Determinants of commercial mortgage-backed securities credit ratings: Australian evidence.

ABSTRACT. Using artificial neural networks (ANN) and ordinal regression (OR) as alternative methods to predict Commercial Mortgage-backed Securities (CMBS) credit ratings, we examine the role that various financial and industry-based variables have on CMBS credit ratings issued by Standard and Poor's from 1999-2005. Our OR results show that rating agencies use only a subset of variables they describe or indicate as important to CMBS credit rating as some of the variables they use were statistically insignificant. Overall, ANN show superior results to OR in predicting CMBS credit ratings.

KEYWORDS: Commercial mortgage-backed securities; Credit rating prediction; Ordinal regression; Artificial neural networks

1. INTRODUCTION

Commercial mortgage-backed securities (CMBSs) have expanded the investment realm of both investors and issuers. They are seen as an alternative to direct investment in property offering advantages of liquidity, diversification, and being an alternative investment to other financial investments. CMBSs are bonds backed by a single commercial mortgage or, more generally, a pool of commercial mortgages (Jacob and Fabozzi, 2003). In Australia, the expansion of the description of CMBSs as a form of securitisation of direct property assets, in addition to traditional definition of the securitisation of mortgages, has gained acceptance in the market (Jones Lang LaSalle, 2001). CMBSs also benefit from the standardised rating agency process that is directly analogous to the corporate bond markets. Corporate bond ratings inform the public of the likelihood of an investor receiving the promised principal and interest payments associated with the bond issue (Shin and Han, 2001). However, issues of proprietorship have resulted in the methodology of rating mostly being shrouded in mystery. The methods and input variables used in rating are not fully disclosed to the public (Shin and Han, 2001).

Generally, the analysis undertaken by Standard and Poor's (2001), Moody's Investors Service (2003) and Fitch Ratings (2005) in rating Australian CMBSs falls into three categories: property characteristics and cash flow analysis; portfolio level analysis; and transaction structure analysis, as elaborated in Appendix 1. The Appendix also includes factors considered and their weighting used by ABN AMRO (Roche, 2002) in ranking CMBSs. Market yields correspond to bond ratings, which indicate an association between rating and risk. The higher the credit quality the lower will be yield and the more successful will be the issue (Alles, 2000; Kose, Lynch and Puri, 2003). As such, studies of rating process are of interest not only to bond holders but also to investors.

Although bond rating agencies claim that their ratings reflect each agency's opinion about an issue's potential default risk and rely heavily on a committee's analysis of the issuer's ability and willingness to repay its debt and therefore researchers would not be able to replicate their ratings quantitatively (Kim, 2005), researchers have still gone ahead and replicated bond ratings on the premise that the financial variables extracted from public financial statements, such as financial ratios, contain a large amount of information about a company's credit risk (Huang et al., 2004). Bond rating studies have traditionally used statistical techniques such as multivariate discriminant analysis (MDA), multiple regression analysis (MRA), probit and logit models to capture and model the expertise of the bond rating process. Recently, however, a number of studies have demonstrated that artificial neural networks (ANN) can be used as an alternative methodology to bond rating.

This study investigates several aspects of the use of ANN as a tool for predicting credit ratings of Australian CMBSs. Tests are undertaken to compare the predictive power of ANN models and ordinal regression models. Maher and Sen (1997) show the following as reasons why predictability of credit rating is useful:

--It provides a firm some insight into the cost of going to the bond market to raise capital, which can be useful in comparing with other sources of funds;

--It can help investors decide where they want to place their money;

--It can provide a modified form of implicit evaluation of the firm in addition to the explicit evaluation of the bond issue; and

--An insight into factors consistent with establishing a firm's bond rating is useful in understanding the value of the firm.

Furthermore, security analysts and investors can use these ratings as the primary source of obtaining information about the quality and marketability of various issues and assess also market risk premium attached to the bonds while investment bankers use the ratings for determining commission rates on undertakings (Kim, 2005).

The paper is structured as follows. Section 2 presents an overview of the Australian CMBS market. Reviews of literature on the use of ANNs in various real estate applications and in corporate bond rating studies are presented in Section 3. Section 4 discusses the data and methodology. Empirical results and analysis are presented in Section 5. Finally, Section 6 concludes and highlights future research direction.

2. AN OVERVIEW OF THE AUSTRALIAN COMMERCIAL MORTGAGE-BACKED SECURITIES MARKET

The Australian CMBS market has undergone significant development since the first transactions came to the market in 1999, with a range of transaction types and issuers now accessing the market. The first CMBSs in Australia were done by Leda Holdings in 1999, the Longreach/Qantas head office securitisation and the David Jones flagship stores deals in 2000. As at the end of 2005 a total of 55 CMBSs had been issued with 137 tranches.

On the whole, global issuance of CMBSs has been on the increase with the USA leading the way. From 1999 to November 2005, CMBSs totalling US$532 billion had been issued in the USA compared to US$184 billion for the rest of the world during the same period as depicted in Figure 1. There has also been an increase in the financing of commercial property through capital markets. Industry data show that in 2005 issuance of commercial CMBS in the United States was around US$170 billion, an 82 per cent increase over the previous year. Strong activity is also evident in Europe, where around US$56 billion of CMBS were issued in 2005, with around three quarters of this amount issued in the United Kingdom. In 2005, A$2.29 billion of newly rated notes were issued in Australia, an increase of 8.03% on the previous year.

The total cumulative Australian and New Zealand CMBS issuance volume since 1999 had reached A$12.6 billion as shown in Figure 2 (Standard & Poor's, 2007). Total notes outstanding as at the end of 2005 was A$10.496 billion, arising from 16 credit lease and 31 CMBS transactions. Table 1 shows the number of tranches by sector issued from 1999-2005. With the overall Australian securitisation market approaching A$200 billion in debts outstanding, CMBS is still a relatively small asset class. Nevertheless, it remains both an important financing tool for commercial property owners and an alternative source of diversification for fixed income-investors.

[FIGURE 1 OMITTED]

[FIGURE 2 OMITTED]

Majority of the issues are in the single borrower multi-property category with over 95% of the total issuance to date. The CPIT 2006 Aurora Bonds CMBS is the only one single borrower single-property issuance to date. Two multi-borrower multi-property issues have been by MCS Capital Pty Limited and Challenger Capital Markets Ltd. ALE Finance Company Pty Ltd--Series 1 issuance is the only whole-business CMBS to date. The diversity of issuance transaction types show the maturity of the market as well as the arranger's confidence in trying out various CMBS structures to suit market needs.

[FIGURE 3 OMITTED]

However, as at the end of 2005 conduit-style CMBSs from large loans securitised in conduit programs which are common in the USA and Europe had not yet been undertaken in Australia. Conduit CMBSs are backed by reasonably large well diversified pools of small-to medium-sized secured property loans. A lot of the commercial mortgages continued to sit on bank balance sheets, and there was limited interest in pursuing securitisation of these assets. Since 2000, the most dominant CMBS issues have been in the office sector (A$3.6 billion), followed by the retail sector (A$2.7 billion). The diversified sector and the industrial sector have had A$2.6 billion and A$1.4 billion worth of CMBS issuance respectively. This is shown is Figure 3.

Given the general appetite for fixed-income securities and the limited supply in the market, CMBS credit spreads have been contracting as shown in Figure 4. In 2005 'AAA' five-year, interest only notes were priced at 20-25 bps (basis points) over three months' bank bill swap (BBSW), and three-year, interest-only notes at 17-20 bps over three-month BBSW. 'BBB' were priced at 60-95 bps over BBSW. These margins were lower than those of 2002, when they priced at least 20 bps wider for 'AAA' and 60 bps wider at 'BBB' level.

