Version of Record: https://www.sciencedirect.

com/science/article/pii/S0959652619304019
Manuscript_18532ea3658ee44271b79ad885fbbcef

The Green Advantage:

Exploring the Convenience of Issuing Green Bonds

Gianfranco Gianfrate (corresponding author)

EDHEC Business School

393 Promenade des Anglais
06200 Nice
France
Gianfranco.gianfrate@edhec.edu

Mattia Peri

Bocconi University
Via Sarfatti 25
20136 Milan
Italy
Mattia.Peri@studbocconi.it

Abstract

The issue of how to finance the transition to a low-carbon economy in order to achieve the Paris
Agreement’s goal is crucial, especially considering the enormous amount of financing
necessary to shift from rhetoric into action. Green bonds have recently emerged as one of the
best candidates to help mobilizing financial resources towards clean and sustainable
investments. Despite the growing relevance of green bonds, there is limited evidences on
whether such bonds are actually convenient in comparison to other bonds with similar
characteristics except for the “greeness”. By adopting a propensity score matching approach,
we study 121 European green bonds issued between 2013 and 2017. We find that green bonds
are more financially convenient than non-green ones. The advantage is larger for corporate
issuers, and it persists in the secondary market. Our findings support the view that these bonds
can potentially play a major role in greening the economy without penalizing financially the
issuers.

Keywords: green bond; sustainability; responsible investment; propensity score; securities issuance.

1
© 2019 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
1. Introduction
While all countries committed under the Paris Agreement to limit global temperature rise to

1.5˚C-2˚C, the major question remains how can the world achieve this temperature goal. IPCC

(2018) finds that "rapid, far-reaching and unprecedented changes in all aspects of society" must

happen to ensure targeted temperature. Those changes will require profound transitions in land,

energy, industry, buildings, transport, and cities.

The financial system will be crucial to support and to accelerate investments in the clean energy

and technologies needed to decarbonise the economy. This is why the 196 participating

countries in the Paris Agreement committed to “make finance flows consistent with a pathway

towards low greenhouse gas emissions and climate-resilient development” in order to hold the

increase in the global average temperature to well below 2° C above pre-industrial levels.

IPCC (2018) estimates that those “finance flows” amount to about $2.4 trillion (roughly, 2.5%

of the global Gross Domestic Product annually) between 2016 and 2035. Such enormous figure

is also consistent with the analysis by the OECD (2017), according to which approximately

$103 trillion of additional investments will be required between 2016 and 2030 to meet global

development needs in a way that is climate compatible. Similarly, McKinsey (Woetzel et al.,

2016) anticipates cumulative needs for about USD 49 trillion, excluding primary energy and

energy efficiency, between 2016 and 2030. Batthacharya et al. (2016) estimates these

infrastructural needs to be between USD 75 and 86 trillion, including primary energy and

energy efficiency. All the estimates imply that a large portion of the global financial system

needs to be activated to prevent the ultimate climatic collapse.

The IPCC report is an alarming warning and it implicitly confirms the unprecedented

investment opportunity that can be unlocked when sustainable finance becomes mainstream.

With banks having restricted lending capabilities and public budgets under strain in many

countries, private sector sources of capital need to be engaged and green bonds are considered

2
among the key instruments to mobilize private financial resources towards the progressive

decarbonisation of the global economy (World Bank, 2015; OECD, 2017).

Green bonds are a relatively new type of bonds defined by the International Capital Markets

Association (ICMA) as “any type of bond instrument where the proceeds will be exclusively

applied to finance or re-finance, in part or in full, new or/and existing eligible green projects”.

In other words, green bonds are conventional bonds – public debt issued by corporates,

municipalities and other governmental entities – with a distinguishing feature: proceeds are

used for environment-friendly projects, primarily related to climate change mitigation and

adaptation. The evolution of this market over the last years confirms the tremendous potential

of this financial instrument. Indeed, since the European Investment Bank (EIB) issued the first

Green bond in 2007, the market has kept growing and becoming more sophisticated. Green

bond issuance is estimated in $250 billion for 2018 by Moody’s and the Climate Bond Initiative

and expected to reach $1 trillion by 2021.

While the relevance of green bonds is widely recognized by financial professionals, little is

known about the convenience of green bonds for corporate and non-corporate issuers. This

paper investigates how the financial market prices green bonds, and whether issuers can lower

their financial costs by issuing a bond labelled as “green” rather than an equivalent non-green

bond (“conventional bond” in the remaining of the paper). Indeed, for companies and non-

corporate organizations the most important driver in investment decisions is the funding cost.

Since the majority of the electricity generation costs are in the financial costs of capital (OECD,

2015), even small differences in the cost of bonds can have a substantial impact on the long-

run sustainability of energy and large industrial facilities. Therefore, assessing the relative

convenience of green bonds, in terms of returns to be paid to investors, is of paramount

importance. Our paper is, to the best of our knowledge, the first measuring the financial cost

for issuers of green bonds and thus estimating the relative convenience to issue bonds labelled

as green versus conventional ones. Our results show that there is a statistically significant
3
advantage, for the issuers, when a bond is labelled as green. Such advantage is quantifiable, on

average, in lower interests paid annually to investors of 18 basis points (meaning 0.18% of the

bond value). Furthermore, such advantage is achieved by both companies and non-corporate

entities like municipalities and governmental agencies. These findings suggest that, even taking

into account the extra-costs needed to obtain a green certification for the issuance, green bonds

are relatively more convenient for the issuers. Hence, green bonds are potentially beneficial not

only to society, but also to the issuers because they can reduce the cost of debt financing.

The remainder of the paper is organised as follows. Section 2 introduces the relevant literature

on the Green bonds market performance in primary and secondary markets. Section 3 describes

the data and the samples that will be used to carry out the analysis. Section 4 presents the main

methodological approach. Section 5 describes the findings obtained using this empirical model.

