Copenhagen Business School Master Thesis 15 Sept 2021
commercial paper and corporate bonds is that the former is typically used for raising short-term funding requirements such as working capital.
Another relevant condition is that the Sukuks bonds are excluded, i.e., specially tailored financial products that conform to the set of ethical and common law– based financial transaction principles laid out in Shari’ah, or Islamic law (Safari, Ariff, & Mohamad, 2014).
An important condition regards the choice of timeline. In fact, the issue date for a bond in the sample selection is set between the 1/1/2012 and 31/12/2020. We chose to start in 2012 as the first ever corporate green bonds have been issued for the first time in 2012.
Since then, the market has exploded, from$10 billion in 2013 to over $40 billion in 2015 and is projected to exceed $100 billion in 2016 (Rosembuj & Bottio, 2016). We ended our timeline in 2020 as the post-issuance performance of corporations is measured in the period following the bond issuance.
The final constraints consist in setting the SIC codes equal to the industries that we wanted to analyse. A list of 32 different SIC codes is presented in 10.1. The result of such research is a dataset of 10895 unique rows called “Universe of bonds”. Each row is a bond issuance which is described by 121 different variables. Summary statistics for the
“Universe of bonds”.
corporate attribute hereafter.
Hereupon, a step-by-step description of the how the analysis will be conducted is out-lined. First of all, I filter the universe of 10896 rows to select only green bonds, which account for 509. Each of the bonds has an issuer who is attributed to a unique code iden-tifier. Eventually, the subset of 509 green bonds determines a list of 155 unique firm-code identifiers.
On the 6th of May 2021, I operated the second data collection for the universe of compa-rable vanilla bonds, using the list of 155 unique issuers as a selection filter. The resulting dataset contains 1502 potentially analogous vanilla bonds, hence, each of the green bonds has about 3 potentially comparable vanilla bonds to be matched with.
Moving to the data manipulation, it has been realized in the working environment of RStudio. There, it is possible to read xlsx files and manipulate the data. The xlsx file containing only green bonds was named Green Universe and consists of 509 observations, whereas the one containing only vanilla bonds is named Vanilla Universe and contains 1502 observations. Both data sets have 121 different character variables. Out of 121 char-acteristics, I have selected the 10, e.g., (SIC, Ticker, Maturity, ‘Coupon‘, ‘Coupon Type‘,
‘Issue Date‘, ‘Yield to Maturity‘, ‘Amount Issued (USD)‘, ‘Green Bond‘, ‘Country of Is-sue‘). The data sets are merged into one and further data manipulation where conducted, e.g., creating a dummy variable for the green bond attribute or mutating the nature of some variables from text into numerical. Also, the data cleaning consists in omitting the missing value, decreasing the data population of about 20%.
To standardize the sample and have a better comparison, I require the bond type to be only “Fixed Income” as opposed to “Variable income”. Indeed, the latter have a return based on some underlying benchmark measure such as short-term interest rates. This process decreased the observations’ number from 1627 to 1464.
Finally, two last manipulations are performed. Firstly, the variable “yield to maturity”
is investigated to find possible outlier. To mitigate the impact of outliers, the variable is winsorized at the 1st and 99th percentiles of its empirical distribution. Secondly, I take the logarithm transformation of the “amount issued in US dollars millions” variable to make it as “normal” as possible so that the statistical analysis results become more valid, as log-transformation reduces skewness from the data.
Initial imbalance
Copenhagen Business School Master Thesis 15 Sept 2021
To conduct the pricing analysis, i.e., comparing the yield at issuance of a green and a corresponding vanilla bond, each green bond needs to be matched to a comparable vanilla bond. Prior to matching, it can be a good idea to view the initial imbalance in one’s data that matching is attempting to eliminate, by looking at Table 6.2 .
Control Treated
All 1217 247
Table (6.1) The table above shows the number of green bond (treated) and the vanilla bond observations (control). The ratio is almost one green bond observation every five vanilla bonds, i.e., each treated has five potential matching.
No matching P4
j=1|SM Dj| 2,681 P4
j=1E(eCDFj) 0,571
Table (6.2) We can see severe imbalances as measured by the standardized mean differ-ences (Std. Mean Diff.), variance ratios (Var. Ratio), and empirical cumulative density function (eCDF) statistics. Values of standardized mean differences and eCDF statistics close to zero and values of variance ratios close to one indicate good balance, and here many of them are far from their ideal values.
After assessing the initial imbalance, the actual matching algorithm is run. It runs by means of the R function “matchit” which comes with the R package called MatchIt (Imai et al., 2021). Although MatchIt can perform matching based on prosperity score or a va-riety of other matching procedures, the intent of this analysis is to match the two nearest neighbors, i.e., finding those subjects whose co-variance matrix are the most similar (Daniel HoKosuke Imai et al., 2021). A matching algorithm, that simultaneously minimizes dis-tances across the four matching covariates (Fr´esard & Valta, 2016), is implemented by means of MatchIt. The formula argument of matchit is
Green Bond=M aturity+Coupon+Issue Date+Log(amount issued in U SD) (6.1) whereGreen Bond is the binary treatment indicator and the rest of the included variables the pre-treatment covariates. The “method” argument specifies a matching method. I opted for the genetic matching method. Such a method is described in the matching framework in the literature review section. Furthermore, the “distance” argument specifies the method used to estimate the distance measure. Although a variety of distances are
Page 68 of 198
available, the Mahalanobis distance is preferred3.