Figure 5 shows the top 10 Australian CMBS issuers, all of which are Listed Property Trusts (LPTs). LPTs have a 65% market share. The single-purpose-vehicle-like characteristics of LPTs have helped in their establishment as major players in the CMBS market. Between 2001 and 2004, LPTs issued CMBSs worth over $3.7B via 27 issues (eg: Mirvac, Macquarie Goodman Industrial, ING Office, ING Industrial, Investa, Macquarie Office) and bonds worth over $4.8B via 40 issues (eg: Gandel, Commonwealth Property, GPT, Stockland, Westfield) (Newell, 2005). This increased participation can partly be attributed to the high demand by institutional investors, mainly superannuation funds, for shares and bonds issued by LPTs in comparison to investing in direct property. The total contribution of asset allocation by Australian superannuation funds to property (both direct and indirect) declined from 17% in 1988 to 9% in 2000-2002, though the contribution of indirect property increased from 3% to 7% over the same period (InTech, 2003). In 2005, 95% of superannuation funds had a specific allocation to property (either direct or indirect) averaging 10% (Newell, 2006). With the drop in public bond issuance, bonds and CMBSs issued by LPT have been an attractive investment option for superannuation funds.

[FIGURE 4 OMITTED]

[FIGURE 5 OMITTED]

The macroeconomic outlook for the Australian market remains benign, with historically low unemployment rates and a low interest environment expected to continue. These stable economic conditions are expected to foster resilience in the supply of securitisable financial receivables.

3. PRIOR RESEARCH IN ARTIFICIAL NEURAL NETWORK SYSTEMS

ANNs are trainable analytical tools that attempt to mimic information processing patterns in the human brain. They are applied to a wide variety of pattern matching, classification, and prediction problems and are useful in many financial applications such as: stock price prediction, development of security trading systems, modelling foreign exchange markets, prediction of bond ratings, forecasting financial distress, and credit fraud detection and prevention. Comprehensive reviews of articles demonstrating the use of ANNs in various finance situations can be found in Fadlalla and Lin (2001), Coakley and Brown (2000), and Krishnaswamy et al. (2000).

Neural networks are regarded by many authoritative commentators as a useful addition to standard statistical techniques, and are in fact themselves based on statistical principles. Frequently these studies are in form of comparative analysis, with researchers contrasting the findings and perceived efficiency of ANNs with more tried and tested statistical methods. Although Salchenberger et al. (1992) and Tam and Kiang (1992) state that ANNs have several advantages over statistical methods, the results of these studies were less than expected because the real data in application is usually unevenly distributed among classes and these applications are limited in dealing with the ordinal nature of bond rating. Unlike statistical models, a neural network does not require priori specification of a function form, but rather attempts to learn from training input-output examples alone.

3.1. Artificial neural network systems in real estate research

ANN has recently earned a popular following amongst real estate researchers covering aspects such as real estate valuation: Tay and Ho (1991), Evans and Collins (1992), Worzala et al. (1995); Kauko (2004); examination of the impact of age on house values: Do and Grudnitski (1992), prediction of house value: McGreal et al. (1998), Nguyen and Cripps (2001) and Lai (2005); forecasting commercial property values: Connellan and James (1998a) and Connellan and James (1998b); and the impact of environmental characteristics on real estate prices: Kauko (2003).

McGreal et al. (1998), Nguyen and Cripps (2001), and Lai (2005); all demonstrated the superiority of ANN over MRA in predicting house values. Worzala et al. (1995) and Lenk et al. (1997), however, noted that ANNS where not necessarily superior. Connellan and James (1998b) also show the superiority of ANNS over MRA in predicting commercial property values.

The increased use of neural networks by academic and commercial analysts in real estate studies is motivated by their recognition of complex patterns of multivariate property data (Connellan and James, 1998a). This increased use of ANN methodology in the commercial real estate research gives credence to its extension to research in predicting CMBS bond ratings.

3.2. Artificial neural network systems in corporate bond rating research

Bond ratings are subjective opinions on the likelihood of an investor receiving the promised interest and principal payments associated with bond issues. They are published by bond rating agencies such as Moody's Investor Service, Standard and Poor's, and Fitch Ratings, in the form of a letter code, ranging from AAA-for excellent financial strength-to D for entities in default.

Rating agencies and some researchers have emphasized the importance of subjective judgement in the bond rating process and criticized the use of simple statistical models and other models derived from artificial intelligence to predict credit ratings, although they agree that such analysis provide a basic ground from judgement in general (Huang et al., 2004). Qualitative judgement, which includes accounting quality, operating efficiency, financial flexibility, industry risk, and market position, is still difficult to measure though. Literature on bond rating prediction has demonstrated that statistical models and artificial intelligence models (mainly neural networks) achieved remarkably good prediction performance and largely captured the characteristics of the bond rating process.

In this sense, various quantitative methods have been applied to bond rating. Statistical methods such as multivariate discriminant analysis (MDA), multiple regression analysis (MRA), probit and logit models have been used in order to capture and model the expertise of the bond rating process.

Several studies show that ANNs can be applied to bond rating: Dutta and Shekhar (1988), Surkan and Singleton (1990), Maher and Sen (1997), Kwon et al. (1997), Daniels and Kamp (1999), Chaveesuk et al. (1999), Yesilyaprak (2004), Huang et al. (2004), and Kim (2005).

Dutta and Shekhar (1988) were the first to investigate the ability of neural networks (NNs) to bond rating. Their sample comprised bonds issued by 47 companies randomly selected from the April 1986 issues of Value Line Index and the Standard and Poor's Bond Guide. They obtained a very high accuracy of 83.3% in discerning AA from non-AA rated bonds. However, the sample was so small that it simply amounted to showing the applicability of neural networks to bond rating.

Surkan and Singleton (1990) also investigated the bond rating abilities of neural networks and linear models. They used MDA, and found that NNs outperformed the linear model for bond rating application.

Maher and Sen (1997) compared the performance of neural networks with that of logistic regression. NN performed better than a traditional logistic regression model. The best performance of the model was 70% (42 out of 60 samples).

Kwon et al. (1997) compared the predictive performance of ordinal pairwise partitioning approach to back propagation neural networks, conventional (CNN) modelling approach and MDA. They used 2365 Korean bond-rating data and demonstrated that NNs with OPP had the highest accuracy (71-73%), followed by CNN (66-67%) and MDA (58-61%).

Chaveesuk et al. (1999) compared the predictive power of three NN paradigms- back propagation (BP), radial basis function (RBF) and learning vector quantisation (LVQ)--with logistic regression models (LRM). Bond issues of 90 companies were randomly selected from the 1997 issues listed by Standard and Poor's. LVQ (36.7%) and RBF (38.3%) had inferior results to BP (51.9%) and LRM (53.3%). BP only performed slightly better than LRM. They concluded came that assignment of bond ratings is one area that is better performed by experienced and specialised experts since neither NN nor LRM produced accurate results.

Daniels and Kamp (1999) modelled the classification of bond rating using NN with one hidden layer; and a linear model using ordinary least squares. Financial figures on bonds issued by 256 companies were selected from Standard and Poor's DataStream. The percentage of correct classification ranged from 60-76% for NN and 48-61% for OLS.

Yesilyaprak (2004) compared ANNs and MDA and multinomial logit (ML) techniques for predicting 921 bonds issued by electric utility (367), gas (259), telephone (110) and manufacturing companies (185). ANNs (57-73%) performed better than both MDA (46-67%) and ML (46-68%) in predicting the bond rating in three samples. ML (68%) performed better in predicting the bond rating (in one sample (electric utility).

Huang et al. (2004) compared back propagation neural networks and vector support machine learning techniques for bond rating in Taiwan and the United States. The data set used in this study was prepared from Standard and Poor's CompuStat financial data. They obtained a prediction accuracy of 80%.