Section 6 discusses the implications of the findings and the limitations of the paper, and Section

7 concludes.

2. Literature Review
Green bonds are a recent phenomenon with a widespread growth across countries started not

earlier that 2013. Consequently, the scholarly literature on green bonds is limited. Ge and Liu

(2015) examining how a firm’s corporate social responsibility (CSR) performance is associated

with the cost of its new bond issues in the US market, found that firms with better CSR

performance are able to issue bonds at lower cost. Similar conclusions have been reached by

Oikonomou, Brooks and Pavelin (2014). Bauer and Hann (2010), analyzing a large cross-

industrial sample of US public corporations, found that environmental concerns are associated

with a higher cost of debt financing and lower credit ratings, and proactive environmental

practices are associated with a lower cost of debt. Stellner, Klein and Zwergel (2015) found

only weak evidence that superior corporate social performance (CSP) results in systematically

reduced credit risk. On the contrary, Menz (2010), focusing on the European corporate bond

4
market, observed that the risk premium for socially responsible firms was, ceteris paribus,

higher than for non-socially responsible companies, although this finding is only slightly

significant. Zerbib (2017) has analysed the green bond advantage focusing on 135 investment

grade green bonds issued worldwide. The paper shows that bondholders pay of 8 basis points

(statistically significant) to buy green bonds after issuance. Evidences have been collected for

non-corporate issuers as well. Karpf and Mendel (2017) investigated green and conventional

bonds in the U.S. municipal bonds market and found that green bonds seem to be penalized by

the market.

All in all, the evidences about the existence of a green advantage in the primary market (when

bonds are initially issued) and in the secondary market (when bonds are traded following the

issuance) are mixed. Further research on this topic is therefore needed, especially with more

data available and apparent growing interests from both issuers and investors. Our study extends

the literature on green bonds by providing evidence of the existence of a significant advantage

for the primary market of European green bonds adopting a propensity score matching (PSM)

methodology and suggesting that this premium persists also after the issuance (the secondary

market).

3. Data description
We set up our samples in order to evaluate, through propensity score matching techniques, the

difference between returns at issuance of green bonds and their conventional peers. Our data

come from “Bond Radar” of Bloomberg. Specifically, our initial sample comprises all the bonds

issued from January 2007 to December 2017. For every bond, Bond Radar provides detailed

information about the bond issues’ and issuers characteristics including the returns offered to

investors.

As of December 2017, Bond Radar reports 7589 public EUR-denominated bonds issued since

January 2007, of which 154 are classified as green. We eliminate from the sample all the bonds
5
with variable interest payments (to avoid the uncertainty that varying payments could have on

the pricing at issuance), all those bonds for which the return is not available or with a size lower

than EUR 200 million (in order to have only liquid bonds), all bonds at high risk of default and

those that are not priced using European rates.

Following these changes, there are 121 green bonds left in the dataset and they are issued by

entities of different nature: corporates, sovereign states, national and multinational agencies,

municipalities, financial institutions. Our comprehensive sample (“All”) comprises 3055 bond

issues, of which 121 are labelled as green. Following an analogous procedure, we define two

subsamples: “Corporate Issuers” and “Non-corporate Issuers”, which contain respectively all

the bonds issued by corporations and all the bonds issued by the other entities. Corporate

Issuers is constituted by 781 observations of which 43 are labelled as Green; Non-corporate

Issuers is constituted by 2155 observations of which 78 are labelled as Green. Table 1 shows

the descriptive statistics for the comprehensive sample.

TABLE 1

Descriptive Statistics

Variable Mean Std. Dev. Min Max

Volume (EUR mln) 620 333 250 2000

Maturity (Years) 8.15 3.95 3.02 30.40
Spread (Basis point) 28 41 -32 140

4. Methodology
To address the question of whether issuing green bonds is convenient, we should compare the

returns of the green bonds with those of their conventional peers. To make the comparison we

6
use propensity score matching techniques which are suited to empirical settings where there is

a “treatment”, a group of treated observations and a group of untreated observations. This is

exactly our case: we refer to “getting the green label” as the treatment, to “Green bonds” as the

treatment group, and to “Conventional bonds” as the untreated group. The change in the

outcome variable (i.e. the return at issuance) due to the treatment is our treatment effect.

The problem of computing the treatment effect is that a real counterfactual framework would

require observing each bond being priced in both states (with and without treatment), and this

is clearly not possible: we can observe only one outcome for each bond. Consequently, given

an observed outcome (e.g. the spread given that the bond is labelled as Green), the

counterfactual outcome has to be estimated. PSM techniques allow to estimate the

counterfactuals. Specifically, in this paper we will estimate the “average treatment effect on the

treated” (ATT).

To obtain the best possible estimation of the counterfactuals and ATT, we would need to build

a control group (a group of conventional bonds) that is ideally identical to the treated group in

everything but for the treatment status. However, treated and untreated bonds usually differ in

other characteristics apart from treatment status, and assignment to treatment and control group

will not be random. For instance, firms that operate in the utility and power sector may have a

higher probability of issuing Green bonds because they are clearly more involved in climate

change and environment-related issues. Hence, comparing the mean values of the returns

between treated and untreated bonds would lead to biased results.

A way for overcoming this problem is to find a control group that is as similar as possible to

the treated group in all relevant (observable) pre-treatment characteristics “X”. That being done,

differences in outcomes of this well selected and thus adequate control group and of treated

group can be attributed to the treatment, i.e. to the Green label. The problem is that, as the

number of characteristics determining selection increases, it is more and more difficult to find

comparable individuals (“curse of dimensionality”). Rosenbaum and Rubin (1983 and 1984)
7
describe how we can bundle such characteristics in a single-index variable, the propensity score,

which makes it possible to achieve consistent estimates of the treatment effect in the same way

as matching on all covariates.

