Policy Research Working Paper 10558 Could Sustainability-Linked Bonds Incentivize Lower Deforestation in Brazil’s Legal Amazon? Dieter Wang Bryan Gurhy Marek Hanusch Philipp Kollenda Macroeconomics, Trade and Investment Global Practice & Finance, Competitiveness and Innovation Global Practice September 2023 Policy Research Working Paper 10558 Abstract This paper proposes a new relative evaluation and bench- by accounting for the real effective exchange rate, global marking framework for performance linked financing commodity prices, and prevalent deforestation trends. The instruments. It argues that the carrots and sticks of sus- results show that policy efforts helped lower deforestation tainability-linked bonds should not use key performance in the 2000s, even after accounting for external factors. The indicators which are solely tied to outcomes. Instead, they trend reversal and acceleration in deforestation since 2012 should be based on its issuer’s level of performance with are partly due to weaker policy and macroeconomic fac- respect to a target. The paper defines performance as the tors. Based on these results, the paper proposes an Amazon part of the outcome that the issuer can influence. Otherwise, sustainability-linked bond, which could allow for a more the issuer may be rewarded or penalized for factors outside effective mechanism to incentivize policy efforts. The paper their control. In such a case, principal-agent theory would also introduces the feasibility and ambitiousness matrix to predict a dilution of the performance-based instrument’s set sustainability performance targets. The matrix is used to incentives. This framework is then applied to deforesta- define the terms low-hanging fruits and long shots and to dis- tion in Brazil’s Legal Amazon and estimate performance cuss why such targets are subject to the risk of greenwashing. This paper is a product of the Macroeconomics, Trade and Investment Global Practice and the Finance, Competitiveness and Innovation Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at dwang5@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Could Sustainability-Linked Bonds Incentivize Lower Deforestation in Brazil’s Legal Amazon? Dieter Wang* 1 , Bryan Gurhy1 , Marek Hanusch1 , and Philipp Kollenda1 1World Bank Group JEL codes: G23, H63, Q23, Q56 Keywords: Sustainability-linked bonds, performance-based finacing, benchmarking, Amazon deforestation, feasibility and ambitiousness. * Corresponding author. E-mail: dwang5@worldbank.org. This paper is a further refined version of a technical background paper for (Hanusch, 2023) "A Balancing Act for Brazil’s Amazonian States: An Economic Memorandum". The authors are grateful for insightful comments provided by André Rodrigues de Aquino, Steen Byskov, Rodrigo Silveira Veiga Cabral, Doerte Doemeland, Jose Antonio Gragani, Farah Hussain, David James Kaczan, Werner Kornexl, Paul Jonathan Martin, Habib Rab, James Seward, Anderson Caputo Silva, Daniel Navia Simón, Fiona Stewart, Claudia Tufani, Carla Cristina Solis Uehara, Niraj Verma, Paolo Verme, Thales Pupo West. We also thank the participants of the “Measuring Development 2023” conference for their feedback. 1 Introduction With sustainability becoming mainstream in financial markets, some investors are increasingly keen to align traditional investment goals with those of responsible or impact investing. Since the European Investment Bank issued the first green bond in 2007, the World Bank in 2008, and the International Finance Corporation in 2013, the green, social, sustainable bonds (GSSB) market has grown exponentially. With GSSBs, also called use-of-proceeds bonds, the proceeds are earmarked to fund predefined projects. However, these instruments face criticism from both issuers and investors, including concerns about “greenwashing” when targets are set too low, the rigidity of the use-of-proceeds model (especially for sovereign issuers), and the lack of measurable additionality. In this environment, issuers and investors are considering alternative mechanisms that tie financing to impact on the ground. A key question is how the investment’s impact should be measured, which is particularly challenging at the sovereign level. Sustainability-linked bonds (SLBs) 1 have emerged as promising instruments to meet this demand (Flugge et al., 2021; Lindner et al., 2023). The instrument is already well established in the corporate market, with US$225 billion SLBs outstanding as of March 2023. The experience of SLBs in the corporate debt asset class provides a motivation to start with high standards for the public sector, given the potential of these instruments to induce changes in public policy (Silva et al., 2023; SSDH, 2023). While market standards for sovereign SLBs (in terms of structure) have not been established, two successful issuances of sovereign SLBs exist as of the time of writing. Chile issued an SLB in March 2022 with a step-up coupon structure linked to greenhouse gas emissions targets and non-conventional renewable energy generation (Ministry of Finance of Chile, 2022). Chile’s structure, with a coupon step-up of 25 basis points, has been the standard in corporate markets (Lindner et al., 2023). However, in October 2022 Uruguay issued an SLB with both a step-up and step-down coupon structure (Ministry of Economy and Finance of Uruguay, 2022). Which payoff structures will become standard in sovereign SLB issuances remains to be seen. The case for sustainability-linked bonds A key difference between SLBs and labeled bonds is that the proceeds of SLBs are not earmarked. Instead, their proceeds can be accounted for like conventional bonds and added to the general government budget. The main stipulation is that the SLB’s coupons are linked to a key performance indicator (KPI), which tracks the issuer’s progress towards a sustainability performance target (SPT). The issuer is incentivized to perform 2 through financial carrots and sticks: missing the target leads to an interest rate penalty (step-up), while achieving the target leads to an interest rate reward (step-down). 3 The less restrictive use of proceeds gives SLBs distinct advantages over labeled bonds. First, the issuer can use the proceeds to fund policies and programs that are not directly linked to an SPT but may benefit it indirectly (Lindner et al., 2023). For instance, a labeled bond intended to curb deforestation in the Amazon may have its proceeds earmarked to support local law enforcement. In contrast, the proceeds of an SLB could be used partly to enhance forest governance and partly to boost productivity, which will foster economic development and shift the economy away from resource extraction, including deforestation, in the longer run (Hanusch, 2023; World Bank 1 In this paper we focus on sustainability-linked bonds but the same insights hold for sustainability-linked loans. 2 We use the term perform to describe efforts the issuer undertakes to reach the SPT. These can include passing necessary legislation, enforcing existing regulations, investing in structural reforms or diversifying the economy. 3 Step-ups and step-downs refer to the practice of increasing or lowering coupon payments of bonds. A step-down means that the debtor pays a lower interest rate, which serves as an incentive to achieve the SPT. Conversely, a step-up would entail higher interest payments. SLBs could be issued with either or both incentive structures, depending on the case. 2 Group, 2023). Such performance-based financing can both be effective in setting incentives and an attractive addition to governments’ fiscal instruments. Second, the possibility to scale the labeled bond market depends on the number of eligible projects. This is not the case for SLBs, which gives them more scalability and also the ability to mobilize more private capital. The financial incentives in the SLB structure assure that issuers have “skin-in-the-game” as achiev- ing the SPT is in their financial interest. This gives SLBs a third powerful advantage: they allow issuers to adapt and explore policy measures during the SLB’s lifetime. From the point of view of economic theory, the SLB structure is an elegant solution to the problem of incomplete contracts. Since it is impossible to completely specify how the issuer should act in all future scenarios to reach the targets, theory suggests that targets can also be achieved if the decision power is delegated to the issuer, as long as the issuer’s incentives are aligned with the target. This decision delegation ensures that issuers will react to a changing environment and adapt policies to achieve the SPT, without having to stipulate how the funds should be used ex-ante. Choosing the appropriate KPI The effectiveness of an SLB in incentivizing issuer performance rests upon the choice of KPIs. We differentiate between outcome KPIs, where performance is measured only using the outcome indicator of interest, and benchmarked KPIs, where performance is assessed relative to a benchmark model. The idea behind a benchmark is to decompose an observed outcome into a performance part, which can be attributed to the issuer’s actions, and an exogenous part that is not in the issuer’s control. In Section 4, for example, lower deforestation may be due to effective government policies, e.g. designating protected areas, or the consequence of exogenous factors that render deforestation unprofitable, e.g. lower global commodity prices or unfavorable effective exchange rates (Arcand et al., 2008; Carvalho et al., 2019; Hanusch, 2023), or both. The benchmark level of deforestation reflects the expected amount of forest loss if the issuer would not undertake any additional efforts to avoid deforestation (see Figure 5). In Section 2, we introduce the Relative Evaluation And benCHmarking (REACH) framework. We discuss the three requirements (R1)-(R3) that benchmarks ideally satisfy in order to facilitate both goals: setting strong performance incentives for the issuer and obtaining evidence-based additionality metrics for the investor. We further explain why all three requirements are necessary in the context of the informativeness principle (Holmström, 2016). An SLB whose KPIs are assessed relative to a benchmark will also have sustainable performance targets that are defined in relative terms. We acknowledge that giving up absolute targets in favor of relative targets may lead to some level of unease as absolute SPTs are simple and easy to understand. However, they dilute performance incentives and hamper impact evaluation, which in turn weakens the SLB’s effectiveness and signaling effect. Moreover, absolute targets pose practical challenges. Suppose we target zero deforestation by 2030. How ambitious or feasible is this target? To answer this question, one would need deforestation predictions until 2030 under the business-as-usual scenario. Since deforestation is path-dependent and influenced by natural and economic factors with feedback effects, this exercise is far from trivial. This does not mean that absolute targets are irrelevant and as we will demonstrate in Section 5 they can for example be used to evaluate near-term targets. Using a benchmark model to deduce performance may not be common in sustainability finance, but the concept is widely used in other areas of finance. Our proposed way of benchmarking performance finds conceptual similarities in factor attribution and empirical asset pricing. For example, the financial performance of investment funds is not assessed in absolute terms. Instead, actively managed funds or passive index funds are evaluated with respect to a benchmark index 3 Figure 1: The components of the REACH framework The framework defines key performance indicators (dark green bars) as the difference between the observed outcome (black, solid line) and the benchmark outcome (blue, dashed line). The difference is zero if the issuer did not perform and generated no additional impact. In this example, the issuer performed consistently (except for period 4) and reaches the benchmarked SPT after five periods. Component Example Benchmark outcome Expected deforestation Observed outcome Observed deforestation = outcome KPI Benchmark – observed Avoided deforestation = benchmarked KPI Sustainability performance Forest preservation goal target (SPT) over five years Note: The benchmark outcome is a counterfactual and is not observed. It needs to be estimated by a statistical benchmark model. 