Kim (2005) used an adaptive learning network (ALN) on a sample of 1080 observations (companies) primarily collected from the CMPUTSTAT database, Dun and Bradstreet database, and Standard and Poor's bond manuals to predict their rating. The overall performance of the model shows that the trained ALN model was successful in predicting 228 (84%) out of 272 cases. The further showed a prediction accuracy of 88% and 91% for investment grade and speculative bonds respectively.

In summary, most studies on ANNs showed promising results than those of other classification methods. The current study attempts to extend the use of ANNs to predict ratings on CMBSs. The predictive capacity of ANNs is further compared to that of OR.

4. METHODOLOGY AND DATA

4.1. Hypotheses

In this paper we hypothesise that loan-to-value ratio (LTV) is negatively related to CMBS credit rating whereas debt-to-service coverage ratio (DSCR) is positively related. The incidence of default rises with increase in LTV; that is, if all other factors are held constant, the probability of default for a loan increases as the LTV increases, but not equal. Unlike the LTV, where the probability of default increases as the LTV rises, the incidence of default is a decreasing function of the DSCR. However, the relationship between the DSCR and the probability of default is weaker than the relationship between the LTV and default. Our motivation for the specified hypothesis stems from Fabozzi and Jacob (1997) and Geltner and Miller (2001), among others, who state that LTV and DSCR are the two mostly widely used commercial mortgage underwriting criteria. Descriptions of LTV and DSCR are found in Section 4.5

We further hypothesise that CMBS issues with a well diversified portfolio both on a property composition and geographic location basis will attract higher credit ratings. The diversity of a portfolio of assets will have an impact on the volatility of the pool's expected loss. By diversifying the mix and location of property, one can mitigate a pool's expected losses. Property diversity mitigates the risk of fall in asset value of the single largest property in the pool. Geographic diversity mitigates the risk single market decline and may reduce any losses associated with this type of risk. In support of our hypotheses, Moody's Investor Service (2003) asserts that CMBS deals also benefit from portfolio diversification.

Additional hypotheses are that size of issue and note tenure are positively and negatively related to the success of bond issues respectively. Larger bond issues are done by bigger firms with strong track records who fall under stricter regulatory regimes such as the Australian Securities and Investment Commission and the Managed Investment Scheme provisions of the Corporations Act 2001, among others, should attract higher credit ratings. Longer note tenures increase the incidence of default and should therefore attract lower credit ratings.

To test the hypotheses, ordinal regressions are applied to the CMBS sample whereas prediction of accuracy in bond rating for ANN evaluates their contribution to the model.

4.2. Description of OR model

There is a general consensus on the inappropriateness of least squares methods to rate bonds as they ignore their ordinal nature (Kamstra, Kennedy and Suan, 2001). OR has been considered appropriate as it accommodates the ordinal nature of the bond rating in the analysis.

The model is similar to the general multiple linear regression model but defines Yi and estimates [beta] differently.

The logistic model computes the probabilities that an observation will fall into each of the various rating categories. The observation is classified into the category with the highest probability. This probability is estimated by the logistic model as:

[log.sub.it] ([p.sub.i]) = log[[P.sub.i]/1-[P.sub.i]] = [[beta].sub.0] + [[beta].sub.1][X.sub.i1] + [[beta].sub.2][X.sub.i2] + + ... [[beta].sub.n][X.sub.in] (1)

where: r = bond rating category; [p.sub.i] = P ([Y.sub.i] = r); i = 1 ... n, where n is the sample size; and [X.sub.i1], ..., [X.sub.in] are predictor variables.

The [beta] s are estimated by maximising the log-likelihood function:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (2)

where: [beta] is the vector of the parameters to be estimated. Once [beta]'s are estimated, [p.sub.i] is estimated by

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

The observation is assigned to the bond rating category with the highest predicted probability. These predictions are compared to the actual bond rating assigned to the issue to calculate classification accuracy for the model.

The observed value on [Y.sub.i] depends on whether or not a particular threshold has been crossed.

[Y.sub.i] = BBB if [Y.sup.*.sub.i] is [less than or equal to] [[beta].sub.1]

[Y.sub.i] = A if [[beta].sub.1] [less than or equal to] [Y.sup.*.sub.i] [less than or equal to] [[beta].sub.2]

[Y.sub.i] = AA if [[beta].sub.2] [less than or equal to] [Y.sup.*.sub.i] [less than or equal to] [[beta].sub.3]

[Y.sub.i] = AAA if [Y.sup.*.sub.i] [less than or equal to] [[beta].sub.3]

OR regressions were where carried out in SPSS[R] version 13.0 (SPSS Inc., 1968).

4.3. Description of ANN model

This subsection contains a gentle introduction to the fundamental theory of ANN. Consider the following model:

[y.sub.t] = g([x.sub.t]; [theta]) + [[epsilon].sub.t] (4)

where g(*) denotes a continuous differentiable function, [x.sub.t] is a k x 1 vector of explanatory variables, which could include the lagged dependent variables, [y.sub.t-1] for some i, [theta] is a l x 1 vector of parameter and [[epsilon].sub.t] is a sequence of independently, identically distributed random variables. In general, the explicit function form of g is unknown. However, it is possible to find a universal approximator, so that the function g can be estimated as accurately as one wish. One such approximator is

F([x.sub.t]; y) = [[phi].sub.0] + [q.summation over (i=1)] [[beta].sub.i]G(X.sub.t];[Y.sub.i]) (5)

where

G(z; v; c) = 1/1 + exp(-v[z-c]) (6)

is the well known logistic function. (Hornik, Stinchcombe and White, 1989, 1990) (see also (Cybenko, 1989), (Carroll and Dickinson, 1989), (Funahashi, 1989)) showed that for any continuous function g([x.sub.t]; 0), every compact subset K of [R.sup.k] and every [delta] > 0, there exists a F([x.sub.t]; y) such that

sup [parallel] F([x.sub.t]; [gamma])-g([x.sub.t]; [theta]) [parallel] < [delta] x[member of][KAPPA] (7)

Following these results, it is straightforward to show that the accuracy of the approximation is determined by the number of hidden layer units, namely, q and the parameter vector y, given a set of k inputs, namely, the k x 1 vector [x.sub.t]. The choice of q can be somewhat arbitrary, it is often a matter of striking a balance between accuracy and over-fitting. Given q, the parameter vector y can be estimated using non-linear least squares:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (8)

Obviously, the computational complexity of this minimisation problem grows as the number of hidden layer units grows. Several studies (See (Weeraprajak, 2007) for a comprehensive review) have suggested that the computational burden can be reduced if it is possible to separate the function F(*) into linear and non-linear components. In this case, the parameters associated with the linear component can be estimated using conventional least squares estimator, which has a closed form solution and the parameters in the non-linear component can be estimated using the nonlinear least squares estimator. This implies the number of parameters required to be estimated by the non-linear estimator is reduced and hence improve computation efficiency.

The graphical representation of the basic ANN model with the three primary components, namely the input layer (the input/explanatory variables, [x.sub.t]), the hidden layer (black box) with multiple units, G([x.sub.t], [y.sub.i]) and the output measure layer (the estimated CMBS rating in this case) can be found in Figure 6.

[FIGURE 6 OMITTED]

The hidden layer(s) contain two processes: the weighted summation function (the linear component); and the transformation function (the nonlinear component). Both of these functions relate the values from the input data (e.g. LTV; DSCR; issue size; bond tenure, property diversity, geographical diversity) to output measures (CMBS rating).

Alyuda Forecaster XL[R] (2001) was used for the ANN experimentation. In the case of our 6 input and 4 output network, the hidden units where automatically set at 29 (model 1), 28 (model 2) and 23 (model 3).