To be more specific, estimating an ATT using propensity score matching involves a two-step

procedure (Wamser, 2014). In the first step, we estimate a propensity score to predict the

probability of bonds of being Green, using a Logit and a Probit function. In the second step, we

match green (treated units) and conventional bonds (control units) and estimate the treatment

effect by computing the difference in the returns (outcome variable) between matched units.

The matching procedure is based on the propensity score, which is a continuous variable, that

we obtained in the first part of the process. Despite all matching estimators compare the

outcome of a treated unit with outcomes of control group members, we need to make sure to

use the appropriate PSM estimators among those available. Moreover, three main conditions

need to be satisfied in order to effectively use PSM techniques. The first one is the “conditional

independence assumption” (CIA), which requires that the outcome variable (the returns) must

be independent of treatment conditional on the propensity score. In other words, it requires that

the common characteristics that affect treatment assignment and outcomes be observable. This

is a strong assumption and it is impossible to verify so that bias resulting from unobservable

characteristics can never be ruled out. This is clearly the main limit of this kind of techniques.

The second condition is the “common support”, i.e. the presence in both groups of units with

similar propensity scores. Implementing the common support is necessary to avoid the

comparison of “incomparable members” of the groups. The third and last condition is that the

propensity score balances the covariates: similar propensity scores have to be based on similar

observed characteristics.

In our analysis, we apply the nearest neighbours matching (NN) with 3, 5, and 8 matches, the

kernel matching and the radius matching with different levels of the radius (“r”).

8
With the nearest neighbours matching the indicated numbers of units from the comparison

group (3, 5 or 8 in our case) are chosen as matching partners for a single treated unit that is

closest in terms of propensity score. In particular, we implement this matching method “with

replacement”, i.e. we allow members of the control group to be used more than once as

matching partners for treated units. Matching with replacement enhances the average quality of

matching and decreases the bias (assuming some re-use occurs) but, at the same time, increases

the variance of the estimator (Smith and Todd, 2005). A possible drawback of this methods is

that the indicated number of matches are assigned to every treated bond, no matter how close

propensity scores actually are, which may result in a rather unsatisfying matching quality.

Radius matching may help to solve this problem: treatment units are matched to control units

only if the propensity scores of the latter are within a certain, pre-definite, range. The smaller

we define the radius (r), which defines the tolerable distance within which units are matched,

the better is the quality of the matches. However, if the propensity scores are “well balanced”

between the treatment and control groups, occurrence of bad matches increases with radius

matching compared to nearest neighbours matching.

Finally, the Kernel matching estimator calculates the weighted averages of all units in the

control group to construct the counterfactual outcome; the closer the propensity score of a given

untreated unit is to the one of the treated unit, the higher its weight will be.

To evaluate different matching methods, we need to take into account the trade-off between the

number of matches (quantity) and their quality. Testing the balancing properties (third

condition) of the various methods that we implement, we find that the most balanced matching

is obtained by applying the nearest neighbours matching with 8 control units for every Green

bond. The results of these tests will be shown in section 5.

9
5. Results
The section is structured as follows: in paragraph 5.1 we analyse the comprehensive sample

and show the results of the nearest neighbours matching with 8 matches for each Green bond;

then, we present and compare the results of the different matching techniques. In paragraph 5.2,

we perform the same analysis on the subsamples of corporate and non-corporate issuers.

Finally, in paragraph 5.3 we look for the persistence of the relatively favourable trading of green

bonds in the secondary market.

5.1 Primary Market

The first stage of the process to estimate the ATT involves obtaining the propensity score. We

then assess if the propensity score (estimated through the Logit function) is properly specified

by applying the “blocking” procedure (Rosenbaum and Rubin, 1983): first, data are sorted by

propensity score and divided into blocks of observations with similar propensity scores; within

each block, it is tested whether the propensity score is balanced between treated and control

observations. If not, blocks are too large and need to be split. If, conditional on the propensity

score being balanced, the covariates are unbalanced, the specification of the propensity score is

not adequate and has to be re-specified.

In our case, the optimal number of blocks, which ensures that the mean propensity score is not

different for treated and controls in each block, is 10 and the balancing property of the

propensity score is completely satisfied (i.e. also the covariates are balanced within each block).

We then conclude that the propensity score is well specified. “Appendix b” shows the inferior

bound, the number of treated and the number of controls for each block.

Table 2 presents propensity score matching results for five different matching procedures (in

columns).

10
 TABLE 2

Primary market spreads treatment effects (in basis points)

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.0005)

ATT -18.47*** -18.56*** -14.90*** -19.40*** -15.17** -16.74***

Std. Err. 4.37 4.76 5.00 5.88 6.74 4.31
# treated 121 121 121 116 100 121
# untreated 535 385 254 1484 893 2934
Notes: The ATT and Std. Err. figures are expressed in basis points. Columns refer to the different matching
methods: Nearest neighbour matching (NN 3,5,8); Radius (r =0.001) matching with 0.1% radius; Radius
(r=0.0005) matching with 0.05% radius; Kernel matching. ATT is the Average Treatment effect on the Treated. #
treated (# untreated) is the number of treated (control) units. The propensity score is based on the logit model
reported in Table 2. (***) (**) (*) indicate significance at the (1%) (5%) (10%) level. In all estimations, a common
probability support of the treated and control units is enforced in order to ensure better comparability of matched
units.

As already highlighted, the propensity score matching approach relies on three basic conditions:

the CIA assumption, the common support, and the propensity score balancing the covariates.

CIA assumption is not testable. On the contrary, common support is implemented and the

results in Table 3 demonstrate that there is an optimal overlap between the treated and untreated

groups. In particular, the 121 Green bonds are all “on support” when nearest neighbours

matching is applied.

The third condition requires that, given random assignment to treatment, bonds with the same

propensity score should have the same distribution of observable variables used to predict the

propensity score. As this balancing property is testable, we provide such tests in Table 3.

