Islamic vs. conventional bond ratings: Determinants and forecastability

Transcription

1. Islamic vs . conventional bond ratings: Determinants and forecastability Sherrihan Radi a,*, Vasileios Pappas b, Antonis Alexandridis c a Kent Business School, University of Kent, Sibson, Canterbury, CT2 7PE, United Kingdom b Kent Business School, University of Kent, Sail and Colour Loft, The Historic Dockyard, Cartham, ME4 4TE United Kingdom c University of Macedonia, Thessaloniki, Greece Abstract Should Islamic bonds be rated the same way as conventional bonds? Using a sample of Malaysian bonds, we answer this question by examining the credit rating determinants of each and testing the significance of Islamic bond (sukuk) features. We lay the foundations for understanding the similarities and differences between Islamic and conventional bond ratings. Our results based on ordered probit and support vector machines show new evidence of the distinction between the two types of bonds, suggesting that their rating methodologies should differ. Sukuk and conventional bond ratings seem to share some common determinants, but their sensitivities to them vary. Moreover, our findings suggest that conventional bond ratings are driven by a smaller set of financial variables, whilst Islamic ones are proven to be triggered by a wider set of variables including Islamic structure ones. Interestingly, the findings also indicate that Malaysian CRAs tend to assign lower credit ratings to sukuk. The most accurate bond rating predictions are achieved using tailor-made individual Islamic and conventional bond rating models. The support vector machines outperform the ordered probit model across all of our samples and increases the bond rating prediction accuracy by more than 20%. Hence, to get the best results we suggest using support vector machines to forecast sukuk and conventional bond ratings, separately. These findings are potentially useful for issuers, investors and analysts that would like to predict sukuk or conventional bond credit ratings or anticipate changes in them. The results also indicate that sukuk differ ‘in practice’ from their conventional counterparts, which investors should take advantage of and explore the benefits they can obtain from them. * Corresponding author. E-mail address: s.radi@kent.ac.uk. We would like to thank Professor Roman Matousek for his valuable comments and suggestions. 1
2. Keywords : Credit rating; Bond rating prediction; Islamic bonds; Sukuk; Support vector machines; Ordered probit; 1 Introduction Corporate bond credit ratings are ordinal measures that provide key information to all capital market investors and regulators. They are often used by investors, issuers and governments as a guide to the riskiness and quality of bond issues. Credit rating agencies (CRAs) play a vital role in the financial credit markets, as their ratings are heavily used to price risky debt instruments and thus influence their marketability and effective interest rate (Amato and Furfine 2004). As a result, investors and fund managers seriously consider credit ratings when making asset allocation decisions of their investment portfolios. The importance of credit ratings is further highlighted by the fact that many regulatory bodies, including the Bank for International Settlements (BIS), base their bond investment regulation and regulatory capital requirements on ratings (Kisgen and Strahan 2010; Löffler 2004). CRAs do not disclose the exact details of their rating methodology, and emphasize that their analysts’ subjective judgement play a significant role the rating process. However, they have been hugely criticised for their lack of transparency, especially post the global financial crisis of 2008, and sluggish response to credit rating adjustments. Given the significance of credit ratings in financial markets, numerous academics study the determinants of ratings and attempt predicting the credit ratings. Despite the credit rating agencies’ claims and denials of the possibility of replicating their ratings, using a variety of methods many researchers have obtained robust results on credit rating prediction (Altman and Katz 1976; Blume, Lim and MacKinlay 1998; Huang et al. 2004; Kim 2005; Lee 2007; Kim and Ahn 2012). Recently, the sukuk1 (Islamic bond) market started to gain popularity as a Sharia-compliant alternative for the existing conventional bond markets. The Accounting and Auditing Organisation for Islamic Financial Institutions (AAOIFI) defines sukuk as “certificates of equal value representing undivided shares in the ownership of tangible assets, usufructs and services or (in the ownership of) the assets of particular projects or special investment activity” (AAOIFI 2008, p.307). Where the underlying assets must be compliant, both in nature and use, “Sukuk” is the plural form of “sakk”, the latter being the Arabic term for a legal instrument or deed. In today’s financial world it is referring to financial instruments, namely, Shariah compliant bonds issued by sovereign or corporate entities. 1 2