4.4. Data

Based on Standard and Poor's Ratings Direct database, our dataset comprised all the CMBSs issued between July 1999 and December 2005 totalling 55. The issues had a combined total of 137 tranches and ratings ranging from AAA, AA, A, BBB+, BBB, BBB-, to NR. In this study, all A and BBB rated tranches were grouped into two groups that is A-rated and BBB-rated respectively. The reclassification of tranches into four classes could enhance model performance because mathematical and statistical approaches have general limits in dealing with ordinal nature of bond rating. It known that as the number of bond classification increases, the predictive power could likely decrease (Kwon, Han and Lee, 1997).

We further excluded unrated tranches, to leave us with 118 tranches (training sample) and 17 tranches (test sample) respectively. Zhang et al. (1998) indicate that literature offers little guidance in selecting the training and test samples, with most authors selecting them based on the rule of 90% vs. 10%, 80% vs. 20% or 70% vs. 30%, etc. They emphasise that the critical issue is to have both the training and the test sets representative of the population or underlying mechanism. The division of training and test sets should depend on the problem characteristics, the data type and the size of the available data. Details of the individual rating categories in each sample are shown in Table 2.

Descriptive statistics of the data used in the experiments is shown in Table 3.

Appendix 2 provides bivariate training sample correlations that exist between the data items.

4.5. Selection of variables

Bond rating recognises the following areas of attention: profitability; liquidity; asset protection; indenture provisions; and quality of management. Bond rating models use independent variables, often calculated as ratios, which are predominantly derived from public financial statements. The assumption is that financial variables extracted from public financial statements, such as financial ratios, contain a large amount of information about a company's credit risk (Huang et al., 2004). Financial ratios used relate to leverage, coverage, liquidity, profitability, and size. Financial and property ratios referred to are in appendix 3. Rating agencies list qualitative factors such as management ability, value of intangible assets, financial flexibility, operating efficiency, industry risk, accounting quality and market position. However, most of these qualitative factors are likely reflected in the quantifiable data such as financial and non-financial variables, and could be assessed indirectly from analysing these quantifiable data (Kim, 2005).

According to Moody's (2003), the credit risk of CMBSs depends the characteristics of the underlying properties, loan structure, loan-to-value (LTV) ratio and the debt service coverage ratio (DSCR) and portfolio diversification. Standard and Poor's (2001) as well state that their basis of rating is the relative risk of the collateral and the ability of the collateral to generate income. The main criterion used to quickly assess credit risk of CMBS deals are the loan-to-value (LTV) ratio and the debt service coverage ratio (DSCR) (Fabozzi and Jacob, 1997). The LTV is calculated by dividing the total amount of the notes issued by the current market value of all the properties. The DSCR is calculated by dividing the total net passing income of the properties by the debt-servicing amount. The debt-servicing amount is derived by multiplying credit rating agencies' stressed interest rate assumption by the notes' issuance amount.

Credit rating agencies establish a stabilised net cash flow and an 'assessed capital value', which are used as the basis of the debt-sizing calculations. The appropriate LTV and DSCR are applied to those values. The capitalisation rate used to determine the 'assessed capital value' is a function of the risk and return of the asset, reflecting its age, quality, location, and competitive position within the market (Standard & Poor's, 2004).

Following Hedander (2005) who used a diversity scoring system based on the Herfindahl Index to measure diversity on a geographic and property type concentration basis in Australian listed property trusts, we adopt a similar procedure to measure diversity in Australian CMBS portfolios. This index effectively converts a pool of issues of uneven size into a measurement of diversity, as if all issues were the same size. A totally focussed CMBS issue has an index equal to one, while the index for a diversified CMBS issue is closer to zero. Appendix 4 shows property and geographical diversity details, among others.

The Herfindahl Geographic Region Index (HHGR) for each respective CMBS issue is calculated as follows:

HHGR = [B.summation over (j=1)][([x.sub.j]/x].sup.2] (9)

where: j = Geographic region: the states in Australia (New South Wales, Victoria, Queensland, South Australia, Western Australia, Northern Territory, Australian Capital Territory (ACT) and Tasmania); [x.sub.j] = Percentage of asset type in portfolio; x = Total portfolio composition.

We wish to acknowledge use of other factors in CMBS rating to deal with transaction and legal risk but have not considered them in this study as there are common or standard features that have been set up to mitigate these risks in all issues.

A number of models are used. Model 1 includes LTV and DSCR as independent variables. Model 2 has an addition of bond tenure and the log of issue size to the independent variables in Model 1. Finally, Model 3 has all the independent variables in Models 1 and 2 in addition to portfolio diversification variables. Tranche rating is the dependent variable in all the models.

5. EMPIRICAL RESULTS AND ANALYSIS

5.1. OR

The results of the ordinal regression analyses are shown in Table 4. To empirically specify the model, three tests were used: the standard technique of likelihood ratio test, the significance of the individual coefficients, explanatory power (pseudo R-Square) and the accuracy of the predicting rate. From the observed significance levels, only LTV is related to CMBS credit ratings being significant at .05 level of confidence in all three models but with anomalous positive coefficients implying that high LTV ratios command higher credit ratings. A negative coefficient for LTV was hypothesised as higher LTV increase the level of default and result in lower credit ratings. Log of issued amount (SIZELN) had the anticipated positive coefficient sign whereas bond tenure (TENURE) and level of property diversity (PD) had the anticipated negative coefficients. DSCR, TENURE, PD and geographic diversity (GD) appear not be related to the rating being insignificant at .05 level of confidence. This is an interesting finding as prior literature has stipulated that LTV and DSCR are the two main predictors of CMBS default risk (Fabozzi and Jacob, 1997). However, recent research by An (2006), Deng et al. (2005) and Grovenstein et al. (2004), among others, find little statistically significant relationship exists between original LTV and DSCR and CMBS default risk, supporting our results. They attribute this to the endogenous nature of original LTV and DSCR to the underwriting process. Lenders frequently respond to higher perceived overall risk (based on a multidimensional analysis including factors other than LTV and DSCR) by limiting the amount they will lend thereby lowering the loan-to-value ratio and increasing the debt service coverage ratio.

The low pseudo R-square in all three models (ranging from 0.018 to 0.039) indicate that there are other factors affecting CMBS bond rating, giving credence to use of other investigative techniques into their rating such as ANN. It should also be noted that addition of variables SIZELN and TENURE (model 2) to the basic model of DSCR and LTV increased the predictive power from 0.018 to 0.033. The full model with all the variables (model 3) showed an over double increase in the predictive power (0.018 to 0.039) over the basic model though there was a marginal increase over model 2 (0.033 to 0.039).

The inclusion of additional variables to the basic model increased chi-square from 7.036 (model 1) to 9.778 and 11.495 (model 2 and 3) respectively though significance levels decreased. Models 1 and 2 chi-square were significant at the 0.05 level and model 3 at the 0.10 level.

These results imply that rating agencies use only a subset of variables they describe or indicate as important to CMBS rating. Further, the suggested variables do not generally (with exception of LTV and to some extent DSCR) discriminate among credit ratings. This is exemplified by Figures la to 6a in Appendix 5. There is a strong relationship between CMBS rating and LTV, whereas a weak relationship exists with DSCR. The other variables show no relationship to CMBS rating.

Table 5 shows the number of ratings correctly predicted. The best results was obtained by model 3 which included all the variables at 53% (63 out of 118 cases) followed by models 1 and 2 at 52% (61 out of 118 cases) each.

The log-likelihood test in this case failed as the estimation of the general model failed to converge. Subsequently we do not believe the test is valid in this case, leading us to conclude that statistical approaches used in corporate bond rating studies have limited replication capabilities in predicting CMBS credit ratings.