11
 TABLE 3

Pstest - Balancing Property

Unmatched Mean %Reduction t-test
Variable Matched Treted Control %Bias |Bias| t |p|>t
Y_2013 U .050 .188 -43.7 -3.87 0.000
M .050 .036 4.2 90.3 0.51 0.608
Y_2014 U .157 .179 -5.9 -0.63 0.531
M .157 .178 -6.1 -2.1 -0.47 0.638
Y_2015 U .182 .211 -7.2 -0.76 0.445
M .182 .176 1.6 78.5 0.13 0.900
Y_2016 U .215 .216 -0.2 -0.02 0.982
M .215 .225 -2.5 -1086.9 -0.19 0.847
Y_2017 U .397 .207 42.2 5.02 0.000
M .397 .383 3.0 92.9 0.21 0.831
ln(Volume) U 63.227 6.557 -41.7 -3.87 0.000
M 63.227 63.155 1.3 96.9 0.11 0.916
Tenor U 81.535 83.259 -3.9 -0.38 0.701
M 81.535 79.422 4.8 -22.6 0.41 0.683
AAA - AA U .364 .463 -20.3 -2.15 0.031
M .364 .367 -0.6 96.9 -0.05 0.960
AA(-) - A U .298 .222 17.4 1.96 0.050
M .298 .285 2.8 83.7 0.21 0.833
A(-) - BBB U .314 .234 18.0 2.03 0.042
M .314 .311 0.7 96.1 0.05 0.959
BBB(-) - BB(+) U .025 .081 -25.4 -2.26 0.024
M .025 .037 -5.6 78.1 -0.55 0.580
Covered U .033 .252 -65.9 -5.53 0.000
M .033 .029 1.2 98.1 0.18 0.854
Western Europe U .868 .838 8.5 0.88 0.379
M .868 .861 2.0 75.9 0.16 0.870
Asia, Australia, New Zealand U .091 .052 15.2 1.88 0.061
M .091 .099 -3.2 78.9 -0.22 0.827
CEEMEA U .0165 .044 -16.2 -1.47 0.141
M .0165 .020 -1.8 88.8 -0.18 0.857
HG Global U .008 .003 6.9 0.98 0.327
M .008 .008 0.0 100.0 -0.00 1.000
North America U .0165 .063 -23.9 -2.09 0.036
M .0165 .012 2.1 91.1 0.27 0.789
Agency - Sovereign U .223 .143 20.6 2.43 0.015
M .223 .220 0.8 96.1 0.06 0.954
Banking U .215 .416 -44.3 -4.43 0.000
M .215 .257 -9.3 79.0 -0.77 0.440
Basic Materials U .025 .0491 -12.9 -1.22 0.221
M .025 .018 3.8 70.2 0.39 0.697
Manufacturing U .008 .087 -37.6 -3.06 0.002
M .008 .001 3.5 90.8 0.82 0.410
Municipality - Local Govt. U .025 .102 -31.9 -2.78 0.005
M .025 .018 3.0 90.6 0.39 0.697
Supra U .182 .082 29.6 3.82 0.000
M .182 .186 -1.2 95.8 -0.08 0.934
Transport and Logistics U .008 .020 -10.0 -0.92 0.358
M .008 .006 1.7 82.6 0.19 0.850
Utilities and Power U .273 .071 55.4 8.16 0.000

12
 M .273 .243 8.2 85.2 0.53 0.596
Real Estate U .041 .029 6.5 0.76 0.446
M .041 .052 -5.6 14.0 -0.38 0.704

Sample Ps R2 LR chi2 p>chi2 MeanBias MedBias B R %Var

Unmatched 0.207 211.12 0.000 23.5 19.2 157.7* 0.37* 100

Matched 0.011 3.64 1.000 3.1 2.7 24.1 1.45 50
* if B>25%, R outside [0.5; 2]
Notes: Tests correspond to the nearest neighbours matching results (NN 8) provided in Table 2. Y_2013, Y_2014,
Y_2015, Y_2016 and Y_2017 are dummy variables indicating whether or not a given bond has been issued in that
year; ln(Volume) is the natural logarithm of the size (amount issued in € million) of a given bond; Tenor is the
time to maturity (in years) of a given bond at launch; AAA – AA, AA(-) – A, A(-) – BBB and BBB(-) - BB(+) are
dummy variables indicating whether or not the credit rating of a given bond falls in that category; Covered is a
dummy variable indicating whether or not a given bond is classified as a “covered bond” (i.e. backed by a separate
group of assets of the issuer); Western Europe, “Asia, Australia, New Zealand”, CEEMEA, HG Global and North
America are dummy variables indicating whether or not the issuer of a given bond has the majority of its operations
(and risks) in that region; Agency – Sovereign, Banking, Basic Materials, Manufacturing, Municipality - Local
Govt., Supranational, Transport and Logistics, Utilities and Power and Real Estate are dummy variables indicating
whether or not the issuer of a given bond operates in that sector.

For all the variables included in the model, we cannot reject the null hypothesis about the

equality of the means between the treated and control groups (see p-values last column).

Notably, following the matching all the “%Bias” are below 10%, with the difference between

the means reduced by more than 80% for the majority of the variables. Moreover, the Rubin’s

B and R are, respectively, lower than 25% and inside the range 0.5 - 2. These tests therefore

demonstrate that also the balancing property is satisfied.

