transfer program with low credit losses. A program with a 4 percent default rate and an SDR of 0.1 is closer to a commercial operation with a modest environmental mandate. These are different instruments serving different purposes, and conflating them in a single "performance" number obscures the distinction.

12.2.5 Outcome-Based versus Activity-Based Evaluation

The deepest design choice in performance evaluation is whether to pay for activities or for outcomes.

Activity-based evaluation counts what was done: loans disbursed, firms served, bonds issued, training delivered. It is easy to measure, hard to game only at the margin, and weakly correlated with environmental results. Outcome-based evaluation counts what changed: tons of CO2 reduced, cubic meters of water saved, hectares of forest conserved. It is hard to measure, easy to game if verification is weak, and directly correlated with the objective.

The trade-off is between measurement cost and incentive strength. Activity metrics are cheap and weak. Outcome metrics are expensive and strong, provided the verification is credible.

A hybrid is often the practical answer: activity metrics for monitoring and disbursement, outcome metrics for evaluation and scaled rewards. The disbursement can proceed on activity verification; the performance bonus or the continuation decision depends on outcome verification. This separates the cash flow cycle from the outcome cycle and reduces the time-inconsistency problem, because the agent is paid for activity promptly but rewarded for outcome later.

The design detail that determines whether the hybrid works is the size of the outcome-linked component. If it is 5 percent of total compensation, it is a rounding error and the agent ignores it. If it is 40 percent, it dominates and the agent optimizes for the outcome metric, including through gaming if verification is weak. The 15–30 percent range is my own working judgment, not a figure established by a single study; it reflects the pattern across the pay-for-performance literature in public services, where small outcome components are routinely ignored and very large ones invite gaming, but no review I know of pins the range down with confidence. It should be treated as a starting point for program design, to be adjusted to the verification strength of the particular instrument, not as a rule.

12.2.6 Evaluating the Evaluators

A final methodological point. Performance evaluation of green finance is itself an activity with incentives. Evaluators are hired by the programs they evaluate, or by the ministries that sponsor the programs. The incentive to report favorable results is structural.

The remedies are familiar from the broader evaluation literature: independent evaluation units with protected budgets, pre-registered evaluation designs, public data release, and replication. None of these is common in green finance evaluation. Most published program evaluations are internal, use before-after comparison, and are released selectively.

The consequence is that the evidence base for what works in green finance is thinner than the volume of program activity suggests. Thousands of green credit programs operate; a few dozen have been evaluated with a credible counterfactual; almost none have been evaluated with a design that would survive peer review in economics. This is not a reason to abandon evaluation. It is a reason to weight the existing evidence by its design quality rather than by the confidence of its conclusions.

12.3 Case Studies of Policy-Guided Instruments

12.3.1 The Green Credit Guarantee Fund: A Design That Failed Its Own Test

Return to the fund from the opening. The design had four components: a government guarantee covering 70 percent of loan principal, an interest subsidy of 1.5 percentage points, a target of RMB 1 billion in new green loans over three years, and an annual evaluation based on loan volume and default rate.

The incentive structure is visible in the design. The guarantee transfers credit risk from the bank to the government, which lowers the bank's screening incentive: if 70 percent of the loss is covered, the bank's expected loss on a bad loan is 30 percent of the principal, and the marginal return to screening falls. The interest subsidy lowers the borrower's cost, which increases demand. The evaluation rewards volume, which pushes the bank to lend. Nothing in the design rewards environmental screening or penalizes label inflation.

The result was predictable in structure and surprising only in magnitude. Loan volume grew 41 percent. Defaults stayed low, because the guarantee absorbed them. The reclassification audit found that 41 percent of the "green" borrowers had no measurable environmental improvement. The fund had done what its contract asked: it had expanded lending to firms that could produce a green label. It had not produced environmental additionality, because no part of the design measured or rewarded it.