5.2. ANN

5.2.1. Prediction accuracy analysis

As pointed out in section 4.5 and following the approach taken to test the explanatory power of OR models to predict credit ratings by composing models with various independent variables, the same approach was adopted using ANN. Three models were run starting with the basic model with two independent variables being LTV and DSCR. Some researchers (Fabozzi and Jacob, 1997) and rating agencies (Moody's Investor Service, 2003) regard these as the most important variables in determine a CMBS credit rating. The second model included bond tenure (TENURE) and log of issue size to the independent variables in Model 1. Finally, Model 3 had all the independent variables used in Models 1 and 2 in addition to portfolio diversity variables. Tranche rating is the dependent variable in all the models.

The predictive capacity of ANNs decreased from 93% (models 1 and 2) to 91% (model 3) for the training set and test and increased from 70% (model 1) to 80% (model 2 and 3) for the test set as shown in Table 6. Further Tables 7 shows the classification of accuracy within individual rating categories. Appendix 6 shows the error distribution.

5.2.2. Variable contribution analysis

Though earlier literature and publications by credit rating agencies state that LTV and DCSR are important property ratios which impact on the achievable credit rating for a CMBS issue, to the best of our knowledge no study has empirically examined the relative contribution of each of these input parameters to a CMBS rating. This study thus evaluates the relative importance of different factors considered in the CMBS rating using a neural network model.

The results of the relative importance of these variables in our full neural network model (model 3) are shown in Figure 7. We do not show the results of the other two models but suffice to state that the following order of importance was revealed though at various percentages: LTV, DSCR, Issued Amount and Bond Tenure.

Our study has shown 62% of CMBS rating is attributable to LTV (38.2%) and DSCR (23.6%); supporting earlier studies which have listed the two as being the most important variables in CMBS rating. The other variables contributions are: CMBS issue size 10.1%; and CMBS tenure 6.7%, geographic diversity 13.5% and property diversity 7.9% respectively.

Our results are comparable to those stated in the ABN AMRO CMBS Ranking Model. Under the model all the property-based factors added up to 75% (asset quality (15%); refinancing risk (20%); lease expiry profile (15%); credit quality of income (15%) and tenancy concentration (10%). All these factors are captured by LTV and DSCR in our model, which have a combined total weighting of 62%. In our model, diversification accounted for 21% whereas the ABN AMRO model had 15%. Differences between our model and the ABN AMRO model with the remaining factors makes difficult to complete the comparisons comprehensively. Our model captures bond tenure and amount issued. The ABN AMRO model captures management experience and growth strategy.

One drawback observable from Figure 2 is that no signs are attached to the calculated weights. Thus the interpretation of the relative weights can be inferred from OR analysis.

6. CONCLUSION, LIMITATIONS AND FUTURE DIRECTIONS

Superior predictive results were obtained from the ANN analysis in comparison to OR. ANN correctly predicted 95% and 91% CMBS rating for the training and test sets respectively whereas OR had 52-53% for the training set across the three models, confirming results obtained in earlier studies on predicting corporate bond rating using the two methodologies. Further, ANNs offer better results classifying across rating classes, while OR perform better only at the AAA class level and perform poorly for lower classes.

While our study has empirically tested variables propagated by credit rating agencies as being important to CMBS rating and found all but LTV to statistically insignificant using OR, we conclude that statistical approaches used in corporate bond rating studies have limited replication capabilities in CMBS rating and that the endogeneity arguments raise significant questions about LTV and DSCR as convenient, short-cut measures of CMBS default risk.

However, ANNs do offer promising predictive results and can be used to facilitate implementation of survey-based CMBS rating systems. This should contribute to making the CMBS rating methodology become more explicit which is advantageous in that both CMBS investors and issuers are provided with greater information and faith in the investment.

However, before these results can be generalised, field studies need to be conducted to compare the interpretation of the bond-rating process we have obtained from our models with bond-rating experts. Deeper market structure analysis is also needed to fully explain the differences we found in our models. Further still, though our results cannot be viewed as definitive due to the small sample size, the can form a basis for future studies. Over time with more CMBS issuances, a larger sample size will enable analysis of various issues backed by different property classes to check for differences, if any. APPENDIX 1. Factors considered in rating Australian CMBSs Moody's CMBS rating Standard and Poor's approach (1) CMBS rating

approach (2) * Property * Property based

characteristics analysis

analysis --Location

--Sustainable cash --Tenancy (tenant

flow profile, lease

--Quality grade maturity risk)

--Property type --Lease

--Tenant quality --Market rental rates

and expenses * Loan structure --Building quality

analysis assessment

--Amortisation profile --Supply and demand

--Floating rate loans considerations

--Seasoning and --Management

Delinquencies

--Cross- * Transaction structure

collateralisation and analysis

Cross-defaulting --Term of debt

--Other loan features --Amortisation profile * Loan-to-value and --Hedging strategy

Debt-service coverage --Cash trap

ratios analysis mechanisms

--Current, Balloon and

Target LTV

--Actual and Hurdle

DSCR * Portfolio level analysis

--Portfolio

diversification

--Other overall

considerations

(legal environment,

quality of service,

liquidity, tail periods,

commingling risk,

insurances) Fitch Ratings CMBS ABN AMRO CMBS rating approach (3) ranking model (4) * Rating analysis * Asset quality (15%) * Quantitative analysis --Location

--Adjustment to Net --Age

operating income (rent --Condition

recognition, vacancy, other --Tenant retention

income, management fee,

real estate taxes, insurance) * Refinancing risk 20%)

--Capital items consideration --Refinancing risk

(leasing costs, replacement --Ownership structure

reserves)

--Interest rate adjustment * Leasing expiry profile

(mortgage constant to (15%)

reflect long-term --Percentage of lease

conventional financing) expiring over debt term

--Debt service coverage ratio --Amount of future cash

--Loan-to-value ratio flow to amortise debt

--Amortisation credit

* Management (10%) * Qualitative analysis --Track record

--Sponsor/manager's track --Growth strategy

record

--Overleverage and Subordinate * Tenancy concentration

debt (10%)

--Collateral quality

(location, access and --Credit worthy of tenant

visibility; design and --Lease profile

construction quality;

tenant quality; economic * Number of assets in pool

and market trends; (15%)

leaseholds --Diversification

--Environmental issues --Number of assets in pool

--Pool-related adjustments

(loan and geographic

diversity) * Structural Issues

--Balloon payments

--Liquidity

--Servicer's experience

--Cross-collateralisation

and Cross-default

--Control of property cash

flow

--Property releases

--Low debt service reserve

--Management replacement

--Insurance coverage * Legal Features

--Special-purpose entity

--Representations and

Warranties APPENDIX 2. Training sample correlations Variable Issued Bond

amount tenure

(A$m) (years) Issued amount (A$m) 1.000 Bond tenure (years) 0.037 1.000 [DSCR.sup.**] 0.236 (**) 0.07 [LTV.sup.**] -0.465 (**) 0.037 Property diversity 0.025 0.108 Geographical diversity -0.089 -0.216 (*) Rating * 0.505 (**) 0.03 Variable [DSCR.sup.**] [LTV.sup.**] Issued amount (A$m) Bond tenure (years) [DSCR.sup.**] 1.000 [LTV.sup.**] -0.689 (**) 1.000 Property diversity -0.146 0.203 (*) Geographical diversity -0.042 0.073 Rating * 0.669 (**) -0.861 (**) Variable Property Geographical

diversity diversity Issued amount (A$m) Bond tenure (years) [DSCR.sup.**] [LTV.sup.**] Property diversity 1.000 Geographical diversity 0.194 (*) 1.000 Rating * -0.138 -0.063 Variable Rating * Issued amount (A$m) Bond tenure (years) [DSCR.sup.**] [LTV.sup.**] Property diversity Geographical diversity Rating * 1.000 ** Correlation is significant at the 0.01 level (2-tailed). * Correlation is significant at the 0.05 level (2-tailed). APPENDIX 3. Financial and property ratios No. Category Description Operating and

financial ratio 1 Size Tangible fixed Total assets

assets 2 Coverage Total size of debt Total debt 3 Leverage Long term capital Total debt/Total