The estimates of the average treatment effect on the treated shown in Table 2 indicate that the

Green label does have a significant impact on bonds pricing in the primary market. Besides,

this finding looks rather robust, irrespective of the matching method used. The estimates are in

the range between -14.8 basis points (NN 3) and -19.4 basis points (radius matching, r=0.0001);

for instance, when using nearest neighbours matching (NN=8 or NN=5), we estimate a

coefficient of about -18.5 basis points, which is significant at the 1% level. Kernel matching

makes the ATT estimate increase by 2 basis points. The biggest average treatment effect on the

treated is estimated when applying radius matching with r=0.1% (-19.4 basis points), while the

greatest standard error is associated with the radius matching with r=0,05%.
13
To recap, findings in Table 2 confirm the existence of a relative convenience of issuing green

versus conventional bonds in the primary market. Green bonds, on average, cost less to issuers

than their conventional peers.

5.2 Primary Market by Issuer type

In this section we try to understand whether or not the results obtained on All are valid

independently on the kind of issuer. In particular, we divide bonds issued by corporations from

those issued by other market players, i.e. banks, governments, local governments,

municipalities, and supranational institutions. Table 4 summarizes the results of the analyses.

TABLE 4

Primary market spreads treatment effects (in basis points)

Sample2 – Corporations

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.01)

ATT -20.80*** -22.51*** -19.71*** -19.80** -21.44*** -21.45***

Std. Err. 5.35 5.31 6.04 10.69 7.40 6.93
# treated 43 43 43 28 38 43
# untreated 164 120 83 128 680 738
Sample3 – Non-corporate Issuers

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.01)

ATT -13.51*** -15.42*** -16.15*** -17.38*** -14.82*** -14.26***

Std. Err. 4.96 5.13 6.39 6.54 4.97 4.55
# treated 78 78 78 66 78 78
# untreated 364 256 166 924 2065 2077
Notes: The ATT and Std. Err. figures are expressed in basis points. Columns refer to the different matching
methods: Nearest neighbour matching (NN 3,5,8); Radius (r =0.001) matching with 0.1% radius; Radius (r=0.01)
matching with 1.0% radius; Kernel matching. ATT is the Average Treatment effect on the Treated. # treated (#
untreated) is the number of treated (control) units. The propensity scores are based on the Logit models reported
in “Appendix c” and “Appendix d”. (***) (**) (*) indicate significance at the (1%) (5%) (10%) level. In all
estimations, a common probability support of the treated and control units is enforced in order to ensure better
comparability of matched units.
14
Interestingly, although the existence of a negative premium is confirmed for both samples, it is

more marked for bonds issued by corporations. Indeed, the ATT for these bonds (the majority

of them operates in the utility and power sector) ranges from -23 basis points (NN=5) to -20

basis points (NN=3) with an average advantage of -21 basis points; on the other hand, the Green

bond advantage for non-corporate issuers ranges from -17 basis points (radius matching with

r=0,1%) to -14 basis points (NN=8) with an average of -15 basis points. These results are

consistent with those obtained on the comprehensive sample: the weighted average of the

average convenience of the two subsamples is the same of the average convenience found in

the full sample (-17 basis points).

Furthermore, all the ATTs but the one estimated through radius matching (r=0.001) for

corporate issuers are statistically significant at the 1% level. In both samples the greatest

standard error is associated with the radius matching (r=0.001) while the lowest is associated

with the nearest neighbours matching with 8 matches.

We conclude this part of the section by running an OLS regression of the spreads on the

variables used to estimate the propensity scores plus an indicator (dummy) variable for Green

bonds; we run such regression on each sample. The results are presented in Table 5.

TABLE 5

Primary market OLS regressions Results

Variable: Coeff. Std. Err. t p>|t| Regression's R2

Sample1
Green dummy -16.63*** 3.51 -4.74 0.000 0.733
Sample2
Green dummy -23.42*** 6.90 -3.40 0.001 0.535
Sample3
Green dummy -10.26** 4.06 -2.52 0.012 0.753
Notes: (***) (**) (*) indicate significance at the (1%) (5%) (10%) level.

15
The coefficients, i.e. the estimates of the Green convenience (lower returns paid to investors),

are statistically significant and in line with the results obtained by using propensity score

matching techniques. However, while for corporate issuers the estimate is lower of about 2-4

basis points than the estimates found through PSM, for non-corporate issuers the value is 3 to

7 basis points higher, depending on which PSM method we consider.

5.3 Secondary market

In this subsection we compare green and conventional bonds pricing in the secondary market.

Before presenting the results, we need to outline some limits of the analysis. The main limit is

that we do not correct the returns for liquidity, i.e. we do not address the problem of a possible

difference in liquidity between bonds (liquidity bias). As noticed in section 3, to carry out the

analysis we download the returns of the bonds from Bloomberg BVAL at different dates. Since

these data are market based, they may be strongly affected by the liquidity of the bonds. Indeed,

the actual problem when dealing with bonds, especially when they are labelled as Green, is that

they are usually bought in the primary market by institutional investors and held until maturity.

Hence, even if they could be potentially liquid, in practice they are not traded in the secondary

market so that their market prices are often not reliable. The second issue is that we just

download the data at three different dates, six months apart from each other: 14 December 2017,

7 July 2017 and 10 January 2017. This implies that we cannot observe the potential volatility

of the returns and its evolution over time. We do not take into consideration earlier data because

there would be too few Green bonds available to effectively implement propensity score

matching techniques. We will consider only the returns as of 14 December 2017 when we will

focus on corporate issues and non-corporate issues because of the lack of Green bonds that had

already been issued in July and January 2017.

16
Table 6 shows the results of the propensity score matching techniques applied on corporate

issues. As can be noticed, as of 14 December 2017 there seems to exist a relative convenience

for Green bonds of about -5 basis points. In particular, the ATT ranges from -3.8 (Kernel

matching) to -7.6 basis points (radius matching with r equal to 0.05%).

