Chapter 12: Performance Evaluation and Incentive Design for Policy-Guided Green Instruments

A provincial green industry guidance fund closes its third year of operation. The fund has committed RMB 2.8 billion across 34 portfolio companies. Its annual report states that the portfolio avoided 410,000 tonnes of CO₂ emissions in the reporting year. The number is printed in bold on page four. On page nineteen, in a footnote set in six-point type, the report explains that the figure was calculated by multiplying each company's revenue by an emission-intensity benchmark drawn from a national industry average, then subtracting the company's self-reported emissions. Fourteen of the thirty-four companies did not report emissions at all. For those, the benchmark was used without subtraction. The 410,000 tonnes, in other words, is partly an estimate of what the companies might have emitted if they had behaved like the average firm in their sector, and partly a record of what some of them said they did.

The fund's supervisory board approved the report. The provincial finance department, which provided 40 percent of the fund's capital, cited the 410,000-tonne figure in its own year-end summary. Nobody lied. The metric was simply constructed so that it could not fail.

This chapter is about that construction. Policy-guided green instruments—guidance funds, subsidized green loans, green bond windows, guarantee schemes, dedicated green banks—are public money deployed for a public purpose. They require evaluation, and evaluation requires metrics. The metrics then shape behavior, because the managers whose compensation and future allocations depend on the numbers will optimize for the numbers. The question is not whether incentives exist. They always do. The question is whether the incentives written into the evaluation system reward the transition the instrument was created to finance, or reward the appearance of it.

Scope first. This is not a chapter about whether green industrial policy works. Chapter 11 examined the market failure in early-stage green capital and the case for public intervention. This chapter takes the intervention as given and asks a narrower question: once a policy-guided instrument exists, how should its performance be measured, and how should the people running it be paid? The two questions are inseparable. A metric that cannot be gamed is rare; a metric that is gamed will corrupt the incentive it was meant to create. Evaluation design and incentive design are the same design problem viewed from two ends.

The chapter begins with metrics and methods—what can be measured, what cannot, and where the measurement error concentrates. It then turns to incentive compatibility, drawing on the multi-task principal-agent literature that began with Holmström and Milgrom (1991), and asks how compensation and allocation can be structured so that the agent's private optimum aligns with the principal's objective. It closes with three policy instruments examined at the level of the contract, to see how evaluation and incentive design played out in practice.

12.1 Evaluation Metrics and Methods

12.1.1 What Is Being Evaluated

Before we can choose metrics, we need to fix the object of measurement. A policy-guided green instrument has at least four distinct outputs, and they are frequently conflated.

The first is capital deployed. How much money moved? This is the easiest number to produce and the least informative. A fund that deploys RMB 3 billion into a solar manufacturer that would have raised the same money from commercial banks has added nothing to the transition; it has added a line to a spreadsheet.

The second is environmental outcome. How much emissions reduction, resource efficiency, or ecosystem service did the financed activity produce? This is the number everyone wants and the number that is hardest to attribute.

The third is financial performance. Did the instrument recover its capital, cover its cost of funds, and avoid losses? A green instrument that loses money is not sustainable, whatever its environmental output; it will not be renewed.

The fourth is market transformation. Did the instrument change the behavior of private capital? Did it demonstrate a risk-return profile that commercial lenders subsequently adopted? Did it build the rating infrastructure, the insurance layer, the intermediation capacity that the transition needs?

These four outputs trade off against one another, and the trade-offs are the substance of evaluation design. A fund optimized for capital deployed will finance large, safe, late-stage projects. A fund optimized for environmental outcome will finance whatever produces the largest measurable reduction per yuan, which usually means mature abatement technologies in heavy industry, not the early-stage innovation the guidance fund was created to support. A fund optimized for financial return will behave like a commercial fund and will not touch the early-stage green capital gap at all. A fund optimized for market transformation will accept losses on demonstration projects and will be judged a failure by any auditor applying standard financial metrics.

The first design decision, then, is which of these four the instrument is for. Most policy documents say "all four." That answer is not operational. An evaluation system that weights four objectives equally gives the manager no guidance when they conflict, and they always conflict. The manager resolves the conflict privately, using whatever criterion the evaluation system does not measure. The unmeasured objective wins.

12.1.2 The Attribution Problem

Suppose a green fund finances a firm that installs a waste-heat recovery unit. The unit reduces the firm's coal consumption by 12,000 tonnes per year. How much of that reduction should be credited to the fund?

The naive answer is all of it. The fund's money paid for the unit. But the firm might have installed the unit anyway, with a commercial loan or from retained earnings; the fund's participation may have accelerated the installation by eighteen months, in which case the fund's contribution is the emissions avoided during those eighteen months, not the lifetime emissions of the unit. Or the fund's money may have crowded out a commercial loan that would have financed the same unit, in which case the fund's contribution is zero and its environmental additionality is zero.

This is the additionality problem, and it is the central methodological difficulty in evaluating any policy-guided instrument. The counterfactual—what would have happened without the instrument—is unobservable. Every evaluation method is an attempt to estimate it.

Three approaches are in common use.

Benchmark comparison. Compare the portfolio's emissions intensity to an industry benchmark. This is what the provincial fund in our opening example did. It is cheap, it requires no counterfactual construction, and it is almost meaningless when the fund selects firms that are already cleaner than the benchmark. A fund that finances only the top decile of firms by emissions intensity will show a large "avoided emissions" figure that reflects selection, not impact.

Difference-in-differences. Compare financed firms to a matched control group of unfinanced firms before and after the intervention. This requires a control group, which requires that some eligible firms were not financed—true when the fund is rationed, false when it is not. It also requires that the matching be credible. In practice, green funds tend to finance firms that are observably different from the ones they reject, and the matching variables (size, sector, age, prior profitability) rarely capture the unobservable differences (management quality, technology readiness) that drive both selection and outcome.

Structural or quasi-experimental estimation. Use an eligibility threshold, a scoring discontinuity, or an exogenous shock to identify the causal effect. This is the most credible approach and the most demanding. It requires that the instrument's allocation rule contain a discontinuity that can be exploited—a score cutoff, a size threshold, a geographic boundary. Where such a discontinuity exists, a regression discontinuity design can estimate the effect of financing on firms just above and just below the cutoff. Where it does not, the evaluator is left with weaker methods.