3. with the Shariah rules and principles . They come in different types2, known as Islamic structures e.g. Ijarah, Musharakah, Mudarabah etc. It is important to highlight that by definition sukuk are not debt certificates with guaranteed fixed income, but instead they certificates representing proportionate ownership of asset(s) or project (Godlewski, Turk-Ariss and Weill 2013; Önder 2016). Hence, sukuk holders typically have ownership rights as opposed to their conventional counterparts. According to the International Islamic Financial Market (IIFM 2018), the global sukuk issuance rapidly increased in 2017 (approximately 32% jump in volume) totalling USD 116.7 billion. This caused the outstanding sukuk issuances to reach its highest historical value of USD 434 billion, which is a clear indication of the consistent growing interest in sukuk and Islamic finance in general. As a result of the subprime mortgage crisis, sukuk became of interest not only for Muslim investors, but also non-Muslim ones that lost trust in the conventional financial markets and would like to diversify their risks or find alternatives to conventional bonds. Along with the popularity and growth of sukuk issuance, sukuk markets face certain challenges. Credit rating is considered a major challenge as it influences the marketability and risk premiums of the assets. Sukuk are typically rated by the well-known globally recognised credit rating agencies: S&P, Moody’s and Fitch, or by local CRAs. However, due to the high costs associated with external rating mechanisms, some bonds and sukuk remain unrated or unupdated (Huang et al. 2004; Arundina, Omar and Kartiwi 2015). The current literature is split into two segments with counter arguments in regards to the application of sukuk in the modern setting. Some studies including Miller, Challoner and Atta (2007) and Wilson (2008) claim that sukuk are often structured on the basis of the conventional rules of securitization, which makes them exact substitutes. In contrast, other scholars explain that sukuk do offer some form of financial innovation and therefore differ into an extent from conventional bonds (Cakir and Raei 2007; Akhtar et al. 2016). Kamali and Abdullah (2014) argue that sukuk credibility should not solely rely on the conventional concept of default, but should also incorporate religious factors and degree of compliance to Shariah. The popularity of sukuk and the on-going debate in literature on the similarity of sukuk and conventional bonds raise serious questions to whether they should be rated in using the same methodology and determinants. Hence, this paper aims to extend the current bond rating 2 AAOIFI (2008) specifies in its Shariah standard (FAS 17) 14 different types of sukuk structures, both tradable and non-tradable ones. 3
4. prediction literature by comparing sukuk credit ratings to their conventional counterparts . To do so, we investigate two main research questions. First, what are the similarities and differences between the determinants of Islamic and conventional bond ratings? Second, is a single model sufficient to predict both sukuk and conventional bonds? Hence, this empirical research examines whether sukuk differ not only in theory but also ‘in practice’ from conventional bonds. To our knowledge, it is the first study to directly contrast between the credit rating of sukuk and conventional bonds. It is potentially useful for investors and issuers that do not want to solely rely on external credit ratings or wait for their changes by the CRAs, which can instead base their investment or capital decisions ‘home made’ internal rating. Such predictions are also essential for portfolio managers that need to assess the credit risk of unrated issuers, and new issuers that seek preliminary estimate of their potential ratings prior to their entrance to the financial markets. To address our research questions, our paper follows the most recent rating prediction literature (Lee 2007; Bellotti, Matousek and Stewart 2011; Huang et al. 2004) and applies the ordered probit model and support vector machines (SVM). It is the first time that SVM is applied to the context of sukuk rating problem. Using a recent sample of Malaysian sukuk and conventional bonds we find that the credit risk of the two bond types are not driven