TABLE 6

Secondary market spreads treatment effects (in basis points)

Sample1 - 14 December 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.0005)

ATT -5.38* -5.75* -6.04* -5.33 -7.61** -3.75

Std. Err. 2.97 3.13 3.46 3.42 3.66 3.14
# treated 118 118 118 117 114 118
# untreated 544 393 257 1511 988 2799
Sample1 - 7 July 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.0005)

ATT -12.69*** -13.73*** -13.44*** -10.51** -13.85*** -9.11***

Std. Err. 2.45 2.97 2.97 4.43 5.20 3.65
# treated 93 93 93 91 84 93
# untreated 433 301 206 1449 969 2307
Sample1 - 10 January 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.001) (r=0.01)

ATT -11.28** -10.90** -11.55** -10.71* -9.05* -8.84*

Std. Err. 4.51 4.59 5.45 6.05 5.17 5.01
# treated 70 70 70 68 70 70
# untreated 353 240 153 832 1726 1740
Notes: The ATT and Std. Err. figures are expressed in basis points. Columns refer to the different matching
methods: Nearest neighbour matching (NN 3,5,8); Radius (r =0.001) matching with 0.1% radius; Radius
(r=0.0005) matching with 0.05% radius; Kernel matching. ATT is the Average Treatment effect on the Treated. #
treated (# untreated) is the number of treated (control) units. (***) (**) (*) indicate significance at the (1%) (5%)
(10%) level. In all estimations, a common probability support of the treated and control units is enforced in order
to ensure better comparability of matched units.

17
The estimates of the average treatment effects are statistically significant at the 10% level using

the nearest neighbours matching and the radius matching with r equal to 0.05%, while “flat

pricing” cannot be rejected when the Kernel matching and the radius matching with r equal to

0.1% are applied. However, the balancing property of the propensity score for the

comprehensive sample are not completely satisfied. Conversely, balancing properties are

satisfied when the two sub-samples are considered; the results are presented in Table 7.

18
 TABLE 7

Secondary market spreads treatment effects (in basis points)

Sample2 - 14 December 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.01) (r=0.005)

ATT -6.40*** -6.53** -6.84** -7.78 -8.04 -7.99

Std. Err. 2.22 2.57 2.60 6.67 7.94 5.89
# treated 43 43 43 39 39 43
# untreated 201 145 92 710 608 720
Sample3 - 14 December 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.01) (r=0.001)

ATT -8.12*** -9.26*** -9.79*** -7.94** -8.92** -7.61**

Std. Err. 2.85 2.88 3.11 3.21 4.19 3.03
# treated 75 75 75 74 71 74
# untreated 319 236 155 1942 1134 1963
Sample3 - 7 July 2017

Neighbours Neighbours Neighbours Radius Radius

Kernel
Matching: matching matching matching matching matching
matching
(NN=8) (NN=5) (NN=3) (r=0.005) (r=0.01)

ATT -13.16*** -14.41*** -13.80*** -11.60*** -10.40*** -10.33***

Std. Err. 3.53 3.64 4.24 3.81 3.58 3.30
# treated 63 63 63 63 63 63
# untreated 358 204 130 1731 1807 1811
Notes: The ATT and Std. Err. figures are expressed in basis points. Columns refer to the different matching
methods: Nearest neighbour matching (NN 3,5,8); Radius (r =0.01) matching with 1.0% radius; Radius (r=0.001)
matching with 0.1% radius; Kernel matching. ATT is the Average Treatment effect on the Treated. # treated (#
untreated) is the number of treated (control) units. (***) (**) (*) indicate significance at the (1%) (5%) (10%)
level. In all estimations, a common probability support of the treated and control units is enforced in order to
ensure better comparability of matched units.

As of 14 December 2017, the ATT for corporations is estimated to be between -7.6 basis points

(Kernel matching) and -9.8 basis points (NN 3), while the ATT for the other issuers is estimated

to be between -10.3 basis points (Kernel matching) and 14.4 basis points (NN5). In both cases,

19
the ATTs estimated through the nearest neighbours matching are statistically significant at least

at the 5% level. On the contrary, radius and Kernel matching gives very high standard errors

when the corporate issuers subsample is analysed so that we cannot say that the corresponding

ATTs are statistically different from zero.

As of 7 July 2017 and 10 January 2017, the ATTs are respectively between -9.1 basis points

and -13.9 basis points, and between -8.8 basis points and -11.5 basis points. Notably, all the

matching methods but the radius matching (r=0.1%) give estimates of the ATT significant at

the 1% level when implemented on the data of July. On the other hand, in January the estimates

are significant at the 5% level when using nearest neighbours matching and statistically

different from zero with a confidence level of 10% when using radius and Kernel matching.

These findings seem to confirm that the Green label does have an impact on the bonds’ returns

also in the secondary market, even if lower than in the primary market. The presence of a Green

convenience in the secondary market is in line with the results of Zerbib (2017). Moreover,

looking at the difference between the ATT of December and the ones of July and January the

question whether the convenience changes over time as the market grows and evolves arises. A

possible explanation of that difference is that, as the supply of Green bonds is surging, the

demand is not growing at the same pace, so that the yields of Green bonds tend to converge

towards those of their conventional peers. In theory, this should also be reflected in the primary

market spreads, but with a PSM approach such a change cannot be spotted, especially

considering the relatively limited number of green bonds to date.

6. Discussion
In this paper we evaluate the convenience of issuing green bonds for companies and non-

corporate entities that want to invest in green projects such as renewable energy plants, energy

and water efficiency, electrification of transport, fuel switching, bioenergy. We show that green

bonds are actually more convenient than conventional bonds, because on average, coeteris

20
paribus, they have to offer to the investors lower returns. Importantly, such result is stronger

for corporate issuers with the implication that private sector - whose support will be necessary

to achieve the Paris Agreement’s temperature goal – are better off financially when they issue

bonds that are labelled as green. Green bonds have some additional transaction cost because

issuers must certify, monitor and report on the green use of proceeds. The magnitude of the

savings for issuers (in terms of interests paid) exceeds the costs to get the green label or rating.