For most policy-guided green instruments in operation today, none of these methods is applied. The evaluation consists of summing self-reported environmental data across the portfolio, applying a benchmark where data are missing, and reporting the total. The number that emerges is not a lie, but it is not an estimate of impact either. It is an estimate of portfolio scale weighted by an assumed intensity.

12.1.3 The Metric Menu

With the attribution problem in view, we can lay out the metrics available to an evaluator and be honest about what each can and cannot support.

Table 12-1: Evaluation Metrics for Policy-Guided Green Instruments

Two columns deserve comment. The leverage ratio is the most politically salient metric and the most abused. A guidance fund that commits RMB 1 billion alongside RMB 4 billion of private capital reports a leverage ratio of 5:1. The ratio says nothing about whether the private capital would have been deployed without the public commitment. In practice, much "mobilized" private capital is capital that was already in the deal and was relabeled when the public fund joined. The leverage ratio measures the size of the syndicate, not the catalytic effect of the public money.

Avoided emissions is the metric most likely to be reported and least likely to be defensible. The calculation requires three inputs: the firm's actual emissions, a baseline emissions level, and an attribution share. Each is contestable. The baseline is usually an industry average, which embeds the assumption that the firm would have performed at the average absent the intervention—an assumption that is false for any firm that was already above average, which is precisely the firms a green fund selects. The attribution share is usually set at 100 percent, which embeds the assumption that the intervention caused the entire reduction. A defensible avoided-emissions figure would report a range, disclose the baseline and attribution assumptions, and distinguish measured from estimated components. Almost none do.

12.1.4 Rating-Based Evaluation

The rating architecture developed in Chapters 3 through 5 offers an alternative to outcome-based metrics for the evaluation of policy-guided instruments. Instead of measuring emissions after the fact, the instrument can be evaluated on the quality of the identification it performed: did it correctly distinguish green from non-green firms, and did it correctly rank them by transition potential?

This is a process metric, not an outcome metric, and it has a specific advantage. Outcome metrics are contaminated by factors outside the instrument's control—a financed firm may fail for technological reasons unrelated to the fund's screening, and its emissions may rise or fall for reasons unrelated to the fund's money. Process metrics measure what the instrument actually does: the accuracy of its green classification and the predictive validity of its transition rating.

The evaluation question becomes: of the firms the instrument rated as high-transition-potential, what fraction subsequently achieved measurable transition? And of the firms it rated low, what fraction did not? This is a classification-accuracy question, and it can be answered with the standard tools of credit scoring validation—the Gini coefficient, the Kolmogorov-Smirnov statistic, the accuracy ratio.

Formally, if we let \(y_i \in \{0,1\}\) indicate whether firm \(i\) achieved a transition outcome (say, a 20 percent reduction in emissions intensity within three years) and \(s_i\) be the instrument's transition score, the discriminatory power of the rating is measured by the area under the receiver operating characteristic curve:

An AUC of 0.5 means the rating is no better than a coin flip. An AUC of 0.75 means that a randomly chosen transitioning firm has a higher score than a randomly chosen non-transitioning firm 75 percent of the time. The Gini coefficient is \(G = 2 \cdot AUC - 1\).

This approach has a practical appeal for policy evaluation. It measures the instrument on the dimension the instrument controls—screening—and it does not require the construction of an unobservable counterfactual for each portfolio firm. It does require that transition outcomes be measured, which brings back the data problem, but the data requirement is lighter: a binary outcome per firm rather than a continuous emissions path.

The approach also has a limitation that must be stated. A rating can be accurate and still be useless if the instrument applies it to a pool of applicants that is already pre-screened by other means. If every firm that applies to a green guidance fund is already green, the rating's AUC will be high and its contribution will be zero. The evaluation must therefore be conducted on the applicant pool, not the funded portfolio. The relevant question is whether the rating correctly ranked the applicants the instrument rejected, not only the ones it accepted. This requires the instrument to retain data on rejected applicants, which most do not. The data architecture for evaluation must be designed at the same time as the instrument, not retrofitted.

12.1.5 The Timing Problem

Green transition outcomes unfold over years. A fund with a seven-year life that finances a firm in year two will observe the firm's emissions trajectory through year seven at the earliest. The evaluation that determines the fund manager's compensation is typically conducted at year five or year six, before the environmental outcome has matured.

The result is a systematic bias toward early, visible, easily counted environmental outputs and away from late, uncertain, hard-to-measure outcomes. A fund manager whose bonus depends on year-five emissions reductions will finance LED retrofits and rooftop solar, which produce immediate measurable reductions, and will decline to finance a process innovation that might reduce emissions by 40 percent in year eight. The evaluation horizon selects the technology.

There is no clean solution to this problem, but there are mitigations. One is to separate the evaluation of process from the evaluation of outcome: compensate the manager on the quality of screening and monitoring during the fund's life, and evaluate environmental outcome separately, with a longer horizon, for the purpose of renewing the instrument rather than paying the manager. A second is to use leading indicators of transition that are observable early and predictive of later outcomes—capital expenditure on decarbonization, certification progress, technology readiness level advancement—and to validate them against realized outcomes in the funds where the horizon is long enough. A third is to structure the manager's compensation with deferred components that vest after the outcome window closes, a structure familiar from private equity carried interest.

The deferred-compensation approach has a known weakness: the manager may have left the fund by the time the deferred component vests, and the deferral may be discounted to near zero by a manager with a high time preference. The leading-indicator approach has a known weakness: the indicators can be gamed if they become the basis of compensation. The process-outcome separation has a known weakness: it requires two evaluation systems, which is expensive.

We need to be clear that no design eliminates the timing problem. The choice is which distortion to accept.

12.1.6 A Composite Evaluation Framework

The discussion so far suggests a composite framework rather than a single metric. The framework has four components, each measuring a different output, each with an explicit weight set by the instrument's mandate.

Component 1: Identification quality. The AUC or Gini of the instrument's green rating, computed on the applicant pool, validated against realized transition outcomes. Weight: high for instruments whose mandate is to build the identification infrastructure.

Component 2: Financial sustainability. The instrument's net return after losses and subsidies, benchmarked against its cost of funds. Weight: high for instruments expected to revolve capital.

Component 3: Environmental outcome. Emissions reduction or resource efficiency, measured with disclosed assumptions and reported as a range, not a point estimate. Weight: high for instruments with a direct abatement mandate.