The audit itself is worth describing in detail, because the method determines how much weight the 41 percent figure can bear. The provincial audit bureau drew a stratified sample of 120 of the 214 funded firms, stratified by loan size and by the date of green classification. For each sampled firm, auditors pulled the environmental registration file from the local ecology and environment bureau, the firm's energy consumption records for the three years before and after disbursement, and the loan file from the lending bank. A firm counted as "no measurable environmental improvement" if its energy consumption per unit of output and its reported emissions both moved by less than 5 percent over the window, and if no new pollution-control equipment had been installed. Two auditors coded each firm independently; disagreements were resolved by a third. The 41 percent figure is the share of the sample meeting that definition, and the 95 percent confidence interval on the sample proportion runs from 32 to 50 percent. The sample was not random across the full population — small loans were oversampled — so the population figure is likely somewhat lower, perhaps 35–45 percent.

Table 12-2 reports borrower characteristics before and after reclassification for the sampled firms, split by whether the firm's environmental performance actually improved.

Table 12-2: Borrower characteristics before and after green reclassification (sample of 120 firms, 2017–2020)

The pattern is not subtle. Firms that improved were larger, older, more likely to have borrowed from the bank before, and more likely to sit in sectors the local taxonomy already designated as green. Firms that did not improve were smaller, newer, and overwhelmingly self-assessed. The reclassification timing is the sharpest signal: 84 percent of the no-improvement firms were reclassified within six months of the fund's launch, against 31 percent of the improvers. A firm that genuinely changes its environmental performance does not do so in six months on a one-page form. A firm that relabels does.

The fix is not to abolish guarantee funds. It is to restructure the contract. Three changes would have changed the equilibrium. First, tie the guarantee coverage to verified environmental performance: 50 percent coverage for unverified borrowers, 70 percent for borrowers with third-party verified taxonomy alignment, 85 percent for borrowers with measured and verified emissions reductions. Second, make the interest subsidy a function of the measured green premium rather than a flat rate, so that the subsidy is calibrated to the gap. Third, replace the volume target with an outcome target: tons of CO2 reduced per RMB of guarantee exposure, evaluated with a matched control group.

None of these changes is technically difficult. Each requires data infrastructure that the fund did not have. This is the practical reason why bad designs persist: the good design requires measurement capacity that the implementing agency lacks, and building that capacity takes longer than the political cycle allows.

12.3.2 China's Green Finance Evaluation System: Aggregation as a Partial Fix

The People's Bank of China's green finance evaluation system, introduced in 2018, scores financial institutions on a composite of green loan volume, green loan share, environmental risk management, and green bond holdings. The score feeds into the central bank's assessment of the institution, which affects reserve requirements, access to central bank lending facilities, and supervisory attention.

The design addresses the common agency problem from Section 12.1.3 by aggregating multiple targets into one score. It addresses the activity-outcome problem only partially: the score is dominated by volume and share, with environmental risk management as a qualitative component. It does not measure emissions outcomes.

The evaluation data show the expected pattern. Green loan balances grew rapidly after 2018 — from roughly RMB 8 trillion to nearly RMB 12 trillion by the end of 2020, about 5 percent of total loans. The growth is real. Whether it represents additionality is not established by the evaluation system, because the system does not measure the counterfactual.

The more interesting question is what the system does to the composition of lending. A bank that needs to raise its green loan share has two options: lend more to green firms, or reclassify existing loans as green. The evaluation system's reliance on self-reported classification, with spot checks rather than universal verification, leaves the second option open. The 5 percent share figure should therefore be read as an upper bound on genuine green lending, not a point estimate.

This is not a criticism of the system's architects, who faced a real constraint: universal third-party verification of every green loan in the Chinese banking system is not feasible at current capacity. It is a statement about what the evaluation can and cannot support. The system is good at measuring activity and weak at measuring additionality. Policy conclusions that require additionality estimates cannot be drawn from it.

12.3.3 Government Guiding Funds: The Return Cap Problem

Government guiding funds for green technology typically combine public capital with private capital in a fund-of-funds structure, with a return cap on the public share and a profit-sharing arrangement that gives private investors a larger share of upside. The design is intended to attract private capital by improving its risk-return profile.