0 1 2 3 4 5 Source: Authors. (e.g. the “market portfolio” in the capital asset pricing model), such as the S&P500 or the MSCI World Index in the equity space. Good performance then means over-performing the benchmark (generating “alpha”), which is very similar to our benchmarking approach. However, while portfolio managers are concerned with risk-adjusted returns, we are interested in additionality and measurable impact. The remainder of this paper is structured as follows. Section 2 introduces the REACH framework and gives examples for each of its components. It stipulates the three requirements for the benchmarking approach. Section 3 lays the theoretical foundation and formalizes the rationale behind paying for performance instead of outcomes. Section 4 shows an application of the REACH framework to Brazil’s Legal Amazon. Section 5 introduces the feasibility and ambitiousness matrix (FAB) and illustrates how to identify ambitious but feasible targets to curb deforestation. Section 6 discusses how the REACH framework and FAB matrix can be used for other instruments and support the development of SLBs in general. 2 The REACH framework The REACH framework consists of five components, as shown in Figure 1: (1) observed outcome, (2) benchmark outcome, (3) benchmark model, (4) benchmarked KPIs, and (5) benchmarked SPTs. This framework is consistent with the International Capital Markets Association (ICMA) guidelines for SLBs and can be seen as a special case of it, where benchmarking takes up a much more central role. In this section, we describe each component and provide examples in the context of deforestation, which sets the stage for the subsequent application in Section 4. 2.1 Observed outcome The observed outcomes are indicators, variables, or measurements that constitute the basic build- ing blocks of the framework. These indicators are objective measures of the sustainability issue 4 of interest. They are usually collected by national statistical offices, government ministries, uni- versities, international organizations or private corporations. Observed outcomes need to meet minimum data quality standards such as sufficient coverage, availability of historical data, acces- sibility to the wider public and transparency about their methodology. Example Annual primary forest loss in km2 . 2.2 Benchmark outcome The benchmark outcomes represent what outcome we would expect without policy intervention in the absence of the SLB issuance. However, the benchmark outcome is not observed because we only observe the outcome with policy intervention. Therefore, this counterfactual outcome has to be estimated using a statistical model, possibly incorporating data from other countries or other indicators. An appropriate, interpretable, robust and transparent benchmark model is essential for the credibility and effectiveness of the REACH framework. Example Expected annual primary forest loss in km2 , controlling for existing policy, earlier deforestation, previous exchange rates and commodity prices. 2.3 Benchmark model To produce a credible counterfactual, the benchmark model needs to provide clear and convincing reasoning why its results can be used for performance evaluation. 4 We suggest that benchmark models should fulfill the following three requirements: (R1) Transparent benchmark model. The benchmark model should articulate which factors it considers as being in the issuer’s control and which it considers exogenous. This clarifies immediately how the SLB defines performance towards the SPT. A transparent model spec- ification is also beneficial for communication and instrument pricing and allows the desired level of performance to be specified by choosing the cutoffs and/or step sizes (see Table 1) given historical performance. (R2) Benchmark probability distribution. The estimated benchmark outcome should be ac- companied by a probability distribution around the preferred estimate to understand the likelihood of alternative outcomes. This is necessary to clarify whether the measured per- formance is due to chance or constitutes actual performance beyond a reasonable statistical doubt. It is worth emphasizing that the benchmark distribution is not merely a tool for performance evaluation. A good understanding of uncertainty around the performance measure is also necessary for accurately setting performance incentives (see Section 3). (R3) Justification for additionality. The benchmark model needs to be supported by a rigorous, evidence-based justification for why it can be used to measure performance and additionality. Ideally, these claims are supported by historical evaluations and back-testing results. Two possible justifications for additionality can be derived from the informativeness principle (see Section 3.4): first, using the controllability principle which stipulates that incentives should only depend on influenceable factors, and second, through relative performance evaluation, where the performance can be gauged through the behavior of relevant peers. 4 Antle et al. (1988) define performance evaluation in their model as being about “producing information relevant to the question of whether the desired behavior was supplied”. 5 Example The benchmark model in Section 4 recognizes the fact that deforestation is not only determined by the government’s policy efforts but also by factors outside the government’s con- trol. This provides a justification via the controllability principle, satisfying (R3). The econometric model (7) articulates which and how external factors are incorporated in the counterfactual bench- mark calculation, which satisfies (R1). Finally, Figure 6 demonstrates how the model can produce a benchmark distribution, as required by (R2). 2.4 Key performance indicator KPIs in REACH serve a more fundamental purpose than only tracking the issuer’s progress towards the SPT. They also measure impact and additionality with a claim to causality (see Section 3.5). More importantly, a well-chosen KPI sets strong incentives for the issuer to act. In Section 3 we will explain why KPIs that only depend on the observed outcome (what we call outcome KPIs) may not set the strongest incentives for the issuer to perform. outcome KPI = observed outcome Instead, the REACH framework advocates for benchmark model-based KPIs, or simply bench- marked KPIs, which are defined as the difference between the observed outcome and the estimated benchmark outcome. benchmarked KPI = observed outcome − benchmark outcome Consequently, benchmarked KPIs have a natural interpretation: The KPI is zero, if there is no evidence of performance. If the observed outcome is better than the benchmark, the KPI will be positive. If it is worse, the KPI will be negative. While we introduced the benchmarking concept in the context of an SLB, the ability to estimate performance can be used for any performance-based financing instrument. Example for outcome KPI: Annual avoided primary forest loss in km2 . Example for benchmarked KPI: Annual avoided primary forest loss in km2 above the benchmark. 2.4.1 What differentiates performance from chance? A positive KPI suggests performance, but it does not imply that it is significant enough to warrant a financial reward. For instance, if observed deforestation is only slightly below the benchmark, it is difficult to interpret this as a strong signal for additionality. An investor may argue that this minor improvement was merely the result of chance rather than policy efforts. Similarly, if observed forest loss is only slightly worse than expected, it does not necessarily mean that the issuer did not exert sufficient effort. The issuer can credibly attribute the negative KPI to chance. Clear identification of effort is important to generate time-consistent government incentives. Without it, governments may not feel sufficiently incentivized at times when exogenous deforestation pressures are strong and deforestation rises in spite of their efforts to prevent even worse outcomes. In the REACH framework, the terms “slightly” and “chance” can be defined in statistical terms thanks to requirement R2, because each benchmark model should be accompanied a benchmark distribution. While the benchmark outcome describes the outcome level we would expect, the benchmark distribution tells us how likely alternative outcomes are. We can then use the bench- mark distribution to assess whether the signal from the KPI is weak (within the 50% confidence 6 Figure 2: Actual performance or coincidence? In Section 2.4.1 we described the need for benchmark models to be able to produce benchmark distributions. These distributions are centered around the benchmark outcome and help us understand if an observed outcome and its associated KPI could be due to chance (within the buffer zone) or constitutes a significant signal for additional performance beyond any reasonable doubt. The 50% and 95% cutoffs are illustrative and can be adjusted in practice. Outcome is Outcome is worse significantly better than benchmark than 95% confidence Outcome interval variable 50% confidence interval Very strong KPI Weak negative KPI Strong negative KPI Very strong negative KPI Source: Authors. interval) or significant (outside the 95% confidence interval). 5 A KPI that falls inside the 50% confidence interval does not make a strong case for either financial reward or penalty because its signal is too weak to rule out the possibility of chance. When a KPI falls outside the 50% confi- dence interval, it is much harder to attribute the result to chance. Hence, a step-up or step-down may be warranted. A KPI that falls outside the 95% confidence interval is a statistically significant signal for over- or underperformance beyond a reasonable doubt. A step-up or step-down is certainly warranted. Table 1 shows a possible incentive structure, where the size of the step-up or step-down scales with the strength of statistical evidence for performance. 2.5 Sustainability performance target An SLB reaches its SPT when the cumulative benchmarked KPIs cross a specific threshold (see lower panel of Figure 1). Targets, therefore, share the same unit as benchmarked KPIs and are expressed in relation to the benchmark outcome. Note that while benchmarked SPTs are set in absolute numbers, they are not interpreted in absolute terms since they are a sum of benchmarked KPIs. 6 The key challenge is how to set targets that are ambitious and feasible. We return to this question in Section 5. Example 10,000 km2 of avoided primary forest loss above benchmark over the next five years. 5 While confidence intervals of 50% and 95% are common reference points in statistical analysis, the precise boundaries may differ for each SLB issuance and change during the structuring process. 