Component 4: Market transformation. Follow-on private capital in portfolio firms, adoption of the instrument's rating methodology by commercial lenders, reduction in the green premium observed in subsequent deals. Weight: high for instruments whose mandate is catalytic.

The weights are a policy choice, and they should be set explicitly, in public, before the instrument is launched. The failure mode observed in practice is not that the weights are wrong. It is that the weights are never set, and the evaluation defaults to whichever component is easiest to measure—usually capital deployed, occasionally environmental outcome computed with a benchmark.

A composite score has its own pathologies. It can average away a catastrophic failure in one component (a fund that loses 60 percent of its capital but reports high identification quality). It can be gamed by optimizing the component with the highest weight and the lowest measurement cost. It creates a single number that invites comparison across instruments with different mandates. These are real costs. They are lower than the cost of no framework at all, which is that the instrument is evaluated on the political salience of its annual report.

12.2 Incentive Compatibility

12.2.1 The Principal-Agent Structure of a Green Instrument

Every policy-guided green instrument embeds a chain of principal-agent relationships. The public principal—a finance ministry, a development bank, a sovereign fund—delegates to an instrument manager, who delegates to portfolio firm managers, who delegate to project operators. At each link, the agent has private information and private objectives, and the principal can observe outcomes only noisily.

The multi-task principal-agent model of Holmström and Milgrom (1991) is the right starting point. In their model, the agent allocates effort across multiple tasks, only some of which are measurable. If the principal compensates the agent on the measurable task, the agent reallocates effort toward it and away from the unmeasured tasks. The model's central result is that the optimal incentive intensity on any task decreases as the number of tasks and the difficulty of measuring them increase. When tasks are substitutes in the agent's effort, paying for one is paying against the others.

A green fund manager has at least four tasks: originate deals, screen for green quality, monitor environmental performance, and exit at a financial return. Only the first and fourth are cheaply measurable. Screening quality is observable only through the accuracy of the rating, which requires outcome data. Monitoring quality is observable only through the firm's environmental reports, which the manager can influence. The model predicts that a compensation contract weighted toward origination and exit will produce a manager who originates aggressively and exits early, and who treats screening and monitoring as compliance overhead.

This prediction matches what we observe. Green fund managers are typically compensated on assets under management and realized returns, with a compliance check for environmental reporting. The compliance check is binary—report filed or not filed—and does not reward accuracy. The manager's effort goes where the money is.

12.2.2 The Multi-Task Contract

The design problem is to write a contract that pays for screening and monitoring without sacrificing origination and exit. The Holmström-Milgrom framework suggests three levers.

Lever 1: Make the unmeasured task measurable. If screening quality can be measured by the rating's AUC against realized outcomes, it can be compensated. The measurement requires a data system that tracks applicants, not only portfolio firms, and that follows outcomes for a fixed window. The cost is real but bounded. The benefit is that the manager's private information about screening quality becomes observable to the principal.

Lever 2: Reduce the number of tasks or assign them to different agents. A fund that both originates and monitors creates a conflict: the origination team has an incentive to understate risk to get the deal done, and the monitoring team inherits the problem. Separating origination from monitoring—different teams, different compensation, different reporting lines—reduces the conflict, at the cost of coordination. Some development finance institutions use this structure; most green funds do not.

Lever 3: Adjust the incentive intensity to the measurement environment. When measurement is noisy, high-powered incentives amplify noise. A manager paid a large bonus on a noisy environmental metric will produce a large, noisy environmental number. A manager paid a modest bonus on the same metric will produce a modest number and will not distort the operation to produce it. The implication is counterintuitive: green instruments with poor measurement infrastructure should use low-powered incentives, not high-powered ones. The temptation is the reverse—to pay aggressively to signal commitment. The result is aggressive gaming.

12.2.3 The Gaming Problem

Gaming is not a moral failure. It is the predictable response of an agent to a metric that is easier to manipulate than the objective it proxies. Goodhart's law—when a measure becomes a target, it ceases to be a good measure—is a statement about incentive design, not about the integrity of managers.

Consider the metrics in Table 12-1 and the gaming strategies each invites.

Capital deployed invites the substitution of large, late-stage, low-risk deals for the early-stage deals the instrument was created to finance. A manager whose bonus depends on deployment will close a RMB 200 million loan to a utility-scale solar farm before closing ten RMB 20 million loans to early-stage technology firms, even if the latter have higher transition potential per yuan.

Leverage ratio invites the relabeling of private capital that was already committed. A fund that co-invests alongside a commercial bank in a deal the bank had already approved reports the bank's commitment as mobilized capital. The ratio rises; the catalytic effect does not.

Avoided emissions invites the choice of a high baseline. A firm compared to the worst-performing decile of its industry shows a large reduction; the same firm compared to the median shows a small one. The manager who chooses the baseline chooses the reported impact.

Green revenue share invites reclassification. A manufacturer of components for electric vehicles may classify its entire revenue as green, including the portion from components that also go into conventional vehicles. The taxonomy classification is a judgment, and the manager makes it.

Default rate invites the avoidance of risk. A manager whose compensation depends on default rate will finance only investment-grade green borrowers and will leave the transition's riskiest, highest-impact segments unfinanced.

The gaming strategies are not independent. A contract that weights deployment and default rate heavily will produce a portfolio of large, safe, green-labeled deals. A contract that weights avoided emissions heavily will produce a portfolio of mature abatement projects with favorable baselines. Neither portfolio will contain the early-stage, hard-to-measure, high-uncertainty firms that the transition requires and that the commercial market will not finance.

12.2.4 Incentive-Compatible Contract Design

An incentive-compatible contract for a green instrument manager must satisfy three conditions.

Condition 1: The agent's payoff increases in the principal's objective. If the principal wants transition, the manager's payoff must rise with transition, not only with deployment. This sounds trivial. It is routinely violated, because the principal's objective is multi-dimensional and the contract is single-dimensional.

Condition 2: The agent cannot increase their payoff by manipulating the metric without changing the objective. This is the informativeness principle: a metric should be included in the contract only if it carries information about the agent's action that is not already in other metrics. A metric that the agent can manipulate at low cost should be excluded or discounted.

Condition 3: The contract is robust to the agent's private information. The manager knows more about the portfolio than the principal. A contract that pays on reported outcomes without verification will be exploited by a manager who reports optimistically. Verification—third-party audit, satellite data, metered emissions—is a cost that must be built into the contract.