by the exact same set of variables, and therefore present new evidence of their distinction. Whilst they share some common determinants, their sensitivities to them differ. Moreover, conventional bond ratings seem to be determined by a smaller set of financial variables, whilst Islamic ones are proven to be triggered by a wider set of variables including Islamic structure ones. Furthermore, our findings suggest that Malaysian CRAs have a tendency of assigning lower credit ratings to sukuk. In line with previous literature, the SVM method consistently achieves a higher rating prediction accuracy than the ordered probit model, making it a more suitable model for credit rating determination. Furthermore, we show that the best bond rating predictions are achieved with individual Islamic and conventional tailor-made rating models. Thus, we recommend treating sukuk and conventional bond as distinct bond types and separating them in credit rating assessments. The rest of the paper is organised as follows. Section 2 provides a brief literature review. Section 3 describes the sample data and variables, while Section 4 explains the methodologies applied. Section 5 discusses the empirical findings and implications, and Section 6 concludes. 4
5. 2 Literature review – a brief overview Credit ratings have been extensively used by investors, analysts, debt issuers and governments as a representative measure of the riskiness of companies and debt issues. Although credit rating agencies do not publically disclose the exact method they follow in assigning the ratings, many researchers found promising results on credit rating prediction. Applying various methodologies, rating models have been developed to predict bond ratings (Kamstra, Kennedy and Suan 2001; Huang et al. 2004; Kim 2005; Cao, Lim and Jingqing 2006; Lee 2007; Kim and Ahn 2012; Reusens and Croux 2017), sovereign credit ratings (Afonso, Gomes and Rother 2011; Reusens and Croux 2017), financial institution or bank ratings (Chen and Shih 2006; Van Gestel et al. 2007; Bellotti, Matousek and Stewart 2011) and other corporate issuer credit ratings (Hwang, Cheng and Lee 2009; Yeh, Lin and Hsu 2012; Mizen and Tsoukas 2012; Hwang 2013a; Hwang 2013b). The empirical credit rating prediction literature attempts to analyse how CRAs utilise public information to set the ratings. Therefore, it studies the importance of various determinants of credit ratings. Most of the studies employ firm specific financial variables in form of financial ratios, typically measuring firm size, leverage, profitability, liquidity, interest coverage and bond issue subordination status (Ederington 1985; Huang et al. 2004; Kim 2005). Other studies also utilize market variables such as market model beta of listed firms (Blume, Lim and MacKinlay 1998; Mizen and Tsoukas 2012), macroeconomic variables including change in GDP, unemployment rate and short-term interest rate (Güttler and Wahrenburg 2007; Hwang 2013a), default risk estimates from structural models and more (Pasiouras, Gaganis and Zopounidis 2006; Hwang, Chung and Chu 2010; Doumpos et al. 2015). The first study addressing the bond rating prediction problem can be traced back to the 1966, when Horrigan (1966) utilised Ordinary Least Squares (OLS) to regress accounting data for long-term credit administration. Since then, there has been an expansion of methodological approaches to corporate bond rating prediction. Generally, the empirical literature can be divided into two strands based on the prediction models applied to predict bond ratings. The first strand employs the traditional statistical methods. For instance, many studies used OLS (Pogue and Soldofsky 1969; West 1970), multiple discriminant analysis (MDA) time (Pinches and Mingo 1973; Pinches and Mingo 1975; Ang and Patel 1975; Altman and Katz 1976; Belkaoui 1980; Perry, Henderson and Cronan 1984), probit and logit models (Kaplan and Urwitz 1979; Ederington 1985; Gentry, Whitford and Newbold 1988; Blume, Lim and 5