6 Some financial products may evaluate social or environmental performance on a yearly basis, combined with a yearly payout structure. SPTs would then also be annual, rather than cumulative. In practice, most instruments will define periodic evaluation windows that are less frequent than annually which necessitates expressing SPTs as cumulative benchmarked KPIs. 7 Table 1: Linking step-up and step-downs to KPI strength This table illustrates how different KPI strengths can be translated into correspondingly sized financial incentives in terms of basis points (bps) changes. Practical implementations may choose cutoffs other than the symmetric 50% and 95% points and/or adjust the step sizes to reflect the goals of the instrument. Cutoffs Within 50% Between 50% and 95% Outside of 95% Description Observed outcome is Observed outcome is Observed outcome is slightly worse/better worse/better significantly worse/better than benchmark than benchmark than benchmark KPI strength Weak Strong Very strong Step-up/-down +/− 0 bps +/− 25 bps +/− 50 bps 3 Paying for performance, not outcomes The terms “performance-based financing”, “outcome-based financing” or “results-based financ- ing” are often used interchangeably. In the REACH framework, however, performance and outcome represent fundamentally different concepts. Outcome is the observed value (e.g. forest loss), which is the result of the government’s efforts combined with a set of external circumstances, such as weather conditions, existing policies or global exchange rates. The effort is, however, unobserved. But, as we will do in Section 4, it is possible to decompose the observed outcome and estimate the share of avoided deforestation that can be attributed to government efforts. This share is called performance. effort outcome ( , ) external factors In this section, we discuss why it is in both the issuer’s and investor’s interest to make this decomposition as clear as possible. From the controllability principle (see Section 3.4) we derive that the better the distinction, the stronger the incentives for the issuing government to perform, and the more attractive the SLB is to investors. The issuer’s perspective Let us consider the case where rewards are linked to observed outcomes instead of performance. Since the outcome does not only depend on the issuer’s performance, the issuer faces two risks. On the one hand, deforestation could slow down because global beef prices drop and logging forests to expand pastureland becomes less profitable. On the other hand, the El Niño-Southern Oscillation could lead to abnormally dry and hot weather, which could drastically accelerate deforestation due to wildfires. However, since neither weather nor global beef prices are under the issuer’s control, the issuer may be rewarded or penalized based on a KPI that is largely unrelated to how much effort they put in. Why, then, should the issuer put in any effort at all? The investor’s perspective A responsible investor is unlikely to be interested in investing in a performance-based product if its returns do not reflect the issuer’s performance. Not only is this against an impact investor’s goal of funding efforts to curb deforestation. Moreover, the investor’s return will depend on weather conditions, exchange rate fluctuations or commodity price swings. 8 Hence, the investor would be exposed to additional risk factors without adequate compensation, effectively raising the SLB’s costs. 3.1 Should the end justify the means? One could argue that investors may be less interested in how much effort the issuer exerted and more interested in the actual outcome on the ground. After all, what should matter is how many hectares of primary forests were protected and how many tons of emissions were avoided. However, we argue that this perspective has three important drawbacks. First, if we maintain the assumption that enacting laws, enforcing conservation policies and reducing the economy’s dependence on extracting natural resources are the most reliable ways to avoid deforestation, then incentivizing the government to act accordingly remains the best bet for the investor. While this does not guarantee a measurable improvement in absolute terms, it certainly raises its likelihood. Second, having a reliable measure of the issuer’s efforts and understanding what share of actual deforestation can be attributed to the government has implications for setting targets in the future. If the SLB targets are set using observed deforestation, but all the issuer can control is their own efforts, then it is likely that targets lie outside the range of outcomes to which the issuer can credibly commit. Finally, the purpose of a performance-based instrument such as an SLB is to promote actions and create conditions to facilitate efforts towards sustainable development goals. Designating protected areas or passing legislature to diversify the economy will have long-lasting effects, beyond the lifetime of the SLB, and affect the economy as a whole. Focusing only on observed deforestation may overly narrow the attention to outcomes in the short term and only to one particular sector. 3.2 Theoretical framework To formalize the lines of reasoning behind REACH, we anchor our analysis using insights from economic theory. In particular, let us consider a simple moral hazard framework in the spirit of Holmström (1979), Mirrlees (1999), and Shavell (1979) to model the principal-agent relationship behind an SLB, or any other performance-linked instrument. 7 The principal (the investor) lends capital to the agent (the issuer), whose actions influence the success probability of the underlying project. Both parties agree on the project target and on the state-contingent instrument. Different from a conventional debt contract, the interest rate of an SLB is not fixed. Instead, depending on the results, the interest rate can either be adjusted with a step-up, Δ > 0, in case the target is missed or a step-down, Δ < 0, if the target is reached. The agent can choose an unobservable level of effort that incurs private costs. The agent is risk-averse and has a utility that depends on the effort and contractual payoff, = ( , Δ) with / < 0 and ( , Δ ) < ( , Δ ). The principal is also risk-averse with utility function, = ( ), that satisfies (Δ ) > (Δ ). For a given set of external factors (also called the “state of nature”) which captures the business cycle, existing legislature, and other factors the agent cannot control, the agent’s effort induces a change in the observable outcome = ( , ). We assume that / > 0 for all states of nature , so exerting effort always improves the outcome. However, neither nor / can be 7 One could look at a green bond contract as a moral hazard problem as well. Namely, after the debt contract is issued, the performance of the issuer is incentivized through monitoring, reporting and verification. This could be modeled through the costly state verification approach (Townsend, 1979). 9 Figure 3: Performance measures with and without additional information The two figures depict the difference between an outcome-based KPI (left), which only uses the outcome to estimate the unobserved effort. The benchmark model-based KPI (right) also incorporates the external factors . Since the observed outcome is determined jointly by unobserved efforts and external factors, including leads to a more accurate estimate of effort. (a) Outcome-based KPI (b) Benchmark model-based KPI effort outcome outcome effort benchmarked outcome E[ | ] ( , ) E[ | , ] ( , ) external factors external factors directly observed. Instead, we must estimate a key performance indicator , which is a noisy signal of the agent’s effort . Thus, while effort itself cannot be observed, it is possible to obtain a statistical distribution that helps determine “whether [the agent’s] supply of inputs is able to affect the probability distribution of the output statistic” (Antle et al., 1988). 8 It is therefore important to accurately characterize the probability distribution around the , and how it depends the included factors, . 3.3 The value of information One could argue that the observed outcome alone is sufficiently informative about the agent’s efforts. This can be expressed as a conditional expectation, = E[ | ] + (1) where is mean zero and with variance 2 . In the special case where is treated as a perfect measure of effort, that is by assuming = = and 2 = 0, we obtain an outcome-based KPI, or simply an outcome KPI. However, alone would typically be an insufficient measure of , since the outcome is jointly determined by the agent’s efforts and the state of nature, = ( , ). Not using in calculating the would be wasting valuable information about the agent’s effort. As both issuer and investor are risk-averse, both prefer a more accurate and informative performance measure. It is therefore always desirable to include in the performance estimation, = E[ | , ] + . (2) Not conditioning on the additional information would increase the estimation uncertainty, i.e. 2 > 2 and give the principal a less informative signal about the agent’s performance, which in turn reduces the risk-averse agent’s incentive to exert effort. 3.4 Informativeness principle Our argumentation for why we advocate benchmarked KPIs over outcome KPIs is consistent with the informativeness principle. As outlined in Holmström’s (2016) Nobel Prize lecture on “Pay For 8 Concretely, Holmström (1979) describes the joint distribution ( , ), where the outcome is and the effort , as being induced on through the relationship = ( , ) where denotes the state of nature. Shavell (1979) introduces the notion of the principal’s observation of effort = ( , ). Hence, for a given , the outcome distributions of different effort levels will differ in the sense of first-order stochastic dominance. 10 Performance and Beyond”, the principle states that an “additional signal [] is valuable if and only if it carries additional information about what the agent did given the signal [ ]”. How the additional signal should enter the cannot be stated in general as it depends on the use case, time period, issuer and the type of effort the SLB intends to incentivize. Nevertheless, the implications of the informativeness principle provide guidance for possible approaches to define E[ | , ]. In particular, the controllability principle and the relative performance evaluation approaches provide statistical frameworks to estimate an unobserved counterfactual benchmark outcome, , which describes the hypothetical outcome if the issuer does not undertake any additional effort. Both approaches can be expressed as a difference between the observed outcome and the benchmark outcome . E[ | , ] = − (3) Inserting this equation into (2) yields = + + (4) Equation 4 shows the decomposition of the observed outcome into a part attributed to the agent’s performance, a part attributed to exogenous factors and a component that represents statistical uncertainty. We discuss how could be specified for both approaches in the following. 3.4.1 Controllability principle The application in Section 4 follows the controllability principle, which states that an agent’s in- centives should only depend on factors under their control. We implement this by estimating the counterfactual benchmark outcome which reflects how much deforestation would have occurred due to external factors alone. In this case, the benchmark model could be expressed as a conditional expectation = [ | ] where ⊆ is a set of exogenous predictors. Thus, equation (4) becomes a regression model of the form = + [ | ] + (5) 3.4.2 Relative performance evaluation Alternatively, one could follow the relative performance evaluation approach and obtain a benchmark outcome by using observed deforestation levels in peer countries. Synthetic control methods (Abadie, 2021) provide a suitable econometric framework. In this case, the benchmark model could be expressed as = [ | ] where denotes the observed outcomes of a set of relevant peers , e.g. similar geographic region or comparable income levels. Correspondingly, equation (4) becomes a regression model of the form = + [ | ] + (6) 3.5 Measuring impact through counterfactuals Setting the right incentives through better performance measurement is a primarily theoretical concern. In practice, the rigorous assessment and quantification of impact are far from trivial and require a sound statistical framework. One fundamental concept, known as the Neyman-Rubin causal model, is that of potential outcomes (Rubin, 1974). It describes the counterfactual situation 11 Figure 4: Annual deforestation in the Legal Amazon (km2 ) 30,000 25,000 20,000 15,000 10,000 5,000 0 1990 1995 2000 2005 2010 2015 2020 Source: Instituto Nacional de Pesquisas Espaciais (INPE) PRODES. that would have occurred in the absence of the evaluated intervention. The fundamental problem of causal inference is that "it is impossible to observe the [potential outcome] on the same unit and, therefore, it is impossible to observe the [treatment] effect" (Holland, 1986). However, researchers across different fields have developed empirical strategies to estimate these counterfactual out- comes, using randomized controlled trials, difference-in-differences, synthetic controls (Abadie, 2021) or matching. While the optimal estimation strategy in any given context depends on the assumptions the researcher or practitioner is willing to make, all methods share the fundamental insight that the causal impact of a program cannot be observed, but needs to be estimated. Relevant for our application of deforestation in Brazil’s Legal Amazon are Baylis et al. (2016), Ferraro (2009), and Ferraro et al. (2014), who review the available causal inference methods and their suitability for applications in environmental policy. 