The Climate Bonds Initiative, for instance, asks a flat fee equal to 0.1 basis points of the issue

value in order to certificate the green label (although it also requires the engagement of third-

party that verifies all the reports and procedures). Moreover, even if the green assessment were

as expensive as normal credit ratings, it would cost up to 3-5 basis points (White 2002), which

is still far lower than the savings we estimate (15-21 basis points).

The relative financial savings obtained by the issuers appear to be the consequence of a strong

demand for these financial products, which reflects the interest of investors willing to fund

green projects. With regard to the possible drivers of this growing demand, it is clear that green

bonds allow investment firms to fulfil the requests of clients sensitive to environment-related

issues. More and more institutional investors are decarbonising their portfolios and redirecting

funds towards environment-friendly investments because they regard climate change as

growing threat to the long-term economic growth. Additionally, in some case the demand has

been sustained also by national regulations; for instance, the French energy transition law forces

institutional investors to report on how they are contributing to reduce CO2 emissions and, more

broadly, on how they are managing climate-related risks (article 173, French Treasury, 2015).

Bank of England and the Securities and Exchange Board of India (UNEP 2016) have issued

requirements to promote and develop the green bond markets. It is likely and desirable that in

the coming years an increasing number of countries will take actions to sustain and promote the

development of this market. For example, governments could offer tax advantages for green

investors in order to help drive the market.

21
Focusing on Europe, a boost to new green bonds’ issuances is expected following the release

of the European Commission’s Action Plan: Financing Sustainable Growth (European

Commission, 2018), especially because of the adoption of a common European Union green

asset taxonomy. Banks are expected to start implementing the reporting recommendations set

out by the Taskforce for Climate-related Financial Disclosure (TCFD, 2017) and monetary

policies that favour the investment in green-labelled assets become more and more likely.

Hence, green bonds not only can help issuers to achieve better financial results and the global

economy to become more sustainable (therefore benefitting the society), but they can also help

investors to comply with the current and future regulation.

The main limitation of our results is the relatively small sample of bonds studied. Considering

the tremendous potential of this kind of financial instruments in facilitating the transition

towards a sustainable economy, we believe that the existence and magnitude of a relative

convenience to issue green bonds should be studied on larger samples and in other geographical

regions. In particular, as the quantity and quality of data will improve, further investigation

could shed light on several aspects relevant for issuers, investors, and policymakers. This

research agenda should encompass whether green bonds’ convenience has materially changed

over time and how much it has varied across industries and regions. Second, it would be

important to better analyse and understand the composition of the demand for these bonds and

its drivers in order to design policies able to sustain and foster the growth of the market. Finally,

the definition of the green label should be studied more in depth. Being new financial

instruments, these bonds suffer of a certain ambiguity in terms of what can be labelled as green

and what cannot. The creation of reliable and consistent criteria for the labelling requires both

scientific analysis and political and standard-setting decisions. This is why, for example, the

European Commission is working towards a formal taxonomy of what is really green in finance

thus avoiding the misuse of the label.

22
7. Conclusion
Limiting warming to 1.5º C is possible but doing so will require unprecedented changes in the

economic and social systems. Financial markets will play a major role in those disruptive

changes and practitioners, policymakers, and scholars are converging in stressing how crucial

the support of finance is in delivering an actual and timely transition to a low carbon economy.

Green bonds are widely assumed to be one of the key instruments to mobilize financial

resources towards to achieve the temperature goal of the Paris Agreement.

Since the European Investment Bank issued the first green bond in 2007, the growth of these

bonds has been exponential. In this paper we investigate whether green bonds as convenient as

similar conventional bonds for the issuers. We specifically study European green bonds using

a propensity score matching approach.

We show that green bonds can represent an effective way for achieving a lower cost of capital

for organizations that need to finance or re-finance green projects. Since there is limited

evidence on corporate decisions in issuing green bonds, our findings also fill in the literature

gap and provide insights the convenience of these financial instruments for companies. This

will also help policymakers identify which policies can deter or encourage the market and

issuers to access innovative green financing instruments.

Acknowledgments
We are grateful for comments from three anonymous referees, Oğuzhan Karakaş (discussant),
Dirk Schoenmaker, and participants at Erasmus University’s Dynamics of Inclusive Prosperity
Conference and at CEP’s Scaling up Green Finance Conference.

References
Pan W. & Bai H. (2015). Propensity score analysis, Concepts and Issues. New York: The
Guilford Press.

Batthacharya A. et al. (2016). Framework for assessing the role of sustainable infrastructure.
Brookings Instituion, Washington DC.

23
Bauer, R. & Hann, D. (2010). Corporate environmental management and credit risk. Working
paper .
URL: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1660470

Bloomberg (2017). Investors are willing to pay a "Green" premium. Bloomberg New Energy
Finance report.

Caliendo, M. & Kopeining, S. (2005). Some Practical Guidance for the Implementation of
Propensity Score Matching. IZA DP No. 1588

Carney, M. (2015). "Breaking the tragedy of the horizon - climate change and financial
stability". speech given at Lloyd's of London, 29 September.

Climate Bonds Initiative (2016). Green bond pricing in the primary market: January 2016 -
March 2017.

Climate Bonds Initiative (2018). Green Bond Highlights 2017.

URL: https://www.climatebonds.net/files/reports/cbi-green-bonds-highlights-2017.pdf

European Commission (2018). European Commission’s Action Plan: Financing Sustainable

Growth. URL: https://ec.europa.eu/info/publications/180308-action-plan-sustainable-
growth_en

French Treasury (2015). Decree no. 2015-1850.

URL: https://www.legifrance.gouv.fr/eli/decret/2015/12/29/2015-1850/jo/texte

Ge, W. & Liu, M. (2015). Corporate social responsibility and the cost of corporate bonds.
Journal of Accounting and Public Policy 34 (6), 597–624.