A contract that satisfies these conditions for a green guidance fund might have the following structure.

Base compensation covers operating costs and a modest management fee on committed capital, set below commercial fund rates to reflect the public mandate. The management fee is on committed, not deployed, capital, to reduce the origination incentive.

Screening bonus is paid on the validated AUC of the fund's green rating against realized outcomes at the portfolio firms, measured at the end of the fund's investment period. The bonus is capped and is paid only if the AUC exceeds a threshold (say, 0.65). This pays for identification quality, which is the instrument's core function.

Environmental performance bonus is paid on a verified emissions reduction metric, computed with a disclosed baseline and reported as a confidence interval. The bonus is proportional to the lower bound of the interval, not the point estimate, to penalize optimistic assumptions. Verification is by a third party with access to metered data, not self-report.

Financial performance bonus is a carried-interest-like share of realized returns above a hurdle, paid after the fund's life. The hurdle is set below commercial rates to reflect the public mandate, and the share is lower than commercial carried interest to reflect the fact that the public capital bears part of the risk.

Clawback applies if the validated screening quality or environmental performance falls below a floor at the final evaluation. The clawback is capped at the bonuses paid, not at the base compensation, to preserve the manager's participation constraint.

The contract has four components because the manager has four tasks. It pays on validated metrics rather than reported ones because the manager can manipulate reports. It defers the screening and environmental bonuses because the outcomes mature late. It caps and floors the bonuses because the measurement is noisy.

12.2.5 The Portfolio Firm's Incentives

The instrument manager is not the only agent. The portfolio firm's manager also responds to incentives, and the green instrument's contract with the firm shapes the firm's behavior.

Consider a green loan with an interest rate tied to the borrower's environmental performance—a sustainability-linked loan. The borrower pays a lower rate if it meets a pre-agreed sustainability target and a higher rate if it does not. The structure is increasingly common: the sustainability-linked loan market reached USD 1.6 trillion in cumulative issuance by 2023, according to BloombergNEF's Sustainability-Linked Loan Market Update (December 2023).

The incentive-compatible design of a sustainability-linked loan requires that the target be material (it addresses the firm's most significant environmental impact), measurable (it can be verified by a third party), and ambitious (it requires improvement beyond business as usual). The failure mode is a target that is measurable but not material or ambitious—a firm commits to reduce water use at an office building while its principal impact is emissions from a factory. The rate adjustment is small (typically 2.5 to 10 basis points), so the financial incentive is weak; the structure works through signaling and reputational channels more than through the price.

For a policy-guided instrument, the rate adjustment can be larger because the public capital can absorb the cost. A green guarantee scheme that reduces the guarantee fee by 50 basis points for firms meeting a verified transition target creates a stronger incentive than a commercial sustainability-linked loan. The design question is whether the target is set by the firm (which will set it low) or by the instrument (which will set it at a level the firm may not accept). The answer is a negotiation, and the negotiation is where the incentive design lives.

12.2.6 Common Agency and the Multiplicity of Principals

A green instrument typically has multiple principals. A provincial guidance fund may be capitalized by the provincial finance department, the provincial development and reform commission, a state-owned bank, and a national guidance fund. Each principal has its own objective: the finance department wants fiscal sustainability, the DRC wants industrial upgrading, the bank wants credit quality, the national fund wants policy alignment.

The common agency literature (Bernheim and Whinston, 1986; Martimort, 1996) shows that when multiple principals contract with a single agent, the equilibrium contract generally differs from the single-principal optimum. The agent can play the principals against each other, extracting rents by satisfying the least demanding principal's objective. The principals, anticipating this, may underinvest in monitoring or may impose conflicting requirements that the agent resolves by doing the minimum.

The practical implication for green instrument design is that the number of principals should be minimized, or their objectives should be aggregated into a single contract before the agent sees them. A fund with four principals and four reporting requirements will produce four annual reports, each optimized for a different audience, and the operational decisions will be driven by none of them. A fund with one consolidated mandate and one reporting line is easier to incentivize, at the cost of political buy-in from the principals whose objectives were merged.

There is no dominant solution. The trade-off is between political support and incentive clarity. What can be said is that the trade-off should be made explicitly. A fund that adds principals to broaden its political base without adjusting its contract is accumulating incentive problems that will surface in the evaluation.

12.2.7 The Ratchet Effect

A final incentive pathology deserves attention: the ratchet. If a manager's performance this period becomes the baseline against which next period's performance is judged, the manager has an incentive to underperform this period. The ratchet is well documented in Soviet enterprise management and in sales force compensation, and it applies directly to green instruments.

Suppose a fund manager's screening bonus is based on the improvement in the fund's AUC relative to the prior evaluation. A manager who achieves an AUC of 0.80 in year three sets the baseline for year four. If the manager expects to be evaluated against 0.80, they may report a lower AUC in year three—by classifying borderline firms conservatively—to preserve room for improvement. The conservative classification reduces the fund's reported impact, which is the opposite of what the principal wants.

The standard remedy is to compensate on the level of the metric, not its improvement, or to commit to a fixed baseline that does not update with performance. The level contract removes the ratchet but removes the incentive to improve beyond the level. The fixed baseline removes the ratchet and preserves the improvement incentive, but the baseline becomes stale as the portfolio changes. Neither is clean. The choice depends on whether the principal is more concerned about the manager sandbagging or about the manager coasting.

12.3 Case Studies of Policy Instruments

The framework is abstract until it is applied. This section examines three policy-guided green instruments at the level of the contract: a provincial green industry guidance fund, a green credit guarantee scheme, and a green bond liquidity facility. Each case is drawn from public documents and from interviews with the managers and officials involved; the names of the specific funds are withheld where the documents are not public, and the figures are rounded to preserve confidentiality. The cases are chosen to illustrate different combinations of evaluation metric and incentive structure, and different failure modes.

12.3.1 Case 1: A Provincial Green Industry Guidance Fund

The fund was established in 2018 by a provincial finance department and a state-owned investment company, with committed capital of RMB 5 billion. Its mandate was to finance early-stage and growth-stage firms in clean technology, energy efficiency, and environmental services. Sixty percent of the capital came from the provincial finance department, forty percent from the investment company. A private fund manager was selected through a competitive process to manage the fund for a seven-year term.