The incentive problem is subtle. A return cap on the public share means that returns above the cap accrue to private investors. This is a subsidy to private investors, and its size depends on the distribution of fund returns. If the fund's return is below the cap, the subsidy is zero. If the return is far above the cap, the subsidy is large. The private investor's expected return is therefore higher than the fund's expected return, and the gap is the expected subsidy.

The design attracts private capital, but it also selects for private investors who are most sensitive to the subsidy — which is to say, those with the least intrinsic interest in the environmental mandate. An investor who would have invested in green technology anyway does not need the return cap; an investor who is marginal does. The return cap therefore shifts the composition of the investor base toward the marginal, which is the intended effect, but it also means that the program's additionality is concentrated in investors whose commitment to the environmental objective is weakest.

A second problem is the evaluation of the fund's environmental performance. Guiding funds typically report the sectoral composition of their portfolios — what share went to renewable energy, energy efficiency, clean transport — rather than the emissions outcomes of the portfolio firms. Sectoral composition is an activity metric. It says what the fund bought, not what the firms did. A fund that invests in solar manufacturers has a green sectoral composition; whether it reduced emissions depends on whether the manufacturers' products displaced fossil generation, which is a different question.

The evaluation design that would answer the question requires following portfolio firms over time and comparing their emissions trajectories with matched non-portfolio firms. This is expensive and slow, and few guiding funds do it. The result is that the largest public equity instruments for green technology are evaluated on the least informative metric available.

12.3.4 Weather Index Insurance Subsidies: A Case Where Outcome Evaluation Is Feasible

Chapter 8 discussed weather index insurance for eco-agriculture. The subsidy design for index insurance offers a useful contrast because the outcome is measurable at low cost.

An index insurance policy pays out when a weather variable — rainfall, temperature, a vegetation index — crosses a threshold. The payout is triggered by the index, not by the farmer's realized loss. This reduces moral hazard and verification cost, at the price of basis risk: the index may not match the farmer's actual loss.

The subsidy for index insurance is typically a share of the premium, often 30–50 percent. The evaluation question is whether the subsidy increases adoption and whether adoption improves farmer welfare and land management.

Both questions are answerable with existing data. Adoption is observable from policy records. Welfare is measurable from household surveys. Land management — the environmental outcome — is measurable from satellite vegetation indices and from survey data on input use, crop choice, and conservation practices. The evaluation can use a randomized or quasi-experimental design because the subsidy rollout is often staggered across counties.

The evidence from the Chinese pilot programs is mixed but informative. Subsidized index insurance increased adoption substantially in the pilot counties. The welfare effect on farmers was positive but smaller than the adoption effect, because basis risk meant that some farmers paid premiums and received no payout in years when they suffered losses. The land management effect was positive where the index was well-correlated with local conditions and weak where it was not.

The lesson for green finance evaluation is that feasibility depends on the outcome. Emissions from a factory are hard to measure and easy to misreport. Vegetation cover on a farm is easy to measure from space and hard to misreport. The design of the evaluation should follow the measurability of the outcome, not a uniform template applied across instruments.

12.3.5 Forest Insurance: Verification Cost as the Binding Constraint

Chapter 9 discussed forest insurance. The evaluation problem here is different again. The environmental outcome — forest cover, carbon stock, biodiversity — is measurable by remote sensing at moderate cost. The financial outcome — claims paid, loss ratios — is measurable from insurance records. The additionality question — whether the insurance caused forest conservation that would not have occurred otherwise — is the hard one.

Forest insurance reduces the owner's risk of loss from fire, pest, and storm. By reducing risk, it may encourage investment in forest management, which can increase carbon stock. It may also create moral hazard: an insured owner has less incentive to invest in fire prevention. The net effect on forest cover is ambiguous and must be estimated.

The Chinese forest insurance programs, which cover roughly 180 million hectares, have been evaluated primarily on uptake and claims, not on forest outcomes. The evaluations that do exist suggest that insurance increases replanting after loss events but has weak effects on preventive management. The moral hazard effect is present but small, because deductibles and co-payments are non-trivial.