4 Deforestation in Brazil’s Legal Amazon We apply the REACH framework to deforestation and construct an econometric benchmark model that accounts for macroeconomic factors. We study deforestation in the nine Brazilian states which comprise the Legal Amazon: Acre, Amapá, Amazonas, Mato Grosso, Pará, Rondônia, Roraima, Tocantins, and (parts of) Maranhão. The rate of deforestation in Brazil’s Legal Amazon is alarming (Figure 4) and constitutes one of the main sources of Brazil’s GHG emissions as well as a major source of biodiversity loss (Hanusch, 2023). Most deforestation in the area is illegal, so resourcing and incentivizing governments to enforce existing laws to protect forests (“Command and Control”) is critical. Whether legal or illegal, deforestation is an economic choice (Hanusch, 2023). Macroeconomic factors can be key determinants of the costs and benefits of that choice and therefore, ultimately, of deforestation outcomes (de Souza Ferreiro Filho et al., 2022). On a local level, deforestation may have different immediate causes (for example cattle ranching, timber extraction, land grabbing, or mining), but on a regional or national level, the literature has found that macroeconomic factors that capture the demand for commodities (e.g. commodity prices) and the external competi- 12 Figure 5: Illustration of the benchmarked KPI for deforestation This figure depicts schematically how observed deforestation (grey) can be attributed to macroeconomic factors (blue), weather variables (orange), and outlier events (red). Most importantly, we also show how much of the forest was at risk of deforestation (green) but the loss never materialized thanks to policy efforts. Incentives should be aligned with the benchmarked KPI (green) rather than the observed outcome KPI (grey). This is only possible if we know the benchmark outcome (dark grey). Forest loss Attribution Macroeconomic factors Weather Outlier events Policy efforts Rising commodity prices, favorable exchange rates Extremely dry weather Strict pandemic The “forest at risk” that was and macroeconomic conditions that benefit the lowers agricultural lockdowns leads to not lost due to conservation profitability of the extractive sector over other yields and raises risks migration from urban efforts, law enforcement and sectors of the economy. of wild fires. to rural areas a diversified economy Observed outcome = Outcome KPI Benchmarked KPI Benchmark outcome Source: Authors. tiveness of domestic producers (e.g. inflation or the exchange rate) can partly explain the loss of primary forests (Arcand et al., 2008; Assunção et al., 2015; Curtis et al., 2018). While global prices and Brazil’s macroeconomic condition likely influence deforestation outcomes in Brazil’s Legal Amazon, the economy of the corresponding Brazilian states only accounts for 9 percent of national GDP in 2019 (de Souza Ferreiro Filho et al., 2022) and 13 percent of the population in 2021 (Hanusch, 2023), so deforestation patterns are unlikely to influence global or national outcomes. This makes deforestation in the Legal Amazon an ideal case to illustrate a benchmarked KPI. Macroeconomic factors matter — but policy matters, too. In many cases, the literature is divided over the effectiveness of different policies over vast territories like the Legal Amazon, as policies’ effectiveness in one area may displace deforestation to other areas. Such displacement effects have been shown, for example, for protected areas and voluntary private sector commitments such as the soy and beef moratoria for the Amazon (Carvalho et al., 2019). Other policies may be more effective at a larger scale, such as satellite-supported law enforcement or the blacklisting of municipalities in high deforestation areas (West et al., 2021). An ideal deforestation baseline would exclude any policy effects, as they are necessary for the performance that is supposed to be rewarded. 13 4.1 Data Observed deforestation for Brazil’s Legal Amazon We use deforestation data from PRODES 9 for the Legal Amazon between 1989 and 2021 for the annual model (Figure 4), and from DETER 10 between May 2015 and January 2021 for the monthly model (see Appendix A). This serves as the outcome variable, . We explore sub-annual and sub-national deforestation data using remote- sensed data with ex-post attribution of deforestation drivers in subsequent research. As discussed in Section 2.4, benchmarked KPIs are defined with respect to the benchmark outcome, which is an estimate from the benchmark model. The main ingredient of the approach in this application is a set of exogenous factors . From the literature, we can identify a set of possible factors that are both exogenous to the Brazilian economy as well as predictive of the deforestation dynamics. However, it is harder to know ex-ante exactly which factors should be included and in what form. We describe the candidate variables below. Real effective exchange rate (REER) As the exchange rate determines a country’s competi- tiveness on the world stage, a weakening exchange rate will accelerate deforestation for wood production and the export of timber products, and/or advance the forest to land conversion for agricultural exports (Arcand et al., 2008; Hanusch, 2023; Richards, 2021). Global commodity prices. Globally, Curtis et al. (2018) estimate that "27±5% of all forest dis- turbance between 2001 and 2015 was associated with commodity-driven deforestation". For our application, we use global market prices of beef, coffee, soy, corn, sugar, soy oil, hardwood logs, and iron ore in Brazilian real. In line with the previous point, higher commodity prices will drive Brazilian deforestation. However, the reverse does not hold, i.e. it is very unlikely that Brazilian forest loss will drive global commodity prices, esp. in the short-run (Arcand et al., 2008; Hanusch, 2023; Richards, 2021). Since commodity prices are highly correlated, we compute their principal components instead of using the variables directly. Previous deforestation rate. To reflect deforestation trends that precede the policy actions de- fined as performance we include lagged deforestation. This variable could be argued to be not fully exogenous. For example, policy decisions about forest conservation affect not only defor- estation in a given year but also in subsequent years. These variables are weakly exogenous or predetermined, meaning that previous deforestation trends determine current deforestation, but not vice versa, especially in the short run. 4.2 Econometric benchmark model Equation (7) shows how the benchmark model under the controllability principle from Section 3.4.1, equation (5), is operationalized in this application. The definition holds for the annual model 9 The PRODES project carries out satellite monitoring of clear-cut deforestation in the Legal Amazon and produces, since 1988, annual deforestation rates in the region, which are used by the Brazilian government to establish public policies. For more information, please see the official website and Valeriano et al. (2004). 10 Developed by the National Institute for Space Research (INPE) as one of the key changes from the Action Plan for the Prevention and Control of Deforestation in the Legal Amazon, DETER is a satellite-based system that captures and processes georeferenced imagery on forest cover in 15-day intervals. See the official website and Assunção et al. (2023). 14 using PRODES data as well as the monthly version which uses DETER data. 11 = + − + − + [ ≥ ] + (7) =1 =1 =1 ∈ Here, is the intercept, is the autocorrelation coefficient of lag , − are the predictors described in the previous section, lagged at least one and up to three years. [ ≥ ] are level- shifting policy dummies that account for known past policy interventions and which remain active for all years after the intervention. The years for which policy dummies are permissible are selected based on knowledge of past policy actions – in this application we choose 2004, 2009 and 2012 (see Figure 6). These dummies are necessary to model the generalizable relationship between exogenous deforestation drivers and benchmark predictors, untainted by known policy effects. This leaves us with a large set of possible predictors. Using a variable selection method (LASSO) we select the most relevant predictors for deforestation (see Figure 12). Policy dummies are not subject to penalization. The model is estimated for annual and monthly deforestation rates. We for variable selection (Tibshirani, 1996), where impose a 1 penalty on the coefficients and the penalty strength is determined through cross-validation. The selection results show that four variables matter the most: previous deforestation, lagged REER percentage change, and lagged commodity prices (see Figure 12). The second and fourth principal components mostly capture the prices of beef, iron ore, sugar 12 and hardwood logs (see Table 3). After selecting the variables, we refit a linear regression model with only the most important predictors (see Table 4) as an illustration. Jointly, they explain about 90% of the variation in the annual model and around 80% for the monthly model. Note that we also include level-shifting policy dummies for the years 2004, 2009 and 2012 for the annual model and seasonal dummies for the monthly model (see Table 5). 4.3 Benchmarking avoided deforestation performance Using the benchmark model (Equation 7) we can ex-post calculate the benchmark level of defor- estation in the Legal Amazon. Figure 6 shows the results. The bars show observed deforestation levels , similarly to the bars in Figure 4. However, they have now been colored to reflect whether was better or worse than the expected benchmark , shown as blue horizontal dashed lines. Since the model also produces a benchmark distribution, we can calculate the confidence inter- vals for each prediction. Interestingly, for the four years where new policies were introduced, we can observe lower than expected deforestation levels: Amazon Region Protected Areas Program (2000), Plan for Prevention and Control of Deforestation in the Legal Amazon (PPCDAm, 2004) and Soy Moratorium (2005), the blacklisting of municipalities and beef moratorium (2009), and the update of the Forest Code (2012). For the blacklisting moratorium and Forest Code, measured deforestation is outside the 95% confidence interval, such that it would constitute a strong KPI (see Table 1). Since 2018, and especially 2019, deforestation has picked up to a rate that exceeds what can be expected given the macroeconomic conditions. However, the increase in the benchmark outcome suggests that not the entire increase in observed deforestation can be attributed to (a lack of) policy 11 The time periods and frequencies of relevant variables are then changed from annual to monthly, accordingly. See Appendix A for monthly results. 