6. MacKinlay 1998 ). Ederington (1985), compares all of the mentioned statistical approaches, his findings suggest that the logit and probit regressions outperform the rest. More recently, Kamstra, Kennedy and Suan (2001) found that a variant of the ordered logit combining method of Kamstra and Kennedy (1998) yields meaningful improvements to the predictions of bond ratings. Generally, the above listed studies employing statistical methods to address the bond rating prediction problem reached a prediction accuracy of approximately 50-70%. Despite the methods’ limitations, previous researchers have shown that relatively simple models with a small number of independent variables based on historical and public information can correctly forecast about two thirds of a sample (holdout) of bond ratings (Huang et al. 2004). Kamstra, Kennedy and Suan (2001) justify the statistical models’ lower prediction power explaining that the actual bond rating process is complex and takes into account other unmeasurable variables, for example technological changes and leadership quality. Accordingly consequent studies suggest that the statistical methods can be used as initial first estimate for the fairly multifaceted and subjective bond rating process (Huang et al. 2004). The second strand of research consists of more recent literature that seeks to improve the accuracy of bond rating forecasts using various artificial intelligence (AI) approaches. Studies applied neural networks (NN) (Dutta and Shekhar 1988; Kim 1993; Moody and Utans 1995; Singleton and Surkan 1995; Kwon, Han and Kun 1997; Maher and Sen 1997), case based reasoning (Shin and Han 1999; Kim and Han 2001), adaptive learning networks (Kim 2005) and support vector machines (SVM) (Huang et al. 2004; Cao, Lim and Jingqing 2006; Lee 2007; Kim and Ahn 2012). Typically, these papers compare the level of prediction accuracy of various AI and statistical methods. For instance, Huang et al. (2004) compare the performance of backpropagation neural networks (BNN) to SVM and Logistic regression models. They found that for both US and Taiwanese corporate bond ratings the BNN and SVM achieve better performance than the logistic regression, with a comparable prediction accuracy of approximately 80%. Using a wider set of methods, Lee (2007) obtained analogous results, though his empirical results clearly show that the SVM outperforms the rest of the methods applied (MDA, case-based reasoning and BNN). In general, it can be stated that the recent studies have shown that that AI techniques, mainly machine learning ones (NN and SVM), achieve better performance than the traditional statistical techniques. 6
7. An extensive of research has been done to predict bond ratings , yet not much attention has been paid to sukuk rating prediction. To our knowledge only four studies explore the case of sukuk credit rating (Elhaj, Muhamed and Ramli 2015; Arundina, Omar and Kartiwi 2015; Azmat, Skully and Brown 2017; Borhan and Ahmad 2018). Consistent with the conventional credit rating literature, their results indicate that sukuk ratings are also driven by firm specific variables (such as leverage, profitability and size). In addition, they found that firm’s share price, corporate governance and Islamic bond feature variables, mainly sukuk structure and Shariah advisor also play a significant role in their credit rating determination (Arundina, Omar and Kartiwi 2015; Azmat, Skully and Brown 2017). From the above studies, only Arundina, Omar and Kartiwi (2015) attempt predicting sukuk credit ratings. Using a sample of 317 Malaysian corporate sukuk issued, they compare the performance of two models: Multinomial Logistic Regression and NN. They show that the NN improves the prediction accuracy (96.18% versus 91.72% accuracy) of sukuk rating. Alongside the growing interest in sukuk, researchers have taken several directions to compare between Islamic and conventional bonds. For instance, a few studies examine the stock market reactions to the announcements Islamic and conventional bond issues (Godlewski, Turk-Ariss and Weill 2013; Alam, Hassan and Haque 2013), the impact of interest rate surprises or regional and global uncertainty factors on the returns and volatility of the two asset classes (Akhtar et al. 2017; Naifar, Mroua and Bahloul 2017). Others attempt investigating the reasons behind firms preferring the issuance of corporate sukuk opposing to conventional bonds (Azmat, Skully and Brown 2014b; Mohamed, Masih and Bacha 2015). The findings of these comparative studies are mixed, suggesting that sukuk have a lot of similarities with conventional bonds but also present several empirical evidences of their special features and distinction. These differences shed a light on the potential usefulness and significance of sukuk in portfolio management for strategic asset allocation and hedging. The studies also show that their differences vary from sukuk structure and market to another, which emphasizes the need for further empirical research in sukuk literature covering larger samples and wider aspects of their characteristics before a clear conclusion can be made. Hence, in this paper aims to extend this comparative literature by studying a different aspect of the bonds, focusing on their credit ratings. 7