The evaluation design that would settle the question requires linking insurance records to forest inventory data at the plot level, which is feasible where both datasets exist and rare in practice. The binding constraint is not method. It is data linkage across agencies that do not share records.

12.4 Designing an Evaluation That Changes Behavior

12.4.1 The Minimum Viable Evaluation

The cases above suggest a pattern. Evaluations fail not because the methods are unknown but because the data infrastructure and the incentive to use the methods are missing. A minimum viable evaluation for a policy-guided green finance instrument has four components, and they work only as a set.

A baseline: measure the environmental performance of funded and comparison firms before disbursement. Without a baseline, no outcome claim is credible. This is the single most common failure and the easiest to fix — require baseline measurement as a condition of disbursement.

A comparison group: identify similar firms or projects that did not receive the instrument. Match on observables, and where possible exploit a threshold or a staggered rollout for a design-based estimate.

A verification protocol with cost proportional to claim size: self-assessment plus random audit for small claims; third-party verification for large claims. Report the audit findings publicly, including the false positive rate.

An outcome-linked component in the implementing agency's compensation. The share should be large enough to matter — 15 percent or more — and the outcome should be measured with a lag that the agency's budget cycle can accommodate, through an escrow or a multi-year agreement.

None of these components is novel. Each is standard practice in some program somewhere. The problem is that they are rarely combined, and the combination is what makes the evaluation credible.

12.4.2 What to Do When the Outcome Cannot Be Measured

Some environmental outcomes are genuinely hard to measure at reasonable cost. Biodiversity is the standard example. Emissions from diffuse agricultural sources are another. For these, the evaluation has three options.

Use a proxy with a documented relationship to the outcome, and report the uncertainty in the proxy. A vegetation index is a proxy for habitat quality; the relationship is imperfect and should be stated.

Use a process metric that is strongly linked to the outcome and hard to game. The share of a portfolio in activities that the taxonomy classifies as enabling — grid infrastructure for renewables, for example — is a process metric with a plausible causal link to emissions reductions.

Concede the limit. Report that the outcome is not measured, state what is measured instead, and do not claim outcome additionality that the data cannot support. This is the least satisfying option and the most honest. Evaluations that concede limits are more useful than evaluations that overclaim, because they tell the reader what to trust.

12.4.3 The Political Economy of Evaluation

Evaluation findings are used, ignored, or suppressed depending on the political configuration. A program that is evaluated favorably will be cited; a program that is evaluated unfavorably will be defended on the grounds that the evaluation missed something.

The design implication is that evaluations should be structured to survive political contestation. This means pre-registering the evaluation design before the program launches, so that the metrics cannot be chosen after the results are known. It means releasing the data, so that critics can reanalyze. It means separating the evaluation unit from the program unit, so that the evaluator's budget does not depend on the program's continuation.

None of these is fully achievable in any political system. They are directions of improvement, not states to be reached. The practical question for a program designer is which of them can be built into the institutional structure at launch, when the political capital to insist on independence is highest.

12.4.4 A Worked Comparison of Evaluation Designs

The argument that design quality should weight the evidence is easier to accept with numbers attached. Table 12-3 compares the three designs from Section 12.2.2 on cost, time, and expected bias, using the guarantee fund from Section 12.3.1 as the reference case: 214 funded firms, a three-year window, and a provincial audit bureau with access to the loan book.

Table 12-3: Cost and bias of three evaluation designs, applied to the guarantee fund case

The numbers are estimates from comparable provincial audits, not from a single published study, and they will shift with the size of the program and the availability of administrative data. The ordering is what matters. Before-after evaluation is roughly a tenth the cost of a matched sample and produces an additionality estimate that is wrong by an order of magnitude more. A program that spends RMB 200,000 on a before-after evaluation and then uses the result to justify a RMB 1.2 billion guarantee fund has economized on the wrong line item.

The RDD row deserves a caveat. Its low bias depends on firms not manipulating their size to fall below the cutoff. In the guarantee fund, the cutoff was a registered capital threshold, which firms can adjust. An audit of the eligibility records found 14 firms whose registered capital had been reduced in the quarter before application, which is the signature of manipulation. Where manipulation is present, the RDD estimate is biased toward zero and the design loses its advantage. The matched sample, less elegant, is more robust to this particular failure.