The contract. The manager received a management fee of 1.5 percent of committed capital per year, below the 2 percent commercial norm, and a carried interest of 15 percent of profits above an 8 percent hurdle, below the 20 percent commercial norm. The contract included an environmental reporting requirement: the manager was to report annually on the portfolio's emissions reduction, computed using a methodology to be agreed with the finance department. No environmental performance bonus was specified. The finance department's evaluation of the manager at the end of the investment period was to consider "policy alignment, financial performance, and environmental contribution," without weights.

The evaluation in practice. The manager's first annual report, for 2019, reported RMB 1.2 billion deployed across 11 portfolio companies. The environmental section reported an estimated 85,000 tonnes of CO₂ avoided, computed by applying a sector-average emissions intensity to each portfolio company's revenue and subtracting the company's reported emissions where available. Four of the eleven companies reported emissions.

The finance department's review raised three questions. First, the avoided-emissions figure was not comparable across years because the sector-average benchmark changed. Second, the four companies that reported emissions were the largest and most established, which meant the benchmark-subtraction method was applied to the firms least likely to be above the benchmark. Third, the fund had not retained data on the applications it rejected, so the finance department could not assess whether the fund's screening was adding value or simply selecting the firms that would have been financed anyway.

The manager's response was that the reporting requirement had not specified a methodology, that the data collection burden on early-stage firms was significant, and that the finance department had not provided guidance on the counterfactual. All three points were correct. The contract had created an obligation without creating the measurement infrastructure to meet it.

What went wrong. The failure was in the contract's silence. The environmental reporting requirement was written as a compliance obligation, not as a performance metric. The manager had no incentive to invest in measurement quality, because measurement quality did not affect compensation. The finance department had no basis for evaluating environmental contribution, because the contract did not specify a method. The result was a number that satisfied the reporting requirement and informed no decision.

The deeper problem was the mismatch between the fund's mandate and its incentive structure. The mandate was early-stage clean technology. The carried interest was on financial return. The manager, facing a 15 percent carry on an 8 percent hurdle, had a strong incentive to avoid early-stage technology risk. By the end of the investment period, the portfolio's average company age was 6.2 years, and only three of the fourteen portfolio companies were pre-revenue at the time of investment. The fund had become a growth-stage fund with a green label.

The redesign. In 2021, the finance department renegotiated the contract. The redesign introduced three changes. First, a screening bonus of up to 0.5 percent of committed capital per year, paid on the validated AUC of the fund's green rating against realized outcomes, with the validation conducted by an independent evaluator on the applicant pool. Second, an environmental performance bonus of up to 0.5 percent of committed capital per year, paid on verified emissions reduction measured with a fixed baseline set at the fund's inception and a third-party verification protocol. Third, a technology-stage allocation requirement: at least 40 percent of committed capital to be invested in companies less than four years old at the time of investment, with the allocation verified at each closing.

The redesign has not yet been evaluated—the fund's investment period runs through 2025—but the early signals are informative. The manager hired two additional analysts with environmental engineering backgrounds. The fund's average investment size fell from RMB 109 million to RMB 62 million. The number of applications rose by 70 percent, and the rejection rate rose from 45 percent to 68 percent. The manager's own characterization, in an interview, was that the fund had become "harder to run and more expensive to report" and that this was "the point."

The lesson. The case illustrates the Holmström-Milgrom prediction in its purest form. A contract that paid on financial return and required environmental reporting as a compliance obligation produced a portfolio optimized for financial return and an environmental number optimized for compliance. The redesign added metrics for the unmeasured tasks and paid on them, at the cost of higher measurement expense and lower financial return. Whether the trade was worth it depends on the weight the finance department places on transition relative to return—a weight the original contract never specified.

12.3.2 Case 2: A Green Credit Guarantee Scheme

The scheme was established in 2020 by a provincial government and a commercial bank to guarantee green loans to small and medium-sized enterprises (SMEs) in the environmental services sector. The government provided a guarantee fund of RMB 800 million; the bank originated the loans; the guarantee covered 70 percent of principal loss on defaulted loans. The guarantee fee was 1 percent per year, below the market rate for SME guarantees, with the difference subsidized by the government.

The incentive structure. The bank's incentive was to originate loans that would not default, because the bank retained 30 percent of the loss. The government's incentive was to maximize the green impact per unit of guarantee capital. The SME's incentive was to obtain credit at a lower cost than the market offered.

The scheme's evaluation metric was the non-performing loan (NPL) ratio of the guaranteed portfolio, with a target of below 3 percent. The government's subsidy to the bank was conditional on the NPL ratio staying below the target. The scheme's green classification was based on the bank's own assessment against a provincial green industry catalogue, with no independent verification.

The first two years. By the end of 2022, the scheme had guaranteed RMB 3.4 billion of loans to 210 SMEs. The NPL ratio was 1.8 percent, well below the target. The scheme was reported as a success in the provincial government's annual work report.

The success was real on its own terms and misleading on the scheme's mandate. The green classification was the bank's own, and the bank had an incentive to classify liberally: a loan classified as green qualified for the guarantee, and a loan not classified as green did not. The provincial green catalogue was broad—it included any firm whose revenue was more than 50 percent from environmental services, a category that in practice included waste collection, sanitation, and equipment leasing. The scheme's portfolio was concentrated in these categories, which are low-risk, asset-backed, and not where the transition's financing gap is largest.

The early-stage green technology firms that the scheme was designed to reach were largely absent. An internal review found that only 12 of the 210 borrowers were technology firms, and that these 12 accounted for 4 percent of guaranteed principal. The bank's loan officers, interviewed for the review, explained the pattern: a technology firm with no collateral and uncertain revenue could not pass the bank's credit committee even with a 70 percent guarantee. The guarantee reduces the loss given default, not the probability of default. The bank's credit committee still applied its standard cash-flow and collateral criteria, and the technology firms that the scheme was meant to reach failed those criteria before the guarantee was considered.

The incentive failure. The scheme's contract paid the bank on the NPL ratio and classified greenness by the bank's own judgment. The bank optimized both. It selectedlow-risk borrowers whose green classification was generous, and it reported an NPL ratio that validated the selection. The government's objective—financing the transition's riskiest segment—was not in the contract, so it was not in the bank's decision.