8. 3 Data description and descriptive statistics The empirical analysis presented in this paper are based on data collected from the Bond Pricing Agency Malaysia Sdn Bhd (BPAM). BPAM was the primary source for all bond ratings and the majority of bond-related and market structure information. Moreover, Bloomberg was mainly used to obtain the issuers’ company financial information. Data were collected for all rated Malaysian corporate bonds active on 29th of December 2017. We chose the Malaysian market as it issues both Islamic and conventional bonds, and dominates 51% of the global sukuk outstanding (IIFM 2018). Only long-term domestic corporate bonds issued between January 2016 and December 2017 were included in the sample. The initial sample consisted 835 bonds (650 sukuk and 185 conventional bonds), which was later due to data availability reduced to 610 bonds (541 sukuk and 69 conventional bonds). Accordingly, to estimate the models we use the full sample (All), and two sub-samples: Islamic sample that includes all sukuk, and conventional sample that includes the conventional bonds. Given that Malaysia is a Muslim country, it is not a surprise that there is a big contrast amongst the subsample sizes. Malaysian firms more commonly issue sukuk that comply with their religious beliefs and are incentivised by government backing and tax advantages offered by the Malaysian government (Securities Commission Malaysia 2018). 3.1 Dependent variable The dependent variable in the bond rating models is the corporate bond credit rating assigned to the bonds by the two local Malaysian credit rating agencies MARC and RAM. In keeping with the standard practice in prior literature, only bonds rated higher than B were considered, and all ratings were categorised without considering the notches or subscripts (i.e. + and – or 1, 2 and 3). The distributions of the credit rating categories of the two sub-samples (Islamic and conventional) and the full sample are presented in Panel A of Table 1. Unlike prior literature, the most common rating assigned by the domestic CRAs is AA rather than A or lower, followed by AAA (Lee 2007; Huang et al. 2004; Kamstra, Kennedy and Suan 2001). This indicates that the Malaysian CRAs seem to be more generous in their rating assignments. This paper considers four ordinal rating categories: AAA, AA, A and BB3. In order to run the models, the rating categories were assigned to numerical values, starting with 1 to BB, 2 to A, up until 4 to AAA. Comparing the two sub-samples to each other, a significantly higher proportion of Islamic bonds (68.58%) were rated as AA, whilst a larger proportion of 3 The sample does not include any BBB bonds for which firm specific financials were available. 8
9. conventional ones (56.52% vs. 27.54%) were classified as AAA. This suggests that Islamic bonds seem to receive lower credit ratings. [Insert Table 1 around here] 3.2 Independent variables and descriptive statistics In our empirical models, we follow both the CRAs practice and bond (including Islamic bonds) rating literature in selecting the potential determinants of credit ratings included in our study. The following subsections explain all explanatory variables we include in our analysis. [Insert Table 2 around here] 3.2.1 Firm specific financial variables CRAs emphasize that the whole rating process is underpinned by the financial analysis of historical financial statements, analytic adjustments and cash flow forecasts (MARC 2016b; MARC 2017; Standard & Poor's Ratings Services 2014). Therefore, the most important set of variables consist of firm financials. The choice of financial variables included into our analysis is guided by previous literature and data availability (Kamstra, Kennedy and Suan 2001; Huang et al. 2004; Kim 2005; Lee 2007; Arundina, Omar and Kartiwi 2015). Six main types of indicators are included: firm size, profitability, leverage, liquidity, market value and cash flow. Since the rating assignments take something between 4 to 6 weeks, thus following Huang et al. (2004) the financial ratios and company financials (apart from share price) variables of the Malaysian issuing firms were collected for two quarters prior to the rating effective date (RAM 2018). To narrow down the selection of variables we input in our model, we analysed spearman’s rank order correlation matrix of the dependent (ratings) and independent variables. We prefer independent variables that have stronger correlations with the bond ratings, and exclude those that cause multi-collinearity. Finally using a backward ordered probit stepwise approach4, we include only significant variables in the model. This approach narrowed down our set of variables form 10 to 7 significant financial variables. Our final set of firm specific financial variables consists of: firm size (SIZE) measured by balance sheet total assets5, profitability given by profit margin (PM), leverage or debt to assets ratio (LEV), interest coverage 4 We set the significance levels as 11% and 10% for the variable removal and addition from the model, respectively. 5 This variable was log transformed, to ensure that larger value inputs do not distort the estimation. 9