intensiveness assets 4 Profitability Short term capital Short term debt/

intensiveness Total assets 5 Liquidity Total liquidity of Current assets/

the firm Current liabilities 6 Coverage Measure of company's Pre-tax interest

ability to pay bond expense/Income

holders 7 Indenture Subordination status (0-1)

provision 8 Efficiency Quality of Net operating

management income/Sales No. Category Property ratio Variable 1 Size Property value V 2 Coverage Debt D 3 Leverage Loan-to-value DIV 4 Profitability Break even (OE+PMT)/GI 5 Liquidity Debt service PMT/NOI

coverage 6 Coverage Interest (NOI-PMT)/NOI

coverage 7 Indenture

provision 8 Efficiency Operating NOI/GI

expenses ratio Source: Author's compilation from Belkaoui (1980); Rowland (1993) and Fischer (2004) APPENDIX 4. CMBS summary details (1999-2005) Sector Issue Issued Note Total

amount tenure lettable

(A$m) (years) area

([m.sup.2]) All

Min 0 1 49.650

Max 350 7 1.008.603

Average 75 4 349.805 Diversified

Min 1 3 97.316

Max 350 6 588.200

Average 62 4 284.666 Industrial

Min 5 1 500.844

Max 185 5 1.008.603

Average 60 3 787.841 Office

Min 10 1 49.65

Max 350 5 431.691

Average 133 3 310.142 Retail

Min 0 3 91.152

Max 240 7 533.343

Average 61 5 189.845

Property Details

Capital value Sector Issue Market S&P Capital

value stressed value All (AU$m) value discount

(AU$m) (AU$m)

Min 200 200 0

Max 1.880 1.660 22,9%

Average 760 672 11,0% Diversified

Min 265 228 7,3%

Max 1.430 1.255 20,2%

Average 688 606 12,0% Industrial

Min 454 399 3,0%

Max 1.147 885 22,9%

Average 808 701 12,2% Office

Min 495 473 4,4%

Max 1.880 1.660 16,4%

Average 1.220 1.084 10,9% Retail

Min 200 200 0%

Max 1.380 1.100 20,3%

Average 524 468 0,10

Property Details

Net income ($m) Sector Issue Market S&P net Net

Net income income

income (AU$m) sdiscount

(AU$m) (%) All

Min 18 17,90 0

Max 142,20 120,30 22,5%

Average 62,00 56,28 9,0% Diversified

Min 21,00 19,50 3,0%

Max 123,87 107,80 13,4%

Average 56,79 50,97 9,3% Industrial

Min 46,00 37,80 2,0%

Max 92,26 84,10 17,8%

Average 74,79 67,53 9,8% Office

Min 34,40 29,30 5,4%

Max 142,20 120,30 22,5%

Average 96,40 83,27 13,6% Retail

Min 17,90 17,90 0%

Max 92,80 85,40 13,9%

Average 41,76 39,06 5,9%

Financial details

Gearing Sector Issue DSCR LTV LF All

Min 12,0 32,0% 1,16%

Max 35,0 76,0% 13,3%

Average 21,4 45,1% 3,1% Diversified

Min 12,9 32,0% 1,9%

Max 35,0 68,0% 4,4%

Average 21,0 46,1% 3,2% Industrial

Min 14,6 33,0% 2,0%

Max 31,0 68,0% 3,3%

Average 24,0 42,6% 2,5% Office

Min 12,8 32,0% 1,2%

Max 24,0 62,0% 3,4%

Average 20,4 41,0% 2,2% Retail

Min 12,0 35,0% 2,0%

Max 33,0 76,0% 13,3%

Average 20,9 0,48 0,04

Tenant/Lease details Sector Issue CQI TC All

Min 0% 20%

Max 100,0% 100,0%

Average 37,5% 45,8% Diversified

Min 17,9% 42,0%

Max 56,0% 67,0%

Average 39,5% 50,9% Industrial

Min 24,2% 24,3%

Max 24,2% 25,0%

Average 24,2% 24,9% Office

Min 13,3% 39,0%

Max 75,0% 79,9%

Average 44,3% 54,2% Retail

Min 0% 20,1%

Max 100,0% 100,0%

Average 0,30 0,45

Tenant/Lease details Sector Issue WALE OR All

Min 3,6 83,0%

Max 30,0 100,0%

Average 7,8 97,2% Diversified

Min 3,6 91,3%

Max 10,0 99,0%

Average 7,1 97,0% Industrial

Min 4,1 94,0%

Max 6,3 99,0%

Average 5,4 97,6% Office

Min 4,1 83,0%

Max 8,0 99,5%

Average 5,7 96,4% Retail

Min 4,0 93,0%

Max 30,0 100,0%

Average 13,9 0,98

No. of assets

Diversity Sector Issue TA PD GD All

Min 1 8,0% 0,20

Max 101 100,0% 1,00

Average 21 29,8% 0,47 Diversified

Min 7 9,7% 0,32

Max 25 60,2% 0,51

Average 19 35,5% 0,40 Industrial

Min 26 8,0% 0,48

Max 39 14,0% 0,79

Average 34 10,2% 0,63 Office

Min 1 11,9% 0,26

Max 21 100,0% 1,00

Average 13 26,3% 0,49 Retail

Min 2 11,0% 0,20

Max 101 64,0% 0,78

Average 20 0,37 0,45 LF: Liquidity facility (% of stressed value) CQI: Credit quality of income (% of income from investment grade tenants) WALE: Weighted average lease expiry (years) OR: Occupancy rate (%) TC: Tenancy concentration (Top 5 tenants as % of total gross income) GD: Geographic diversity Herfindahl index PD: Property diversity (% of portfolio value) TA: Total number of properties Source: Author's compilation from various Standard and Poor's CMBS presale reports

APPENDIX 5. Variable scatter plots

[FIGURE 1a. OMITTED]

[FIGURE 2a. OMITTED]

[FIGURE 3a. OMITTED]

[FIGURE 4a. OMITTED]

[FIGURE 5a. OMITTED]

[FIGURE 6a. OMITTED] APPENDIX 6. ANN error distribution Model 1 Class # Cases # Errors % Errors AAA 59 4 6.78% AA 23 1 4.35% A 17 6 35.29% BBB 19 0 0.00% Total 118 11 9.32% Model 2 Class # Cases # Errors % Errors AAA 59 0 0.00% AA 23 2 8.70% A 17 6 35.29% BBB 19 1 5.26% Total 118 9 7.63% Model 3 Class # Cases # Errors % Errors AAA 59 2 3.39% AA 23 3 13.04% A 17 5 29.41% BBB 19 1 5.26% Total 118 11 9.32%

ACKNOWLEDGEMENTS

An earlier draft of this paper was presented at the 13th Pacific Rim Real Estate Society Conference in Fremantle, Western Australia, January 21-24, 2007. The authors gratefully acknowledge valuable comments from various conference participants and Professor Graeme Newell. Further gratitude goes to Standard and Poor's for giving access to their Ratings Direct database.

Received 8 August 2007; accepted 18 March 2008

REFERENCES

Alles, L. (2000) Asset securitisation in Australia: how and why it works. Curtin Business School, Perth.

Alyuda Research Inc. (2001) Forecaster XL. Fremont, CA.

An, X., Deng, Y. and Sanders, A.B. (2006) Subordination level as a predictor of credit risk. Cambridge-UNC Charlotte Symposium 12-14 June, Madingley Hall, University of Cambridge.

Belkaoui, A. (1980) Industrial bond ratings: A new look, Financial Management, 9(3), pp. 44-51.

Carroll, S.M. and Dickinson, B.W. (1989) Construction of neural nets using the radon transform. Proceedings of the IEEE Conference on Neural Networks, Washington DC, pp. 607-611.