12 Though sugar is not grown in the Legal Amazon, its cultivation in other parts of Brazil can lead to displacement effects and affect land choices across the country. 15 Figure 6: Better or worse than expected? This figure shows the expected outcomes (center of distributions, blue dashed lines) and the observed deforestation levels (colored bars), which are red if deforestation is worse and green if better than expected. The text boxes demarcate the years when a policy was introduced to combat deforestation. For monthly results, see Figure 13, Appendix. 35,000 Amazon Region Protected Areas Program (2002) 30,000 Plan for Prevention and Control of Deforestation in the Legal Amazon (2004) Annual Amazonian deforestation (km2) 25,000 Amazon Soy moratorium (2006) 20,000 Blacklisting of municipalities (2008) Cattle moratorium (2009) 15,000 Updated forest code (2012) 10,000 5,000 0 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 Benchmark deforestation Observed deforestation benchmark outcome slightly below benchmark: weak KPI slightly above benchmark: weak KPI 50% confidence interval below benchmark: strong KPI above benchmark: strong KPI 95% confidence interval significantly below benchmark: very strong KPI significantly above benchmark: very strong KPI Source: Instituto Nacional de Pesquisas Espaciais (INPE) PRODES, calculation by authors. action under the federal government. This finding is in line with (de Smit, 2022) who also finds a mostly positive but not significant effect using synthetic control methods. 4.4 Incentives under constant benchmarks The main argument behind the above analysis is that estimating benchmark outcomes requires robust and transparent empirical methods instead of simple extrapolation of historical trends. The various implementations of forest reference levels/forest reference emissions levels (FRLs/FRELs) are an example for the latter. 13 FRLs/FRELs form the baseline for results-based financing that penalizes emissions from deforestation and forest degradation, rewards forest carbon stock con- servation and enhancement, and sustainable forest management (UNFCCC, 2011, paragraph 70). Payments are contingent upon the observed difference between FRLs/FRELs and the observed deforestation (Granziera et al., 2021). This is consistent with our suggestion not to use observed outcomes as KPIs, but instead compare the outcomes with a benchmark. In fact, the guidelines in UNFCCC (2014) assess how historical data were taken into account when establishing the FRLs/FRELs. However, some reference levels discussed in Granziera et al. (2021) are simple historical averages over 4-year to 20-year periods, reassessed periodically. Various reference levels were used for results-based financing (see Figure 7), which correspond to historical averages over an expanding window since 1996. These remain constant for 5-year periods, after which the FRELs are updated. 13 FRLs/FRELs are submitted by developing country parties, assessed by LULUCF experts and published on the Lima REDD+ Information Hub. 16 Figure 7: Brazil’s forest reference emissions levels (FRELs) This figure depicts Brazil’s submitted FRELs for the Amazonia biome (Ministry of the Environment of Brazil, 2018; Ministry of the Environment of Brazil et al., 2014), and reflects how the benchmarks are computed as historical averages over an expanding window since 1996, which remain constant within 5-year periods. FREL A, reference period: 1996 2005 1,400 FREL B, reference period: 1996 2010 FREL C, reference period: 1996 2015 1,200 FREL A, reference level: 2006 2010 FREL A FREL B, reference level: 2011 2015 FREL C, reference level: 2016 2020 1,000 FREL B 800 FREL C 600 400 200 0 1996 2000 2004 2008 2012 2016 2020 Source: Ministry of the Environment; Ministry of Science, Technology and Innovation of Brazil. Within the 5-year periods, the results were evaluated against the prevalent benchmark. This approach to estimating a baseline scenario has been criticized in the past as too simplistic (Fletcher et al., 2016; Huettner et al., 2009; Mertz et al., 2018), see Arts et al. (2019) for an overview. From a principal-agent perspective, a constant benchmark that does not satisfy the controllability principle (see Section 3.4.1) will weaken the performance incentives for the issuer. If the benchmark does not reflect external factors, then the issuer’s rewards or penalties will be decoupled from its performance. Moreover, funding providers may end up rewarding results which may have taken place regardless of the issuer’s efforts. For instance, the spike in deforestation in Brazil’s Legal Amazon until 2004 implies a high baseline deforestation rate for all FRELs used in projects that started in subsequent years, as shown in Figure 7. As a result, West et al. (2020), West et al. (2023) argue that a significant part of the carbon credits claimed through such projects do not reflect additional emissions reductions. This highlights the importance of appropriate and robust benchmark choices, also for the credibility of carbon offsets. The authors also show that during the implementation phase of such projects, deforestation decreased in most areas of the Brazilian Amazon (in part) due to macroeconomic conditions. 4.5 Discussion If we compare the FREL benchmarks described Figure 7 with the benchmarked KPI in Figure 6, we can make several observations with respect to the REACH framework. Both approaches satisfy the requirement of a “transparent benchmark model”. In the case of FRELs, the benchmark is a historical average over a reference period, updated every five years. For our proposed benchmarked KPIs, the econometric benchmark model is described in Section 4.2. However, different from the benchmarked KPI approach, FRELs do not provide any probability distributions 17 for their benchmarks. This makes it difficult to assess if the outcomes were due to chance or actual efforts. Most importantly, the FRELs do not provide a clear justification for additionality. From Brazil’s submissions under the UNFCCC 14 we can gather a potential justification, since the submissions explicitly state that the FRELs “[do] not include assumptions on potential future changes to domestic policies.” This is different for the benchmarked KPIs, which satisfies the informativeness principle. The benefit of the latter is that it disentangles which part of deforestation was under the issuer’s control and which part was not, which in turn satisfies the controllability principle. An alternative approach that satisfies the relative performance evaluation is the work of Guizar- Coutiño et al. (2022) and West et al. (2020) who used statistical methods based on matching to estimate the additionality of conservation projects on a sub-national project level. While Guizar- Coutiño et al. (2022) find that deforestation in 40 conservation projects areas in nine countries was reduced by 47%, on average, compared to matched control pixels, West et al. (2020) find little evidence that such projects in the Brazilian Amazon “have mitigated forest loss” compared to matched control regions using synthetic control methods. This method was also explored by de Smit (2022) on a country level, though the author did not find any statistically significant results. 5 Setting ambitious, yet feasible targets A promising SPT strikes a good balance between ambitiousness and feasibility. 15 So far, we have described how the framework assesses performance and sets incentives through benchmarked KPIs. In this section, we discuss another essential aspect of the framework: setting ambitious SPTs that remain feasible to achieve. The ICMA guidelines define ambitious targets as those that “represent a material improvement [...] beyond a ‘Business as Usual’ trajectory” (ICMA, 2020). At the same time, the guidelines emphasize the need for feasible targets based on “benchmarking approaches” such as the past performance of the issuer, their peers, or science-based scenarios. The Feasibility-Ambitiousness (FAB) matrix summarizes how we assess sustainability perfor- mance targets. Figure 8 shows the interplay between feasibility and ambitiousness. The ideal targets are those that are highly ambitious and become increasingly feasible due to SLB fund- ing. Such targets imply additionality and thereby send the strongest signals for both issuers and investors. 5.1 Low-hanging fruits and long shots Low-hanging fruits Issuers of SLBs with unambitious targets are vulnerable to greenwashing allegations. If little to no effort is necessary to reach the SPT, then such targets would have likely been realized without the SLB financing. This is especially true for SPTs that the issuer was already making good progress towards. Additionally, in the event of reaching an SPT, investors will face lower financial returns due to the triggered step-down. Responsible or impact investors may find such an SLB unattractive since they would have to justify the financial and reputational consequences if the social or environmental impact achieved was only a low-hanging fruit. Long shots If achieving the SPT would require an unrealistic improvement, projects may be regarded as unattainable and overzealous. Issuers would likely have to pay a penalty fee since 14 See Ministry of the Environment of Brazil (2018) and Ministry of the Environment of Brazil et al. (2014). 15 Visit https://esgdata.worldbank.org/tools/fab to analyze indicators and countries on the Sovereign ESG Data Portal. 18 Figure 8: The Feasibility–Ambitiousness (FAB) matrix The FAB shows that setting promising targets requires the consideration of both ambitiousness and feasi- bility. Neither low-hanging fruits nor long shots are desirable as they may lead to greenwashing allegations (see Section 5.1) The ideal targets are both feasible and ambitious but require additional SLB funding to become realizable. forward-looking extrapolates business-as-usual Ambitiousness trends with no policy changes Feasibility Low High Very high Highly ambitious backward-looking but low chance based on historical Low Long Shots of success precedents of Target unlikely to relevant peers be achieved given High lack of historical precedence Reachable Low-Hanging Targets Very high Fruits Highly feasible but unambitious Target requires little to no effort and Source: Wang, Gurhy, may have been reached in any case Hanusch, Kollenda (2023) Source: Authors. the SLB structure stipulates a step-up in the case of missing the SPT. Such overambitious SPTs may be therefore only look good on paper and never translate into reality, leading to possible greenwashing allegations. Investors that buy such an SLB face two types of reputational risks. The first type stems from funding unrealistic projects whose targets may never be reached. This is compounded by the second risk, namely that they will need to justify benefiting from a step- up when the issuer fails to meet an unrealistic SPT – especially, if it was known that the SPT’s feasibility was low. We illustrate how SPTs for reduced deforestation in Brazil’s Legal Amazon could be set using the FAB matrix. The methodology to evaluate feasibility relies on a peer country comparison with other tropical forest countries. For these countries, we use country-level primary forest loss data from the Global Forest Watch rather than INPE’s data, which is only available for Brazil’s Legal Amazon. We set the SPT horizon to four years for this application. From a collection of all tropical forest countries, we selected 19 comparable peer countries with non-negligible deforestation rates in the past. 16 Because of the disproportional size of Brazil’s forest area, our deforestation indicator is standardized to represent annual forest loss relative to the forest area in the year 2000. Figure 9 shows the selected countries and annual relative forest losses since 2000. In the two decades until 2021, Brazil’s Legal Amazon has lost 7 percent of its primary forest area – a loss of 261,567 square kilometers, almost the total land area of Ecuador. 