10. (INT_COV), liquidity (LIQUID) measured by the quick ratio, and two market value measures: price-to-earnings ratio (PE) and percentage change in firm share price6 (PRICE_CHANGE). Table 2 lists and defines all of the variables included in the estimation. 3.2.2 Bond specific variables Four bond specific variables are included in the estimation samples (see Table 2). The first bond characteristic (ISLAMIC) variable indicates the principle of the bond; whether it is an Islamic (sukuk) or conventional bond. This variable will show whether the CRAs distinguish between the two types of bonds in their rating assessment. MARC (2016a) state that they analyse both internal credit enhancements (e.g. collateral value) and external credit support provided the issuer or third party. Therefore, the next two variables (GUARANTEE and SECURED) address these supports. GUARANTEE reflects whether the bond is guaranteed by a corporation, financial institution, bank or other supports, whilst SECURED refers to whether the bond is secured with some form of a collateral. Furthermore, FIXED_RATE variable specifies if the bond has a fixed rate (coupon), which limits the uncertainty of the cash stream investors will receive. Studies have shown that CRAs assess credit ratings differently, thus our last bond specific variable attempt capturing the effect of credit rating agency firms on credit quality (Cantor and Packer 1997; Livingston, Naranjo and Zhou 2007; Livingston, Naranjo and Zhou 2008). However, since in the Malaysian market bonds are either rated by MARC or RAM, but not both, therefore we incorporate a dummy variable (MARC) that distinguishes between the rating agencies. 3.2.3 Market structure variables There is evidence that, industrial sectors characterized with high-risk are expected to receive lower credit ratings i.e. adversely affect ratings, and vice versa (Mizen and Tsoukas 2012). In order to capture the industry effect on credit ratings, the sectoral classifications of bond issuers were obtained from BPAM. After filtering for missing values, the total number of sectors included in the full sample are 10. However, due to the small number of observations in certain sectors, we aggregated six sectors that included 6 bonds or less and constructed an ‘other sector’ variable (OTHERSEC). As a result the number of sector classifications is reduced from 10 to 5. Table 1 shows that most Islamic bonds are issued by infrastructure and utility, and 6 The change is calculated as the percentage difference between the share price on the rating effective date and the share price 200 days before it. 10
11. construction and engineering firms , which is typical given their specific features (structures) (Ayub 2007). On the other hand, conventional bonds seem to be mostly issued to finance property and real estate or trading and transportation services firms. Following prior studies (Mizen and Tsoukas 2012; Hwang 2013b) the industrial sectors are included as dummy variables that take a value of 1 if the bond issuer belongs to the sector and 0 if it does not. A total number of 4 dummies are included as shown in Table 2, which are benchmarked against the infrastructures and utilities industrial sector with the highest number of observations. 3.2.4 Islamic bond structure variables As mentioned earlier, sukuk come in different Islamic structures (types) with distinct characteristics, where certain structures might inherit more credit risk than others. To capture this, we following previous sukuk rating studies we include binary indicators for different sukuk structures (Borhan and Ahmad 2018; Elhaj, Muhamed and Ramli 2015; Arundina, Omar and Kartiwi 2015). From the data provided by BPAM, the sample includes eight different sukuk structures as shown in Table 1, where two of them (Wakalah and Istisna’a) have not been previously studied in prior sukuk rating literature. These variables are included only in the Islamic sample, to analyse whether and how do the sukuk structures influence the credit ratings. A total number of 7 dummies are included (see Table 2), where they are benchmarked against the Murabahah sukuk that are often referred to as Islamic debt bonds that are argued to be the closest to conventional ones (Azmat, Skully and Brown 2014a). From Table 1, it is evident that the most common sukuk structure included in our sample are hybrid (mix of two or more structures) and Mudarabah sukuk. In contrast, the least common contractual agreement is Bai Bithaman Ajil (BBA), which is due to Shariah principles restricted from secondary market trading. 