Chaveesuk, R., Srivaree-Ratana, C. and Smith, A.E. (1999) Alternative neural network approaches to corporate bond rating, Journal of Engineering Valuation and Cost Analysis, 2(2), pp. 117-131.

Coakley, J.R. and Brown, C.E. (2000) Artificial neural networks in accounting and finance: Modelling issues, Intelligent Systems in Accounting, Finance and Management, 9(2), pp. 119-144.

Connellan, O. and James, H. (1998a) Estimated realisation price (ERP) by neural networks: Forecasting commercial property values, Journal of Property Valuation and Investment, 16(1), pp. 71-86.

Connellan, O. and James, H. (1998b) Forecasting commercial property values in the short term. RIGS Cutting Edge Conference 1998, RICS Research, Leicester.

Cybenko, G. (1989) Approximation by superpositions of a sigmoidal function, Mathematics of Control Signals and Systems, 2, pp. 303-314.

Daniels, H. and Kamp, B. (1999) Application of MLP networks to bond rating and house pricing, Neural Computing & Applications, 8(8), pp. 226-234.

Deng, Y., Quigley, J.M. and Sanders, A.B. (2005) Commercial terminations: Evidence from CMBS', Annual American Real Estate and Ur ban Economics Association (AREUEA) Meetings.

Do, A.Q. and Grudnitski, G. (1992) A neural network approach to residential property appraisal, Real Estate Appraiser, 58(3), pp. 38-45.

Dutta, S. and Shekhar, S. (1988) Bond rating: a non-conservative application of neural networks. Proceedings of the IEEE International Conference on Neural Networks, San Diego, vol. 2, pp. 443-450.

Ederington, L.H. (1985) Classification models and bond ratings, The Financial Review, 20(4), pp. 237-262.

Evans, A.H.J. and Collins, A. (1992) Artificial neural networks: An application to residential valuation in the UK, Journal of Property Valuation and Investment, 11(2), pp. 195-203.

Fabozzi, F.J. and Jacob, D.P. (1997) The Handbook of Commercial Mortgage-backed Securities, Frank Fabozzi Associates, New Hope.

Fadlalla, A. and Lin, C. (2001) An analysis of the applications of neural networks in finance, Interfaces, 31(4), pp. 112-122.

Fischer, D. (2004) Income Property Analysis. D. Fischer, Perth.

Fitch Ratings (2005) Rating Single-Borrower Commercial Mortgage Transactions. Fitch Ratings, New York.

Funahashi, K. (1989) On the approximate realization of continuous mappings by neural networks, Neural Networks, 2(3), pp. 183-192.

Geltner, D. and Miller, N.G. (2001) Commercial Real Estate Analysis and Investments. South-Western, Mason, Ohio.

Grovenstein, R.A., Harding, J.P., Sirmans, C.F., Thebpanya, S. and Turnbull, G.K. (2004) Commercial Mortgage Underwriting: How Well Do Lenders Manage Risks? School of Business, University of Connecticut, Storrs.

Hedander, J. (2005) An empirical study of listed property trusts in Australia. 12th Pacific Rim Real Estate Society (PPRES) Conference, Melbourne 23-27 January.

Hornik, K., Stinchcombe, M. and White, H. (1989) Multilayer feedforward networks are universal approximators, Neural Networks, 2(5), pp. 359-366.

Hornik, K., Stinchcombe, M. and White, H. (1990) Universal approximation of an unknown mapping and its derivatives using multilayer feedforward networks, Neural Networks, 3(5), pp. 551-576.

Huang, Z., Chen, H., Hsu, C., Chen, W. and Wu, S. (2004) Credit rating analysis with support vector machines and neural networks: A market comparative study, Decision Support Systems, 37(4), pp. 543-558.

InTech (2003) Institutional Listed Property Exposure. JB Were Investment Series.

Jacob, D.P. and Fabozzi, F.J. (2003) The impact of structuring on CMBS class performance, Journal of Portfolio Management, Special Real Estate Issue, pp. 76-86.

Jones Lang LaSalle (2001) Commercial Mortgage Backed Securities: The New Kid on the Block. Sydney.

Kamstra, M., Kennedy, P. and Suan, T.-K. (2001) Combining bond rating forecasts using logit, The Financial Review, 36(2), pp. 75-96.

Kauko, T. (2003) Residential property value and locational externalities, Journal of Property Investment & Finance, 21(3), pp. 250-270.

Kauko, T. (2004) Towards the 4th generation--an essay on innovations in residential property value modelling expertise, Journal of Property Research, 21(1), pp. 75-97.

Kim, K.S. (2005) Predicting bond ratings using publicly available information, Expert Systems with Applications, 29(1), pp. 75-81.

Kose, J., Lynch, A.W. and Puri, M. (2003) Credit ratings, collateral, and loan characteristics: Implications for yield, The Journal of Business, 76(3), pp. 371-409.

Krishnaswamy, C.R., Gilbert, E.W. and Pashley, M.M. (2000) Neural network applications in finance: A practical innovation, Journal of Financial Practice and Education, 10(1), pp. 75-84.

Kwon, Y.S., Han, I. and Lee, K.C. (1997) Ordinal pairwise partitioning (OPP) approach to neural networks training in bond rating, Intelligent Systems in Accounting, Finance and Management, 6(1), pp. 23-40.

Lai, P.P. (2005) An exploration of neural networks and its application to Taipei's housing price. 10th Asian Real Estate Society Annual Conference, Sydney, 18-21 July.

Lenk, M.M., Worzala, E.M. and Silva, A. (1997) High-tech valuation: Should artificial neural networks bypass the human valuer, Journal of Property Valuation and Investment, 15(1), pp. 8-26.

Maher, J.J. and Sen, T.K. (1997) Predicting bond ratings using neural networks: A comparison with logistic regression, Intelligent Systems in Accounting, Finance and Management, 6(1), pp. 59-72.

McGreal, S., Adair, A., McBurney, D. and Patterson, D. (1998) Neural networks: The prediction of residential values, Journal of Property Valuation and Investment, 16(1), pp. 57-70.

Moody's Investor Service (2003) CMBS: Moody's Approach to Rating Australian CMBS. Moody's Investors Service, Sydney.

Newell, G. (2005) The changing dynamics of Australian commercial property portfolios, Australian Property Journal, 38(7), pp. 553-558.

Newell, G. (2006) The significance of property in industry based superannuation funds in Australia. 12th Annual Conference of the Pacific Rim Real Estate Society, January 22 to 25, 2006, Auckland, New Zealand.

Nguyen, N. and Cripps, A. (2001) Predicting housing value: a comparison of multiple regression analysis and artificial neural networks, Journal of Real Estate Research, 22(3), pp. 313-336.

Ovnerud-Potter, P. (2003) CMBS: Moody's Approach to Rating Australian CMBS. Moody's Investors Service, Sydney.

Roche, T. (2002) Commercial mortgage-backed securities: A homogeneous asset class? Australian Property Journal, August, pp. 170-174.

Rowland, P.J. (1993) Property Investments and Their Financing. The Law Book Company Limited, Sydney.

Salchenberger, L.M., Cinar, E.M. and Lash, N.A. (1992) Neural networks: A new tool for predicting thrift failures, Decision Sciences, 23(4), pp. 899-915.

Shin, K. and Han, I. (2001) A case-based approach using inductive indexing for corporate bond rating, Decision Support Systems, 32(1), pp. 41-52.

SPSS Inc. (1968) SPSS 13.0. Chicago, IL.

Standard & Poor's (2001) Australian Commercial Mortgage-Backed Securitization--The Rating Process. Standard & Poor's, Melbourne.

Standard & Poor's (2004) CMBS Property Evaluation Criteria. Standard & Poor's, New York.

Standard & Poor's (2007) Fourth-Quarter and Year-End 2006 Australia and New Zealand Structured Finance Performance Trends. Standard & Poor's, Melbourne.