5.2 Ambitiousness Ambitiousness is a forward-looking concept based on projecting a business-as-usual (BAU) scenario. Given the discussion on macroeconomic drivers of deforestation in Section 4 we construct a BAU scenario by applying equation (7) to predict deforestation with forecasted macroeconomic 16 Angola, Argentina, Bolivia, Brazil, Cameroon, Republic of Congo, Colombia, Ecuador, Gabon, Guyana, Indonesia, India, Mexico, Peru, Suriname, Tanzania, Uganda, República Bolivariano de Venezuela, and Zambia. 19 Figure 9: Annual forest loss (relative to 2000) Annual forest loss as a share of the forest area in 2000 is at the national level for peer countries and only for the Legal Amazon for Brazil. The peer countries are other tropical forest countries. Source: Global Forest Watch, PRODES, calculation by authors. conditions from 2022-2025 from the World Bank’s Macro-Fiscal Model (Burns et al., 2019). 17 Figure 10 shows that given forecasted macroeconomic conditions, deforestation in Brazil’s Legal Amazon is expected to reduce significantly below the rate of 2021, even in the absence of any policy changes. However, the BAU values of roughly 0.24% (∼8,500 square kilometers, see Table 2) from 2023-2025 still reflect a higher rate of deforestation than the period between 2010-2015 where macroeconomic conditions and policy efforts led to relatively low deforestation rates. The forecast intervals reflect the uncertainty around the annual forest loss trajectory. We can use the forecast intervals to differentiate between targets of low and high ambition. Outcomes within the 50% forecast interval have a high likelihood of happening even under a BAU scenario and are therefore classified as low ambitiousness. In contrast, targets that fall below the 90% forecast interval are unlikely to be realized in a BAU scenario and are therefore classified as highly ambitious. 18 This is in line with ICMA’s criteria of setting targets that “represent a material improvement [...] beyond a ‘Business as Usual’ trajectory” (ICMA, 2020). Pathways significantly below the BAU trajectory contain promising targets. For example, the lower bound of the 90% interval of forecasted forest loss could serve as guidance for an ambitious SPT. In our sustainability-linked performance framework, KPIs and SPTs are always defined in relation to a benchmarked outcome. During the target setting, the average forecast from the ambitiousness analysis can then be used as a benchmark to define a relative SPT. In Section 5.4 and Table 2 we give a numerical example of how targets could be set in this application. 17 Alternative modeling choices for BAU scenarios could be based on computable general equilibrium models, time series econometrics or dynamic factor models. The best choice will depend on the indicator of interest. 18 Note that, because lower annual forest loss is better, the differentiation is not necessarily symmetric around the median forecast path. Instead, we do not classify targets that would result in forest loss rates significantly above the BAU forecast as those would reflect a material worsening and hence not an applicable SPT. 20 Figure 10: Ambitiousness analysis for deforestation reduction targets The forecasts are formed by fitting the benchmarking model based on macroeconomic conditions (Equation 7 and using forecasted macroeconomic conditions to predict annual deforestation rates in Brazil’s Legal Amazon. The average forecast (dark green line) represents the BAU scenario. The forecast intervals show the likelihood of alternative outcomes. Annual forest loss is expressed relative to the year 2000 area and standardized to be 0 in 2021. Source: PRODES, calculation by authors. 5.3 Feasibility Feasibility is a backwards-looking concept which looks for historical precedents to identify a range of realistic improvements. This approach connects to ICMA’s criteria of setting targets using benchmarking, specifically with respect to issuers’ peers and over time or in reference to science- based scenarios. Here, the key idea is that we regard a target as feasible if other countries have been able to show a similar improvement over a comparable time horizon in the past. In the application to deforestation in Brazil’s Legal Amazon, the ideal peers are tropical forest countries with a similar rate of deforestation as Brazil had in 2021, which shared a similar set of circumstances in terms of climate, income and region. By tracing the forest loss that these peer countries experienced over the following four years we can quantify the range of possible outcomes based on historical precedents. A feasible SPT should then not be too far away from what peer countries have achieved in the past. Delineating the feasibility dimension has two steps: identify relevant peers and select those that had similar historical value(s). First, we select peer countries that are comparable to the issuer, for example in terms of geography, income or climate. In this application, we use the 19 selected tropical forest countries (see footnote 16) as possible candidates. More refined methods can consider other factors, such as geographic region or income level. Second, we select historical periods where peer countries shared a similar indicator level as the issuer in 2021. This process generates several historical paths that can serve as precedents for the issuer. 19 Each precedent 19 A country may contribute several precedents if it frequently had an indicator value close to the most recent value for the issuer. To reduce double-counting, our preferred specification selects the latest available path per country. Alternatively, one can choose the path which represents the biggest improvement or aggregate all paths per country using the average or the median. 21 Figure 11: Feasibility analysis for deforestation reduction targets The grey lines are deforestation trajectories from selected tropical forest countries with similar forest loss levels as Brazil (Legal Amazon) in 2021. These historical precedents depict the range of what has been achieved within four years in the past. The blue line shows historical forest loss for Brazil’s Legal Amazon between 2015 and 2021, standardized to 0 in 2021. Annual forest loss is expressed relative to the forest area in peers’ base year and as the change from Brazil’s Legal Amazon’s most recent value: 0.37%. Source: PRODES, calculation by authors. draws a path and shows what other countries were or were not able to accomplish within four years. From historical precedents, we can deduce feasibility intervals, which can be used to differentiate between SPTs with low or high feasibility. Figure 11 shows the feasibility precedents and the median development path for annual forest loss. The median development path summarizes all past historical developments in peer countries with deforestation rates similar to Brazil’s in 2021. It shows that in the following four years, the median peer countries’ annual deforestation had risen to 0.43%. SPTs within the 50% feasibility interval have high feasibility, whereas SPTs below the 90% feasibility interval are considered low feasibility, with scant historical evidence. 20 This approach can be further refined following the same rationale as in Section 2 by accounting for factors that are outside the control of individual peer countries. For example, the results shown in Figure 11 do not account for individual country effects, such as through policy dummies (see Section 4.2). Moreover, one could also control for macroeconomic determinants, such as we have done in Section 4. We leave these extensions for future work. Alternative approaches to feasibility While the present analysis relies on historical precedents from peer countries to characterize what is feasible, it does not preclude other ways to assess feasibility. The peer country approach has the advantage of being general enough to be used systematically for different types of indicators. However, historical precedents may not always be the best benchmark. New technologies, which were not available in the past, may expand 20 Again, targets above the 50% feasibility interval are considered not applicable, as the goal is to identify realistic reductions in forest loss compared to the 2021 level. 22 Table 2: FAB analysis for Brazil’s Legal Amazon until 2025 This table shows an illustration of the FAB matrix analysis. It shows the values from the ambitiousness and feasibility categories and how the two analyses can inform a target setting exercise. Values are expressed as annual forest loss relative to the forest area in the year 2000 and converted into square kilometers in parentheses. For feasibility, relative forest loss is transformed to square kilometers using the forest area of Brazil’s Legal Amazon. 2021 2022 2023 2024 2025 0.37% Most recent value (13,235 km2 ) Ambitiousness 0.31% 0.25% 0.24% 0.24% Average BAU scenario forecast (11,178 km2 ) (8,738 km2 ) (8,469 km2 ) (8,470 km2 ) 0.30% 0.25% 0.24% 0.24% Cutoff: low ambitiousness / ambitious (10,640 km2 ) (8,749 km2 ) (8,461 km2 ) (8,426 km2 ) 0.28% 0.23% 0.23% 0.23% Cutoff: ambitious / highly ambitious (9,802 km2 ) (8,122 km2 ) (8,016 km2 ) (8,083 km2 ) Feasibility 0.40% 0.37% 0.46% 0.43% Median development path (14,235 km2 ) (13,108 km2 ) (16,464 km2 ) (15,169 km2 ) 0.30% 0.30% 0.31% 0.30% Cutoff: feasible / highly feasible (10,689 km2 ) (10,636 km2 ) (11,260 km2 ) (10,769 km2 ) 0.21% 0.21% 0.26% 0.21% Cutoff: low feasibility / feasible (7,549 km2 ) (7,392 km2 ) (9,361 km2 ) (7,321 km2 ) the range of realistic targets. Formal feasibility studies or rigorous impact evaluations of policy options can take the local context into account to an extent that a general approach cannot. Such alternative approaches may, in some situations, be more credible than a general approach. Nevertheless, in order to adhere to the REACH principles, a study of the feasibility of different outcome ranges should be transparent and use evidence-based methodologies. For example, Assunção et al. (2023) describe how recent advances in satellite technology have contributed to monitoring and environmental enforcement in the Amazon. They find that “increasing monitoring and law enforcement by half decreases municipal deforestation by an estimated 25 percent”. The authors used cloud coverage as an instrument for law enforcement and were thereby able to construct a model to estimate counterfactuals for different scenarios. This framework could serve as a benchmark for feasible improvements for other tropical forest countries that consider deforestation-related SPTs paired with similar policy efforts. 5.4 Discussion We end the exercise in target setting by combining our insights along the ambitiousness and feasibility dimensions into the feasibility-ambitiousness matrix and identifying a promising target. Table 2 summarizes the relevant values from the ambitiousness (Figure 10) and feasibility (Figure 11) analyses. From the ambitiousness analysis, we considered the lower bound of the 90% forecast interval as a potential ambitious target. This translates to a target of an annual forest loss between 0.28% (9,802 km2 ) in 2022 and 0.23% (8,083 km2 ) in 2025. From the feasibility analysis, we can see that such a trajectory would lie within the range of feasible outcomes, which taking 2025 as an example ranges from 0.21% (7,321 km2 ) to 0.30% (10,769 km2 ). Hence, the trajectory laid out by the lower bound of the 90% ambitiousness forecast interval is 23 a reachable (ambitious and feasible) target (see Figure 8). So far, we have expressed the target in absolute annual values and not in relation to a baseline outcome. To set a relative target, one would have to define an appropriate ex-ante benchmark and then take the difference between the benchmark trajectory and the absolute target path to calculate the target in relation to the benchmark. In our example, an appropriate benchmark could be the average BAU forecast (see Table 2. 