3.2.5 Descriptive statistics The descriptive statistics of the selected financial variables from our Islamic, conventional, and full (combined) samples are reported in Table 3. The table also reports the t-tests (MannWhitney test) on the mean (median) difference of the Islamic and conventional bond issuers’ financials. [Insert Table 3 around here] Most of the statistics are in line with previous literature. Observing the averages of the size factor, we verify that Islamic firms that issue sukuk are significantly smaller than conventional 11
12. firms or conventional bond issuers (total assets of RM 47.98 bn against RM 208.16 bn). This is consistent with the findings of Mohamed, Masih and Bacha (2015) and Grassa and Miniaoui (2018), as well as comparative studies of Islamic and conventional banks (Olson and Zoubi 2008; Pappas et al. 2017). Moreover, the first profitability measure, profit margin, show that sukuk issuers are also less profitable than conventional bond issuers (15.59% and 19.16%), which is supports the view that firms with lower profit expectations seek resort to sukuk (mainly partnership based structures) issuance to be able to share the losses in the event of default (Godlewski, Turk-Ariss and Weill 2013; Alam, Hassan and Haque 2013). Nevertheless, investors seem to be slightly more optimistic about the Malaysian Islamic firms’ (sukuk issuers’) future prospects indicated by their higher price-to-earnings ratio. In terms of leverage, Table 3 indicates that the Islamic sample firms have higher debt to asset ratios (37.52% versus 23.82%). However, it is noteworthy to mention that this ratio does not take into account the total amount of liabilities firms owe. In fact, when taking into account the total amount of liabilities, the average leverage ratio of Islamic bond issuers become lower than the conventional ones (61.51% against 73.74%), and accordingly they are also more capitalised (38.49% versus 26.26%). These results support the findings of other studies that claim Islamic banks have higher capitalization (see e.g. Beck, Demirgüç-Kunt and Merrouche 2013). Despite the Islamic sample firm’s lower profitability, their lower liability levels (and this lower interest expenses) allow them to have greater ability to cover their interest payments (15.55 times against 8.75 times). Furthermore, Islamic banking researches prove that Islamic institutions are characterised by higher levels of liquidity. Our descriptive statistics for quick ratio (LIQUID) show consistent evidence, where Islamic bond issuers have a liquidity ratio exceeding 1, opposing to the conventional bond issuers which have a ratio below one (1.12 versus 0.97). This indicates that the Islamic sample firms have a sufficient proportion of liquid assets to instantly cover their current liabilities. Beck, Demirgüç-Kunt and Merrouche (2013) argue that the higher liquidity reserves and capitalisation levels of Islamic banks could potentially explain their reasonably better performance during the subprime mortgage crisis. Lastly, the summary statistics show no significant difference in the growth of share prices of Islamic and conventional firms. All outliers observed in the full sample (e.g. PM and PE) have been treated by winzorising at the 1st and 99th percentile. 12
13. 4 Methodology We use two methods for predicting the bond credit ratings: ordered probit (statistical method) and support vector machines (machine learning method). We perform two types of predictions, in-sample and out-of-sample predictions. For the last section of analysis 5.2.1, we split the datatset into two subsets: a training set 70% (Islamic: 378, conventional: 48 and All7: 427) and a holdout set 30% (Islamic: 163, conventional: 21 and all: 183) of the total sample (610), respectively. The split was done in a way to keep the original proportions of the rating classes (AAA to B, see Table 1) in each of the subsets. Moreover, for the support vector machine estimation and prediction, we scaled the firm specific financial independent variables to the range [-1, 1] to avoid having certain values dominating others or any computational difficulties. The same scaling method was used for both subsets. 4.1 Ordered probit We estimate an ordered probit (OP) model which is appropriate for ordinal dependent variables, such as credit ratings (Kamstra, Kennedy and Suan 2001; Wooldridge 2010). The model assumes