The scheme also created a crowding-out problem that the evaluation did not detect. Several of the guaranteed borrowers had previously obtained unguaranteed commercial loans at market rates. The guarantee lowered their cost of capital without changing their environmental behavior. The subsidy flowed to infra-marginal borrowers, a standard result in credit guarantee schemes documented by Honohan (2010) in a cross-country study of credit guarantee funds and by Beck, Klapper, and Mendoza (2010) in an assessment of partial credit guarantees for SMEs. The scheme's leverage ratio—RMB 3.4 billion guaranteed on an RMB 800 million fund, or 4.25:1—was reported as evidence of effectiveness. It was evidence of scale, not of additionality.

The redesign. The 2023 revision made two changes. First, the green classification was moved to an independent third-party assessor using the rating methodology described in Chapters 3 and 4, with a sample audit by the provincial development and reform commission. Second, the guarantee fee was differentiated by borrower type: 0.5 percent for early-stage technology firms, 1.0 percent for established environmental services firms, and 1.5 percent for firms in transition-intensive sectors. The differentiation was designed to redirect the subsidy toward the segment the scheme was created to serve.

The early results are mixed. The share of technology firms in new originations rose from 4 percent to 19 percent in the first year of the redesign. The NPL ratio rose to 2.9 percent, still below the target but close to it. The bank's credit committee, according to the scheme's administrator, has begun to treat the guarantee as a substitute for collateral in the technology segment, which is the behavioral change the redesign was meant to produce. Whether the NPL ratio stays below 3 percent as the technology share rises is the open question. A guarantee scheme that reaches the target segment and breaches its NPL target has traded one failure for another; the design problem is to set the NPL target at a level that reflects the risk of the mandated segment rather than the risk of the portfolio the bank would have chosen anyway.

The lesson. The case shows how a single evaluation metric—the NPL ratio—can dominate a multi-dimensional mandate. The bank was not acting in bad faith. It was responding to the only number in its contract that carried a financial consequence. The green classification, unverified, followed the path of least resistance. The redesign's independent assessment and differentiated fee are an attempt to put the missing dimensions into the contract. The NPL target remains a blunt instrument; a better design would set the target conditional on the portfolio's technology-stage mix, so that the bank is not penalized for financing the segment it was asked to finance.

12.3.3 Case 3: A Green Bond Liquidity Facility

The third case is a national-level facility established in 2021 to support the issuance of green bonds by smaller issuers. The facility provided a committed backstop purchase of up to 20 percent of any qualifying green bond at issuance, at a spread of 15 basis points over the benchmark, with the commitment expiring after two years. The purpose was to reduce the issuance cost for first-time green issuers, who faced a liquidity premium because their bonds were small and thinly traded.

The evaluation design. The facility's performance was evaluated on three metrics: the number of qualifying issuances, the average issuance spread relative to a matched sample of conventional bonds, and the volume of secondary-market trading in the supported bonds after the commitment expired. The first two were measured at issuance; the third was measured at the end of the second year.

The design is notable because it attempted to measure market transformation directly—the trading volume metric was intended to capture whether the facility had built a liquid market for small green bonds, not merely subsidized a few issuances. The metric was chosen with the correct intuition: a liquidity facility that leaves the market as illiquid as it found it has not transformed anything.

The outcome. In the first eighteen months, the facility supported 47 issuances totaling RMB 28 billion. The average issuance spread was 22 basis points below the matched conventional sample, a larger discount than the 15 basis point facility spread, which suggested that the facility's commitment was valued beyond its direct cost—issuers and investors priced the backstop as a signal of quality. The trading volume metric, however, told a different story. The supported bonds traded at an average of 0.8 percent of outstanding volume per month in the second year, compared to 2.4 percent for a matched sample of larger green bonds and 1.1 percent for matched conventional bonds of similar size. The facility had lowered the cost of issuance without improving liquidity.

The interpretation. The facility's design created an incentive for issuance, not for liquidity. An issuer who wanted the spread benefit had to obtain the commitment; once the bond was issued and the commitment expired, the issuer had no reason to support trading. The investors who bought the bonds at issuance had obtained the spread benefit and had no obligation to make a market. The facility had subsidized the primary market and left the secondary market unchanged.

The evaluation caught this because it measured the right thing. Many facilities would have reported the 47 issuances and the 22 basis point spread reduction and declared success. The trading volume metric, specified in advance, made the failure visible.

The redesign. The 2023 revision added a market-making obligation: issuers receiving the facility commitment were required to designate a market maker for the bond for the two-year commitment period, with the market maker obligated to quote two-way prices within a specified spread. The facility paid the market maker a fee of 5 basis points per year on the outstanding volume. The trading volume metric was retained and tightened: the target was raised from 0.8 percent to 1.5 percent of outstanding volume per month.

The redesign converts the facility from a primary-market subsidy into a market-structure intervention. It costs more, because the market-making fee is a recurring expense rather than a one-time spread concession. It addresses the actual constraint—the absence of a market maker for small green bonds—rather than the symptom the facility was originally designed to treat.

The lesson. The case is the most encouraging of the three, because the evaluation metric was specified before the facility was launched and was capable of detecting failure. The facility's initial design was wrong, but the evaluation was right, and the redesign followed from the evaluation. The general principle is that the value of an evaluation system lies not in the number it produces but in its capacity to falsify the instrument's theory of change. A metric that can only confirm success is not an evaluation; it is a reporting exercise.

12.3.4 Cross-Case Observations

Three observations emerge from the cases.

First, the contract's silences matter more than its explicit terms. In Case 1, the environmental reporting requirement was explicit and the methodology was silent; the manager produced a number that satisfied the requirement and informed nothing. In Case 2, the green classification was explicit and the verification was silent; the bank classified liberally. In Case 3, the trading volume metric was explicit and the market-making obligation was absent; the facility subsidized issuance without building liquidity. In each case, the failure was located in what the contract did not say.

Second, the evaluation metric and the incentive contract are the same object viewed from different angles. Case 1's manager was paid on financial return and evaluated on environmental contribution; the two were disconnected, and the manager optimized the one that paid. Case 2's bank was paid on the NPL ratio and evaluated on green volume; the two were connected in the wrong direction, and the bank optimized the NPL ratio by selecting the wrong borrowers. Case 3's facility was evaluated on trading volume and paid on issuance; the evaluation was right and the payment was wrong, and the market outcome followed the payment.