6 Conclusion In this paper we provided a new way of looking at sustainability-linked financing instruments and investigated how SLBs could be an effective tool to curb deforestation in Brazil’s Legal Amazon. We introduce the REACH framework – relative evaluation and benchmarking – which makes the crucial distinction between outcome measurement and performance measurement through the use of benchmarking methods. The key insight is that it is in both the issuer’s and the investor’s interests to reward the issuer’s performance and not the observed outcomes. This follows directly from principal-agent theory and the controllability principle. To give the strongest incentives for the issuer to perform, the carrots and sticks of an SLB should only be linked to factors the issuer can control, e.g. passing legislation to curb deforestation, supporting the forest police or diversifying the economy. Conversely, external factors which the issuer cannot control but determine the ultimate outcome, e.g. weather conditions or global commodity prices, should be stripped away from the performance metric. The use of benchmarks is in line with the guidelines proposed by ICMA, which suggest that KPIs should be measured with respect to a benchmark, which can be in the form of “an external reference or definitions” (ICMA, 2020). REACH imposes stricter requirements on benchmarks, since they determine what qualifies as impact or additionality. In other words, benchmarks are necessary to quantify the causal effect beyond any statistical doubt. We discuss why this cannot be done without knowing counterfactual outcomes, i.e. how much deforestation would have happened if the issuer did not perform any better or worse than expected. The REACH framework lists three requirements for the benchmarks: (R1) transparent model specification, (R2) benchmark probability distribution, and (R3) justification for additionality. Using the REACH framework we turn our attention to Brazil’s Legal Amazon. We specify a bench- mark model that satisfies the requirements (R1)-(R3) and estimate this historical performance of a hypothetical SLB. The analysis shows the difference between outcome KPIs, which equate per- formance with outcomes (“20,000 km2 of observed forest loss”), and benchmarked KPIs, which calculate performance relative to the effect of external factors (“2,000 km2 less deforestation than would be expected given the prevalent macroeconomic conditions”). We also discuss why bench- marks which do not fulfil the REACH requirements, such as forest reference emissions levels, dilute the incentives that benchmarked KPIs preserve. To complement REACH, which is primarily about performance measurement, we introduce the Feasibility-Ambitiousness (FAB) matrix which is about target setting. SLBs specify which sustain- ability goals they aim to achieve and within which time frame. However, it is difficult to assess whether a target is ambitious. Issuers have used second-party opinions to answer this question. We argue that one must also consider the feasibility of the targets to avoid greenwashing accusa- tions. According to the FAB, one should avoid low hanging fruits – which are highly feasible but not ambitious – and long shots – which are highly ambitious but not feasible. The FAB delineates statistical methods to assess both dimensions and we give an example of an SPT that would be 24 feasible and ambitious in Brazil’s Legal Amazon. While we introduced the frameworks in the context of SLBs, the REACH and FAB can be applied more widely. For instance, it can be used to assess the impact of labeled bonds or to track the additionality of nature-based solutions or debt-for-nature swaps. SLBs just happen to be ideally suited to connect the evaluation with performance incentives. For issuers and investors, REACH and FAB can form the basis of discussions and roadshows, while also facilitating the subsequent structuring and pricing process. 6.1 Practical considerations State-contingent debt instruments The structure of SLBs is reminiscent of state-contingent debt instruments (SCDIs), which share the fundamental idea to link debt service with a state variable. 21 Traditionally, SCDIs were attractive to investors due to their counter-cyclical, diversification and hedging properties. They were also appealing to issuers since debt repayments are linked to the fiscal capacity and because SCDIs could broaden the country’s investor base (IMF, 2017). However, SCDIs enjoyed limited uptake due to liquidity and pricing concerns, and measurement issues such as delayed reporting. In the current climate of sovereign debt distress and associated debt restructuring negotiations, linking debt repayments to a state variable could be worth exploring, especially for sovereigns exposed to climate change risks and environmental degradation (Volz, 2023). The benchmarked KPIs developed under the REACH framework could inform debt (re-)structuring decisions. Sim- ilar to catastrophe bonds or other disaster risk financing instruments, SLBs could thereby provide downside protection in the case of natural disasters, since the benchmarked KPIs account for exter- nal factors. At the same time, SLBs could expand governments fiscal space and aid in their mission to reach their Sustainable Development Goals (SDGs) or Nationally Determined Contributions (NDCs). Geospatial deforestation data It is worth emphasizing that the viability of both evidence-based frameworks hinges on the data environment. In the application, annual data was sufficient to demonstrate the framework. However, for model validation and reliability, more granular data would be needed. For deforestation, remote-sensed satellite data is a highly promising way to provide objective, regular and reliable observations. Subannual and subnational geospatial data on land use and land cover transitions can help us understand the drivers, locations and trends in deforestation and degradation, which could in turn produce more sophisticated and effective KPIs. An SLB that targets deforestation due to cropland expansion may necessitate a different set of government entities than an SLB that targets deforestation due to urban expansion. The World Bank and the European Space Agency are currently exploring how SLBs may leverage new deforestation monitoring technologies with cloud-penetrating capabilities. Private capital mobilization Multilateral development banks (MDBs) have a clear and essential role to play to mobilize private capital. The REACH framework aims at setting better performance incentives by accounting for what the issuer can be reasonably expected to achieve. The size of the step-ups or step-downs could then be scaled beyond the commonly used 25 basis points. However, while this sends strong signals about the commitment of both issuer and investor, the interest rate differentials may be difficult to stomach, especially when the market for sovereign 21 The SLBs discussed in the REACH framework could be seen as floater-type SCDIs, which have a fixed principal and a variable coupon. For the SLB, the coupon would be linked to the benchmarked KPI, the state variable. 25 SLBs is still nascent. MDBs can help derisk the transaction on either side through concessional finance or guarantees. 22 Moreover, SPTs are often aligned with existing development operations and MDBs can provide technical assistance, capacity building and knowledge transfer to support issuers in reaching their goals. A strong coupling of incentives to performance, as in the REACH framework, only strengthens the case for accompanying non-financial support as these efforts are now more directly liked to financial rewards. Use-of-proceeds with exclusions We have argued that, in theory, one of the SLB’s strengths is that a well-chosen KPI obviates the need for identifying a basket of eligible projects and tag the expenses as is common for GSSBs. In practice, however, an entirely unrestricted use of proceeds may lead to greenwashing concerns as well. For instance, when proceeds are used to fund activities that pollute the environment or violate human rights. In order to conserve room for sovereigns to navigate, on the one hand, and preventing reputational concerns due to the use of proceeds, on the other hand, it may be prudent to stipulate which activities SLB funds cannot be used for. The UN Global Compact principles could serve as a reference for exclusionary criteria. Pricing of SLBs Since SLBs are still new, pricing is a particular hurdle. For instance, it would be difficult to find comparable instruments to serve as references. Conventional fixed income pricing approaches may not be appropriate given the SLB’s state-contingent repayment mechanisms. Instead, SLBs constructed using REACH can leverage the (R2) requirement of always providing a benchmark probability distribution. This opens up the toolbox of insurance or derivative pricing. Concretely, benchmarked KPIs are the underlying, SPTs define the strike price, and the benchmark distributions provide the necessary moments to compute variance, skewness and kurtosis. Closely related is also the literature on real option pricing, which could inform the price of forest carbon offsets. We leave the pricing exercise to future research. 6.2 Outlook REACH and FAB together pave the way for the standardization of SLB issuances in the future. While the selection of KPIs, definition of benchmark models and target-setting procedures will need to be tailored for each country and instrument, the structured approach of both REACH and FAB can be generalized. Moreover, since statistical models underpin the frameworks, discussions about whether or not a particular KPI constitutes actual performance can be easily supported or disproved with empirical evidence. Making the data and code publicly accessible will fur- ther reduce information asymmetries, foster uptake, support coordination between ministries, encourage critical discussion, and ultimately lead to better SLBs. 22 In the context of contract theory, MDBs can help relax the participation constraints for issuers. 