Hirtenstein, A. (2017). Evidence Mounts for Green Bonds Outperforming Conventional: HSBC.
Bloomberg Markets, 5 September 2017.
URL: https://www.bloomberg.com/news/articles/2017-09-05/evidence-mounts-for-Green-
bonds-outperforming-conventional-hsbc

I4CE (2016). Enhancing Green bonds’ contribution to the low-carbon and climate resilient
transition. Institute for Climate Economics.

Ibikunle, G. & Steffen, T. (2015). European Green Mutual Fund Performance: A Comparative
Analysis with their Conventional and Black Peers. Journal of Business Ethics.

IIGCC (2017). Implementing the Paris Agreement: Increasing engagement and driving
disclosure. 2017 Year in Review.
URL: https://www.climatebonds.net/market/investor-appetite

Karpf, A. & Mandel, A. (2017). Does it pay to be Green? Working Paper.

Kovner, A. & Wei, C. (2014). The Private Premium in Public Bonds. Federal Reserve Bank of
New York No. 553.

KPMG Advisory N.V. (2016). Mainstreaming the green bond market: pathway towards
common standards.

24
Menz, K. (2010). Corporate social responsibility: is it rewarded by the corporate bond market?
A critical note. Journal of Business Ethics 96, 117–134.

Morgan Stanley (2017). Why We Think Green Bonds Will Continue to Grow: A Primer. Morgan
Stanley report.
URL: https://www.morganstanley.com/ideas/Green-bond-boom

Natixis (2017). Is the EIB paving the way for a "Green Premium"? Natixis Report.

OECD (2015). Mapping Channels to Mobilise Institutional Investment in Sustainable Energy.

Green Finance and Investment, OECD Publishing, Paris .
URL: https://www.oecd.org/g20/topics/energy-environment-Green-growth/mapping-
channels-to-mobilise-institutional-investment-in-sustainable-energy-9789264224582-en.htm

OECD (2017a). Green bonds: Mobilising Bond Markets for a low-carbon transition, Green
Finance and Investments. OECD Publishing, Paris.
URL: http://www.oecd.org/env/mobilising-bond-markets-for-a-low-carbon-transition-
9789264272323-en.htm

OECD (2017b). Investing in Climate, Investing in Growth. OECD Publishing, Paris.

URL: https://www.oecd.org/environment/cc/g20-climate/synthesis-investing-in-climate-
investing-in-growth.pdf

Oikonomou, I., Brooks, C. & Pavelin, S. (2014). The effects of corporate social performance
on the cost of corporate debt and credit ratings. Financ. Rev. 49, 49-75.

Rosenbaum, P. R. & Rubin, D. B. (1983). The central role of the propensity score in
observational studies for causal effects. Biometrika, Vol. 70, pp. 41–55.

Rosenbaum, P. R. & Rubin, D. B. (1984). Reducing bias in observational studies using

subclassification on the propensity score. Journal of the American Statistical Association, Vol.
79, pp. 516–524.

Calder, A., Kolodzie, M. & Selot, V. (2017). Green Bonds. RBC Capital Markets. RBC Report.

Sean Kidney (2017). Address by Climate Bonds CEO Sean Kidney to Luxembourg Stock
Exchange. Climate Bonds Initiative, Luxembourg.
URL: https://www.climatebonds.net/address-climate-bonds-ceo-sean-kidney-luxembourg-
stock-exchange

Smith, J. & Todd, P. (2005). Does Matching Overcome LaLonde's Critique

of Nonexperimental Estimators?. Journal of Econometrics, 125(1-2), 305-353.

Stellner, C., Klein, C. & Zwergel, B. (2015). Corporate social responsibility and Eurozone
corporate bonds: The moderating role of country sustainability. Journal of Banking & Finance.
59, 538–549.
URL: 10.1016/j.jbankfin.2015.04.032.

TFCD (2017). The Task Force on Climate-Related Financial Disclosures

URL: https://www.fsb-tcfd.org/publications/final-recommendations-report/

25
The new climate economy (2016). The sustainable infrastructure imperative. The new climate
economy report. URL: http://newclimateeconomy.report/2016/

United Nations (1992). United Nations Framework Convention on Climate Change, article 2.
URL: https://unfccc.int/resource/docs/convkp/conveng.pdf

United Nations / Framework Convention on Climate Change (2015). Adoption of the Paris
Agreement, 21st Conference of the Parties, Paris: United Nations.

UNEP (2016). Delivering a Sustainable Financial System in India. UNEP Inquiry - Federation
of Indian Chambers of Commerce and Industry.
URL:http://unepinquiry.org/wp-
content/uploads/2016/04/Delivering_a_Sustainable_Financial_System_in_India.pdf

Wamser, G. (2014). The Impact of Thin-Capitalization Rules on External Debt Usage – A

Propensity Score Matching Approach. Oxford bulletin of economics and statistics, 76, 5 (2014)
0305–9049.

White, L. (2002). The credit rating industry: an industrial organisation analysis. In Ratings,
Rating Agencies, and the Global Financial System, ed. Levich, R. M., Reinhart, C. & Majnoni,
G. Boston: Kluwer.

Woetzel et al. (2016). Bridging global infrastructure gaps, McKinsey Global Institute.
URL: https://www.un.org/pga/71/wp-content/uploads/sites/40/2017/06/Bridging-Global-
Infrastructure-Gaps-Full-report-June-2016.pdf

26
Appendixes
Appendix a

Source: Climate Bonds Initiative, Green Bond Highlights 2017.

27
 Appendix b

Inferior bound, number of treated and number of controls

for each block

Inferior
of block Green
of pscore 0 1 Total

0 985 1 986
.00625 499 5 504
.0125 426 6 432
.025 221 3 224
.0375 183 9 192
.05 173 10 183
.075 122 27 149
.1 234 45 279
.2 88 15 103
.4 3 0 3

Total 2,934 121 3,055

28