Third, redesign is possible and expensive. All three instruments were revised within three years of launch. The revisions added measurement infrastructure, independent verification, and differentiated pricing. They raised the administrative cost of the instruments by an estimated 15 to 30 percent. The cost is the price of connecting the evaluation to the incentive. An instrument that is cheap to administer and impossible to evaluate is not cheap; it is expensive in a currency that does not appear on the budget line.

12.4 The Evaluation-Incentive Loop

The three cases suggest a general structure that we can now state explicitly. Evaluation and incentive design form a loop, not a sequence. The evaluation metric determines what the agent optimizes; the agent's optimization determines what the evaluation observes; the observation determines the next period's contract. The loop can be virtuous or vicious, and the difference is whether the metric is connected to the objective in a way that survives the agent's response.

flowchart TD
    A[Public mandate: transition outcome] --> B[Evaluation metric]
    B --> C[Incentive contract]
    C --> D[Agent behavior]
    D --> E[Observed outcome]
    E --> F[Validation against objective]
    F --> G{Metric still informative?}
    G -->|Yes| B
    G -->|No| H[Redesign metric and contract]
    H --> B
    D --> I[Gaming strategies]
    I --> F

The loop has a stability condition. A metric is incentive-stable if the agent's optimal response to the incentive contract produces an outcome that the principal would have chosen absent the contract. Formally, let \(a\) be the agent's action, \(e(a)\) the true objective, and \(m(a)\) the measured metric. The contract pays \(w(m(a))\). The metric is incentive-stable if

for all actions available to the agent. In practice, the equality rarely holds exactly. The design problem is to choose \(m\) and \(w\) so that the divergence between the two argmax sets is small relative to the stakes.

Three design principles follow.

Principle 1: Measure the objective, not the input. Capital deployed, leverage ratio, and number of issuances are inputs. Emissions reduction, transition progress, and market liquidity are objectives. An input metric is stable only if the input is perfectly correlated with the objective, which it never is. Where the objective is unmeasurable, measure a proxy and validate the proxy against the objective at regular intervals.

Principle 2: Pay on validated metrics, not reported ones. The agent reports the metric; the principal validates it. Validation can be third-party audit, satellite or metered data, or statistical comparison against independent sources. The cost of validation is part of the instrument's design cost, not an overhead to be minimized.

Principle 3: Set the incentive intensity to the measurement quality. High-powered incentives on noisy metrics produce noise amplification. Low-powered incentives on precise metrics leave value on the table. The intensity should rise with the precision of the metric and fall with the number of tasks the agent performs. This is the Holmström-Milgrom result applied to green instruments, and it cuts against the intuition that commitment to the transition requires aggressive pay-for-performance.

12.5 The Political Economy of Evaluation

The design principles are necessary and not sufficient. An evaluation system operates in a political environment, and the environment shapes what can be measured and what can be said.

A provincial finance department that has committed RMB 3 billion to a green fund has an interest in the fund's success. The evaluation it commissions is not a neutral scientific exercise; it is an input to a decision about whether to renew the fund, and the officials who commissioned it are the officials who will defend the renewal. The evaluator knows this. The result is a predictable drift toward metrics that confirm and away from metrics that falsify.

This is not a story about individual integrity. It is a story about institutional incentives, and it applies to the evaluator as much as to the fund manager. An independent evaluator who reports that the fund failed to achieve additionality will not be re-engaged. An evaluator who reports that the fund achieved a 4.25:1 leverage ratio will be. The market for evaluation services selects for confirmation.

Three institutional designs can counteract the drift.

Pre-commitment. The evaluation methodology, including the metrics, the baselines, the validation protocol, and the reporting format, is specified and published before the instrument is launched. The evaluator is bound by the pre-committed methodology and cannot adjust it after seeing the results. Case 3's trading volume metric is an example; it was specified in advance and it falsified the facility's theory of change. Pre-commitment does not eliminate the political pressure, but it raises the cost of yielding to it: changing the methodology after the fact is visible.

Adversarial evaluation. Two evaluators are commissioned, one by the instrument's sponsor and one by an independent body—a legislature, a central bank, an academic institution. The two reports are published together. The adversarial structure does not guarantee truth, but it guarantees that a confirming evaluation will be contested. The cost is duplication, and the benefit is that the sponsor cannot control the narrative.

Outcome-linked renewal. The instrument's renewal is conditioned on pre-specified outcome thresholds, not on a discretionary review. If the fund's validated AUC falls below 0.60, or its verified emissions reduction falls below the lower bound of the pre-committed range, the fund is not renewed. The condition is written into the instrument's establishing document, so that the renewal decision is a rule rather than a negotiation. The rule can be waived, but the waiver is a public act.

None of these designs is common in current practice. Most green instruments are evaluated by the sponsoring agency, using a methodology agreed after launch, with renewal decided through a budgetary negotiation. The result is that the evaluation loop in the diagram above is open at the validation step: the observed outcome is not compared against the objective in a way that can trigger redesign. The loop becomes a reporting cycle.

12.6 Designing the Evaluation System: A Practical Sequence

For an instrument being designed now, the sequence of decisions is as follows. We set it out as a sequence because the decisions are interdependent and the order matters: a metric chosen before the mandate is fixed will be the wrong metric, and a contract written before the metric is chosen will pay for the wrong thing.

Step 1: Fix the mandate in one dimension. The instrument's primary objective should be stated as a single dimension—transition outcome, or financial sustainability, or market transformation—with the others as constraints. A mandate with two primary objectives has no primary objective. The provincial fund in Case 1 had a mandate that read as three objectives and functioned as a financial-return mandate, because financial return was the only one with a number attached.

Step 2: Choose the metric that measures the mandate, and specify its validation. If the mandate is transition outcome, the metric is emissions reduction or a transition-progress index, validated by third-party verification against a pre-committed baseline. If the mandate is identification quality, the metric is the AUC of the green rating against realized outcomes, validated on the applicant pool. The validation protocol is part of the metric definition, not an afterthought.

Step 3: Estimate the measurement cost and build it into the instrument's budget. Measurement is not free. Third-party verification, data systems, and evaluator fees typically add 1 to 3 percent of the instrument's capital to its lifetime cost. An instrument designed without this line item will under-measure, and under-measurement will be resolved by substituting self-reported data, which is the failure mode of Case 1.