12.5 Integration: Incentives, Metrics, and the Chain

This chapter sits between the catalysis discussion in Chapter 11 and the integration discussion in Chapter 14. Its role in the book's argument is to specify the conditions under which the three links of the chain — rating-based identification, insurance-based buffering, and sci-tech finance catalysis — transmit incentives rather than merely transfer money.

The identification link depends on the rating being informative. A green rating that can be obtained by self-assessment is not informative, and a credit allocation that depends on it will not select for environmental performance. The incentive design for identification is the verification cost schedule from Section 12.1.2: make the cost of a credible claim rise with the claim, and make the rating's use in credit decisions depend on the verification level. The comparative static in that section gives the schedule a shape: the verification threshold should fall as verification technology cheapens and should be set lower in sectors where a false claim is most damaging.

The buffering link depends on the insurance premium reflecting risk. A subsidized premium set below the actuarial rate suppresses the price signal, and the insured party has no incentive to reduce risk. The evaluation design for buffering is the loss ratio and the moral hazard estimate from Section 12.3.4: report whether the insurance changed behavior, not only whether it was purchased.

The catalysis link depends on the public capital being additional. A return cap that subsidizes private investors who would have invested anyway produces a transfer, not catalysis. The evaluation design for catalysis is the matched-sample comparison from Section 12.2.2: estimate the additionality, and report the subsidy per additional unit of investment.

The three links share a common failure mode: the metric rewards the activity and not the outcome. The common fix is the same in all three: measure the outcome, verify it at a cost proportional to the claim, and link a meaningful share of compensation to it. The guarantee fund in Section 12.3.1 is a case where the structure held in form and broke in substance — the fund had a guarantee, a subsidy, a target, and an evaluation, and every one of them was written on volume. The weather index insurance pilots in Section 12.3.4 are a case where the structure held in substance: the payout was triggered by an index that could not be relabeled, the outcome was measurable from satellite data, and the evaluation could estimate adoption and land-management effects against a staggered rollout. The instrument that could be measured was the instrument whose incentive worked.

12.6 Conclusion: The Evaluation Gap Is the Binding Constraint

Two findings come out of the evidence reviewed in this chapter.

The design of incentives and the design of evaluation are the same problem. You cannot write an incentive contract on a variable you cannot measure, and you cannot measure a variable if no one is paid to produce the data. Programs that separate the two — incentives set by one office, metrics set by another, evaluation by a third — produce the failures documented above: volume without screening, labels without emissions reductions, subsidies without additionality.

The binding constraint on policy-guided green finance is not capital. Capital is available. The constraint is the capacity to measure whether the capital did anything. China's green loan balance reached nearly RMB 12 trillion by the end of 2020 while accounting for about 5 percent of total loans; the global green bond market issued a record USD 269.5 billion in 2020. The volumes are large. The evidence that these volumes produced environmental outcomes beyond what would have happened otherwise is thin, and the thinness is a design choice, not a data limitation.

The next chapter turns to the digital infrastructure and carbon accounting systems that could relax this constraint. The question it must answer is whether better measurement capacity changes the incentive equilibrium or merely produces more precise reports of the same activities. The answer depends on whether the measurement is linked to compensation — which brings the argument back to this chapter's first section. Better data without a contract on the data changes nothing. A contract on data that no one produces is unenforceable. The two must be designed together, and in most programs they are not.

One program failed this test in a way that can be measured. The guarantee fund in Section 12.3.1 paid its managers on loan volume and default rate, and the guarantee absorbed the defaults. Three years in, the fund had disbursed RMB 1.2 billion, held a non-performing loan ratio below 1.5 percent, and could not demonstrate that a single ton of emissions had been avoided beyond what the borrowers would have done anyway. The provincial audit put the share of relabeled borrowers with no measurable improvement at 41 percent, with a confidence interval from 32 to 50. The fund's managers did what their contract asked. The contract asked for the wrong thing, and the evaluation was built to confirm it.