26 A Appendix Figure 12: Selecting the most important predictors with LASSO (annual model) Using the least absolute shrinkage and selection operator (LASSO) we identify the most predictive variables, i.e. the variables in colored boxes. Variables that enter the active set first (non-zero coefficient value), as the penalty decreases from left to right, reflect higher importance. The red dots demarcate the mean-squared error (MSE) associated with predictions made using variables in the active set. The black diamond marks the lowest MSE. The relative changes of the real effective exchange rate (REER) is always selected across different specifications. Mean-squared error (MSE) Minimum Coefficient values 0 Highest penalty Lowest penalty (1) previous | Deforestation | t-1 (11) macros | gdppc.growth | t-1 (21) commodities | pca.1 | t-2 (2) macros | reer | t-1 (12) previous | Deforestation | t-2 (22) macros | reer | t-3 (3) commodities | pca.5 | t-1 (13) commodities | pca.4 | t-1 (23) macros | gdppc.growth | t-2 (4) commodities | pca.3 | t-1 (14) macros | reer | t-2 (24) macros | cpi-inflation | t-1 (5) commodities | pca.5 | t-2 (15) commodities | pca.4 | t-3 (6) commodities | pca.5 | t-3 (16) macros | cpi-inflation | t-3 (7) commodities | pca.1 | t-3 (17) commodities | pca.2 | t-1 (8) commodities | pca.3 | t-2 previous | Deforestation | t-3 (18) (9) macros | gdppc.growth | t-3 (19) commodities | pca.1 | t-1 (10) commodities | pca.2 | t-2 (20) commodities | pca.2 | t-3 Source: Calculation by authors. Table 3: Principal components of commodity prices This table shows the first four principal components of global commodity prices in Brazilian reals between 1996–2020 (annual frequency). The data was obtained from FRED, Federal Reserve Bank of St. Louis. Principal component 1 2 3 4 Beef 0.3522 0.4019 -0.0203 0.5664 Corn 0.3641 -0.0979 -0.2659 -0.0701 Hard Logs 0.3586 0.3387 -0.1922 -0.0990 Coffee (Arabica) 0.3526 0.1899 0.4833 0.3263 Soybeans 0.3606 -0.0139 -0.4148 -0.1546 Sugar 0.3450 0.0784 0.6011 -0.6206 Soybeans Oil 0.3619 -0.1311 -0.3096 -0.2356 Iron Ore 0.3323 -0.8091 0.1706 0.3057 Explained variance 90.06% 93.46% 96.13% 97.97% 27 Table 4: Post-LASSO regression results for the annual model The results were obtained from fitting an linear regression model on the selected variables from the LASSO model. Hence, the standard errors are invalid and only serve an indicative purpose due to issues surrounding post selection inference. Note that a variable being selected by the LASSO does not imply that the variable is also statistically significant. Group Predictor Lag Coef. S.E. t-stat p-val. Deforestation − 1 0.605 0.150 4.026 0.001 *** Macroeconomic Δ (REER) − 1 -0.237 0.141 -1.677 0.114 Commodities Principal comp. 2 − 1 0.191 0.090 2.127 0.050 * Principal comp. 4 − 1 0.170 0.090 1.883 0.079 * Dummies 2004 onwards . -0.028 0.053 -0.523 0.608 2009 onwards . -0.093 0.070 -1.333 0.202 2012 onwards . -0.004 0.060 -0.073 0.943 Constant . 1.659 0.638 2.600 0.020 ** R-squared . 93.7% . . . . R-squared adj. . 90.8% . . . . F-stat. . 31.997 . . . . Prob(F-stat.) . 0.000 . . . . Observations . 23 . . . . Table 5: Post-LASSO regression results for the monthly model The results were obtained from fitting an linear regression model on the selected variables from the LASSO model. Hence, the standard errors are invalid and only serve an indicative purpose due to issues surrounding post selection inference. Note that a variable being selected by the LASSO does not imply that the variable is also statistically significant. Group Predictor Lag Coef S.E. t-stat p-val * Deforestation − 1 0.468 0.110 4.240 0.000 *** Macroeconomic Δ (REER) − 13 -2.487 0.887 -2.804 0.007 *** − 27 -1.694 0.874 -1.938 0.059 * − 35 1.090 0.946 1.153 0.255 Commodities Principal component 1 − 24 1.614 1.357 1.190 0.240 − 35 1.455 1.422 1.024 0.311 Dummies Constant . 0.622 0.296 2.102 0.041 ** Seasonal dummies . . . . . *** R-squared . 82.3% . . . . R-squared adj. . 75.9% . . . . F-stat. . 12.847 . . . . Prob(F-stat.) . 0.000 . . . . Observations . 65 . . . . 28 Figure 13: Better or worse than expected? (Monthly frequency) This figure shows the expected outcomes (blue dashed lines in the box plots) and the observed deforestation levels (colored bars), which are red if deforestation is worse and green if better than expected. Note that the results are not directly comparable with those from Figure 6, which uses annual PRODES data. While PRODES assess land cover changes by using satellite images of previous years, DETER is a daily alert system that is updated every five days. Deforested area in km² Expected deforestation 95% confidence interval 2,000 50% confidence interval Observed deforestation (lower than expected) Observed deforestation (higher than expected) 1,500 1,000 500 0 2016 2017 2018 2019 2020 2021 Source: Instituto Nacional de Pesquisas Espaciais (INPE) DETER, Calculation by authors. 29 References Abadie, A. (2021). Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects. Journal of Economic Literature, 59(2), 391–425. Antle, R., & Demski, J. S. (1988). The Controllability Principle in Responsibility Accounting. The Accounting Review, 63(4), 700–718. Arcand, J.-L., Guillaumont, P., & Jeanneney, S. G. (2008). Deforestation and the real exchange rate. Journal of Development Economics, 86(2), 242–262. Arts, B., Ingram, V., & Brockhaus, M. (2019). The Performance of REDD+: From Global Governance to Local Practices. Forests, 10(10), 837. Assunção, J., Gandour, C., & Rocha, R. (2023). DETER-ing Deforestation in the Amazon: Environ- mental Monitoring and Law Enforcement. American Economic Journal: Applied Economics, 15(2), 125–156. Assunção, J., Gandour, C., & Rocha, R. (2015). Deforestation slowdown in the Brazilian Amazon: prices or policies? Environment and Development Economics, 20(6), 697–722. Baylis, K., Honey-Rosés, J., Börner, J., Corbera, E., Ezzine-de-Blas, D., Ferraro, P. J., Lapeyre, R., Persson, U. M., Pfaff, A., & Wunder, S. (2016). Mainstreaming Impact Evaluation in Nature Conservation. Conservation Letters, 9(1), 58–64. Burns, A., Campagne, B., Jooste, C., Stephan, D., & Bui, T. T. (2019, August). The World Bank Macro-Fiscal Model Technical Description. World Bank Group. Carvalho, W. D., Mustin, K., Hilário, R. R., Vasconcelos, I. M., Eilers, V., & Fearnside, P. M. (2019). Deforestation control in the Brazilian Amazon: A conservation struggle being lost as agreements and regulations are subverted and bypassed. Perspectives in Ecology and Conservation, 17(3), 122–130. Curtis, P. G., Slay, C. M., Harris, N. L., Tyukavina, A., & Hansen, M. C. (2018). Classifying drivers of global forest loss. Science, 361(6407), 1108–1111. de Smit, V. (2022). Investigating the Bolsonaro Effect on Brazilian Deforestation (Master Thesis). Vrije Universiteit Amsterdam. Amsterdam. de Souza Ferreiro Filho, J. B., & Hanusch, M. (2022). A Macroeconomic Perspective of Structural Defor- estation in Brazil’s Legal Amazon (World Bank Policy Research Working Papers No. 10162). World Bank Group. Washington D.C. Ferraro, P. J. (2009). Counterfactual thinking and impact evaluation in environmental policy. New Directions for Evaluation, 2009(122), 75–84. Ferraro, P. J., & Hanauer, M. M. (2014). Advances in Measuring the Environmental and Social Impacts of Environmental Programs. Annual Review of Environment and Resources, 39(1), 495–517. Fletcher, R., Dressler, W., Büscher, B., & Anderson, Z. R. (2016). Questioning REDD+ and the future of market-based conservation. Conservation Biology, 30(3), 673–675. Flugge, M. L., Mok, R. C. K., & Stewart, F. E. (2021, November). Striking the Right Note: Key Performance Indicators for Sovereign Sustainability-Linked Bonds. World Bank. Washington, DC. 30 Granziera, B., Hamrick, K., & Comstock, M. (2021). Eligibility Requirements for REDD+ Standards and Financing. The Nature Conservancy and Conservation International. Guizar-Coutiño, A., Jones, J. P. G., Balmford, A., Carmenta, R., & Coomes, D. A. (2022). A global evaluation of the effectiveness of voluntary REDD+ projects at reducing deforestation and degradation in the moist tropics. Conservation Biology, 36(6), e13970. Hanusch, M. (Ed.). (2023). A Balancing Act for Brazil’s Amazonian States: An Economic Memorandum. The World Bank. Holland, P. W. (1986). Statistics and Causal Inference. Journal of the American Statistical Association, 81(396), 945–960. Holmström, B. (1979). Moral Hazard and Observability. The Bell Journal of Economics, 10(1), 74. Holmström, B. (2016, December 8). Pay For Performance and Beyond [Prize Lecture]. Huettner, M., Leemans, R., Kok, K., & Ebeling, J. (2009). A comparison of baseline methodolo- gies for ’Reducing Emissions from Deforestation and Degradation’. Carbon Balance and Management, 4(1), 4. ICMA. (2020, June). Sustainability-Linked Bond Principles. International Capital Market Association. Paris. IMF. (2017). State-Contingent Debt Instruments for Sovereigns. IMF Policy Papers, 17(23). Lindner, P., & Chung, K. (2023). Sovereign ESG Bond Issuance (IMF Working Paper No. 2023/058). International Monetary Fund. Washington, DC. Mertz, O., Grogan, K., Pflugmacher, D., Lestrelin, G., Castella, J.-C., Vongvisouk, T., Hett, C., Fensholt, R., Sun, Z., Berry, N., & Müller, D. (2018). Uncertainty in establishing forest reference levels and predicting future forest-based carbon stocks for REDD+. Journal of Land Use Science, 13(1), 1–15. Ministry of Economy and Finance of Uruguay. (2022). Uruguay’s Sovereign Sustainability-Linked Bond (SSLB) Framework. Uruguay. Montevideo. Ministry of Finance of Chile. (2022, February). Chile’s Sustainability-Linked Bond Framework. Ministry of Finance of Chile. Ministry of the Environment of Brazil. (2018, January). Brazil’s submission of a Forest Reference Emis- sion Level (FREL) for reducing emissions from deforestation in the Amazonia biome for REDD+ results-based payments under the UNFCCC from 2016 to 2020. Ministry of the Environment. Brasília. Ministry of the Environment of Brazil & Ministry of Science, Technology and Innovation of Brazil. (2014, September). Brazil’s submission of a Forest Reference Emission Level (FREL) for reducing emissions from deforestation in the Amazonia biome for REDD+ results-based payments under the UNFCCC. Ministry of the Environment; Ministry of Science, Technology and Innovation. Brasília. Mirrlees, J. A. (1999). The Theory of Moral Hazard and Unobservable Behaviour: Part I. The Review of Economic Studies, 66(1), 3–21. Richards, P. (2021). A Key Ingredient in Deforestation Slowdowns? A Strong Brazilian Economy. Frontiers in Forests and Global Change, 4. 31 Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66, 688–701. Shavell, S. (1979). Risk Sharing and Incentives in the Principal and Agent Relationship. The Bell Journal of Economics, 10(1), 55–73. Silva, A. C., Stewart, F. E., Baghdassarian, W., & Gurhy, B. (2023). Credit Enhanced Sustainability- Linked Funds and Bonds: a proposed approach to scale [forthcoming]. World Bank Group. Washington D.C. SSDH. (2023). More for Less: Scaling Sustainability-linked Sovereign Debt. Sustainability-linked Sovereign Debt Hub. Geneva, Switzerland. Tibshirani, R. (1996). Regression Shrinkage and Selection via the Lasso. Journal of the Royal Statistical Society. Series B (Methodological), 58(1), 267–288. Townsend, R. M. (1979). Optimal contracts and competitive markets with costly state verification. Journal of Economic Theory, 21(2), 265–293. UNFCCC. (2011, March 15). Report of the Conference of the Parties on its sixteenth session, held in Cancun from 29 November to 10 December 2010. Addendum. Part two: Action taken by the Conference of the Parties at its sixteenth session (FCCC/CP/2010/7/Add.1). United Nations. UNFCCC. (2014, January 31). Report of the Conference of the Parties on its nineteenth session, held in Warsaw from 11 to 23 November 2013. Addendum. Part two: Action taken by the Conference of the Parties at its nineteenth session (FCCC/CP/2013/10/Add.1). United Nations. Valeriano, D. d. M., Mello, E. M. K., Moreira, J. C., Shimabukuro, Y. E., & Duarte, V. (2004). Monitoring Tropical Forest from Space: The PRODES Digital Project. 35 part. b, 272–274. Volz, U. (2023). On the Potential of Sovereign State-Contingent Debt in Contributing to Better Public Debt Management and Enhancing Sustainability Outcomes. Journal of Globalization and Development, 13(2), 379–409. West, T. A. P., Börner, J., Sills, E. O., & Kontoleon, A. (2020). Overstated carbon emission reduc- tions from voluntary REDD+ projects in the Brazilian Amazon. Proceedings of the National Academy of Sciences, 117(39), 24188–24194. West, T. A. P., & Fearnside, P. M. (2021). Brazil’s conservation reform and the reduction of defor- estation in Amazonia. Land Use Policy, 100, 105072. West, T. A. P., Wunder, S., Sills, E. O., Börner, J., Rifai, S. W., Neidermeier, A. N., Frey, G. P., & Kontoleon, A. (2023). Action needed to make carbon offsets from forest conservation work for climate change mitigation [Publisher: American Association for the Advancement of Science]. Science, 381(6660), 873–877. World Bank Group. (2023). Brazil Country Climate and Development Report (CCDR Series). World Bank Group. Washington DC. 32