Step 4: Write the incentive contract on the validated metric, with intensity set to measurement quality. If the metric is precise and independently verified, the incentive can be high-powered. If the metric is noisy or self-reported, the incentive should be low-powered, and the contract should invest in improving the metric before raising the intensity. The sequence is measure, then pay; not pay, then measure.

Step 5: Specify the evaluation horizon and the deferral structure. Outcomes that mature after the manager's tenure require deferred compensation or a separate outcome evaluation. The choice between the two depends on whether the principal can commit to paying a departed manager, which depends on the instrument's legal form.

Step 6: Pre-commit the renewal rule. The instrument's establishing document should specify the outcome thresholds at which renewal is automatic, at which it is discretionary, and at which it is denied. The rule should be published. The publication is the commitment device.

Step 7: Commission the adversarial evaluation. At least one evaluator should be independent of the sponsoring agency, and the reports should be published together. The cost is small relative to the instrument's capital and large relative to the value of an uncontested evaluation.

The sequence is a template, not a straitjacket. Its purpose is to force the design team to confront the trade-offs in order, rather than discovering them after the instrument is launched and the contract is signed.

12.7 What the Rating-Insurance-Catalysis Sequence Implies for Evaluation

The integrated mechanism developed in this book—rating-based identification, insurance-based risk buffering, sci-tech finance catalysis—has a specific implication for evaluation design that is worth stating separately.

In the sequence, the three components are complements, not substitutes. Rating identifies the firm; insurance buffers the risk that the rating cannot eliminate; catalysis channels capital to the identified and buffered firm. The performance of the sequence is not the sum of the components' performances. It is the performance of the chain, and a chain is as strong as its weakest link.

This means that the evaluation of a policy-guided instrument that implements the sequence must measure linkage, not only component performance. Three linkage metrics matter.

Rating-to-insurance linkage. What fraction of firms rated as high-transition-potential obtained insurance, and at what premium? A high rating that does not translate into insurability is an identification that the risk market does not accept. The gap between the rating and the insurance uptake is a measure of the rating's credibility to the risk carrier.

Insurance-to-credit linkage. What fraction of insured firms obtained credit, and at what spread? The insurance is valuable only if it changes the lender's decision. A premium paid by a firm that would have obtained credit anyway is a transfer, not a buffer.

Credit-to-transition linkage. What fraction of credited firms achieved verified transition outcomes? This is the terminal metric, and it closes the loop to the rating: the firms that the rating identified, the insurance buffered, and the credit financed should be the firms that transitioned. If they are not, one of the links is broken.

The three linkage metrics can be computed from the instrument's own records if the records are designed to track firms across the sequence. Most instruments do not track across the sequence, because the rating, the insurance, and the credit are administered by different agencies with different data systems. The evaluation of the sequence therefore requires a data architecture that spans the agencies—a topic developed in Chapter 14.

There is a further implication for incentive design. If the sequence is the instrument's theory of change, the instrument manager should be compensated on the sequence's terminal outcome, not on the volume of any single component. A manager paid on the number of ratings issued will issue ratings; a manager paid on the number of insured firms will push insurance; a manager paid on the transition outcome of the fully linked firms will build the linkages. The multi-task problem reappears at the level of the sequence, and the solution is the same: pay on the objective, validate the metric, and set the intensity to the measurement quality.

12.8 Objections and Qualifications

Three objections to the framework developed in this chapter deserve a response.

Objection 1: The measurement cost is too high for small instruments. A RMB 200 million green guarantee scheme cannot afford a 3 percent measurement budget and an independent evaluator. The framework is designed for instruments of scale.

The objection is correct and the response is proportional. A small instrument should not attempt the full framework. It should choose one metric, validate it cheaply—a sample audit rather than a full verification—and accept a lower incentive intensity. The alternative, which is common, is to choose no metric and rely on the sponsoring agency's judgment. The judgment is cheaper and less informative. The proportional framework is a middle path: one metric, cheap validation, modest incentive.

Objection 2: The framework assumes the principal knows the objective. In practice, the principal is a coalition of agencies with different objectives, and the objective emerges from a political process that the evaluation cannot capture.

This is the common agency problem discussed in Section 12.2.6, and the objection is valid. The response is that the framework does not require a single objective; it requires that the objective be stated. A coalition that states its objective as a weighted combination of transition outcome and financial return, with the weights published, can use the framework. A coalition that states its objective as "policy alignment, financial performance, and environmental contribution," without weights, cannot—and the Case 1 fund is the illustration of what happens next. The framework's demand is procedural, not substantive: state the objective, whatever it is, and measure it.

Objection 3: Validation is gameable too. The agent can influence the validator, select the validation sample, or time the validation to a favorable period.

True. Validation raises the cost of gaming; it does not eliminate it. The response is to make the validator independent, to randomize the validation sample, and to conduct validation at pre-committed times. Each of these is a partial remedy. The general point is that no evaluation system is game-proof, and the design objective is to raise the cost of gaming above the cost of performing. An agent who can game a metric at low cost will game it; an agent who must corrupt an independent validator, at reputational and legal risk, may not.

12.9 The Evaluation of the Evaluator

A final reflexivity. The evaluation system itself is an instrument, and it should be evaluated. The questions are the same ones we have asked of the green instruments: does the evaluation system change behavior in the intended direction, and can it be gamed?

The evidence from the cases is that evaluation systems change behavior, sometimes in the intended direction and sometimes not. Case 1's original evaluation changed the manager's behavior toward financial return and away from early-stage technology. Case 3's evaluation changed the facility's behavior toward market-making, once the metric was tightened. The evaluation system is not a neutral observer; it is part of the intervention.

The implication is that the evaluation system should be designed with the same care as the instrument, by the same team, at the same time. In practice, evaluation is often an afterthought—a reporting requirement added at the end of the design process, when the contract is already written and the incentive structure is fixed. The afterthought produces the compliance number. The design-time evaluation produces the incentive.

The provincial fund in our opening example reported 410,000 tonnes of avoided emissions. The number was constructed so that it could not fail. The redesign described in Case 1 replaced it with a verified metric that could fail, and the fund's manager responded by changing the portfolio. That is the test of an evaluation system: not whether it produces a number, but whether the number can be wrong, and whether anyone acts when it is.

The evaluation framework in this chapter sets the requirements. Meeting them will require data systems that track firms across the rating, insurance, and credit sequence—systems that most instruments do not yet have, and whose design is the subject of Chapter 14.