Surkan, A.J. and Singleton, J.C. (1990) Neural networks for bond rating improved by multiple hidden layers. Proceedings of the IEEE Inter national Conference on Neural Networks, Revised edn, San Diego, pp. 163-168.

Tam, K.Y. and Kiang, M.Y. (1992) Managerial applications of neural networks: The case of bank failure predictions, Management Science, 38(7), pp. 926-947.

Tay, D.P.H. and Ho, D.K.H. (1991) Artificial intelligence and the mass appraisal of residential apartments, Journal of Property Valuation and Investment, 10(2), pp. 525-540.

Weeraprajak, I. (2007) Faster Adaptive Network Based on Fuzzy Inference System. University of Canterbury, New Zealand.

Worzala, E.M., Lenk, M.M. and Silva, A. (1995) An exploration of neural networks and its application to real estate valuation, Journal of Real Estate Research, 10(2), pp. 185-202.

Yesilyaprak, A. (2004) Bond ratings with artificial neural networks and econometric models, American Business Review, 22(1), pp. 113-123.

Zhang, G.B., Patuwo, E. and Hu, M.Y. (1998) Forecasting with artificial neural networks: The state of the art, International Journal of Forecasting, 14(1), pp. 35-62.

Bwembya CHIKOLWA (1) (email) and Felix CHAN (2)

(1) School of Urban Development, Queensland University of Technology, Brisbane 4001, Queensland, Australia E-mail: bwembya.chikolwa@qut.edu.au; tel.: +61 7 3138 4072; fax: +61 7 3138 1170

(2) School of Economics and Finance, Curtin University of Technology, Perth WA 6845, Western Australia

SANTRAUKA

KOMERCINE HIPOTEKA UZTIKRINTU VERTYBINIU POPIERIU KREDITO REITINGU EMPIRINE ANALIZE: AUSTRALIJOS PAVYZDYS

Bwembya CHIKOLWA, Felix CHAN

Sisteminant komercine hipoteka uztikrintu vertybiniu popieriu prekybos sandorius, svarbiausias tikslas--gauti auksta kredito reitinga, nes tai daro poveiki pelningumui ir emitento sekmei. Kredito reitingu agenturos teigia, kad ju vertinimai isreiskia kiekvienos agenturos nuomone apie potencialia emitento nemokumo rizika ir daugiausia remiasi emitento gebejimo bei noro grazinti savo skola analize, kuria atlieka komitetas, taigi tyrinetojams ju reitingu kiekybiskai replikuoti nepavyktu. Taciau tyrinetojai replikavo obligaciju reitingus, remdamiesi prielaida, kad finansiniai koeficientai turi daug informacijos apie imones kredito rizika. Prognozuodami komercine hipoteka uztikrintul vertybiniu popieriu reitingus, kaip alternatyvius metodus naudojame dirbtinius neuroninius tinklus ir rangine regresija,. Rangines regresijos rezultatai rodo, kad reitingu agenturos naudoja tik ta kintamuju poaibi, kuriuos jos apibudina arba nurodo kaip svarbius komercine hipoteka uztikrintu vertybiniu popieriu reitingui, nes kai kurie is naudojamu kintamuju statistiskai nereiksmingi. Apskritai dirbtiniu neuroniniu tinklu rezultatai, prognozuojant komercine hipoteka uztikrintu vertybiniu popieriu reitingus, geresni nei rangines regresijos. Table 1. Number of Australian CMBS issues by sector (2000-2005) Sector 2000 2001 2002 2003 Diversi fied 1 2 11 7 Industrial 4 3 6 12 Office 0 3 4 5 Retail 0 0 15 9 Total 5 8 36 33 Sector 2004 2005 2000-2005 Diversified 7 14 42 Industrial 4 3 32 Office 9 10 31 Retail 0 8 32 Total 20 35 137 Source: Author's compilation from Standard and Poor's presale reports Table 2. Observations per CMBS rating Rating Training sample Test sample

Count Proportion Count Proportion A 17 14% 4 23% AA 25 21% 3 18% AAA 62 53% 3 18% BBB 14 12% 7 41% Total 118 100% 17 100% Table 3. Descriptive statistics Training sample

Issued Bond [DSCR.sup.**]

amount tenure

(A$m) (years) Mean 79.87 3.97 2.14 Standard error 7.36 0.12 0.05 Standard deviation 79.9 1.31 0.51 Minimum 1 1 1.28 Maximum 350 7 3.5 Test sample

Issued Bond [DSCR.sup.**]

amount tenure

(A$m) (years) Mean 47.59 4.94 1.81 Standard Error 13.33 0.06 0.09 Standard Deviation 54.96 0.24 0.36 Minimum 3 4 1.2 Maximum 190 5 2.7 Training sample

[LTV.sup.**] Property Geographical

diversity diversity Mean 0.46 0.29 0.48 Standard error 0.01 0.02 0.01 Standard deviation 0.1 0.18 0.15 Minimum 0.31 0.08 0.2 Maximum 0.76 1 1 Test sample

[LTV.sup.**] Property Geographical

diversity diversity Mean 0.48 0.32 0.51 Standard Error 0.02 0.04 0.06 Standard Deviation 0.07 0.18 0.26 Minimum 0.36 0.11 0.21 Maximum 0.61 0.55 0.78 Table 4. OR results Variable Model 1 (Expected sign) A 1.980 (0.310) [1.031] AA 3.053 (0.118) [1.952] AAA 5.515 (0.006) [2.006] DSCR (+) 0.471 (0.321) [0.983] LTV (-) 6.268 (0.011) [6.548] SIZELN (+) TENURE (-) PD (-) GD (+) Chi-Square 7.036 (0.030) * Pseudo 0.018 R-Square Variable Model 2 (Expected sign) A 3.861 (0.100) [2.700] AA 4.959 (0.035) [4.428] AAA 7.481 (0.002) [9.545] DSCR (+) 0.622 (0.207) [1.593] LTV (-) 8.307 (0.003) [9.004] SIZELN (+) 0.590 (0.122) [0.331] TENURE (-) -0.079 (0.565) [2.394] PD (-) GD (+) Chi-Square 9.778 (0.044) * Pseudo 0.033 R-Square Variable Model 3 (Expected sign) A 4.115 (0.088) [2.914] AA 5.221 (0.031) [4.664] AAA 7.757 (0.002) [9.768] DSCR (+) 0.801 (0.122) [2.393] LTV (-) 9.512 (0.001) [10.401] SIZELN (+) 0.693 (0.077) [3.130] TENURE (-) -0.087 (0.553) [0.353] PD (-) -1.255 (0.230) [1.438] GD (+) -0.949 (0.446) [0.580] Chi-Square 11.495 (0.074) * Pseudo 0.039 R-Square * We utilise McFadden's pseudo R-Square based on Ederington (1985) well as theoretically of all others. Regression coefficients provided with significance levels (in parenthesis) and Wald chi-square [in brackets]. Table 5. OR classifi cation accuracy of models 1-3 Model 1 Actual Predicted CMBS rating CMBS rating

AAA BBB Total A 17 0 17 AA 23 0 23 AAA 59 0 59 BBB 17 2 19 Total 116 2 118 Model 2 Actual Predicted CMBS rating CMBS rating

AAA BBB Total A 17 0 17 AA 23 0 23 AAA 58 1 59 BBB 16 3 19 Total 114 4 118 Model 3 Actual Predicted CMBS rating CMBS rating

AAA BBB Total A 17 0 17 AA 23 0 23 AAA 59 0 59 BBB 15 4 19 Total 114 4 118 Table 6. Summary of ANN results Model Training sample

No. of good predictions No. of bad predictions Model 1 93(95%) 5(5%) Model 2 93(95%) 5(5%) Model 3 91(93%) 7(7%) Model Test sample

No. of good predictions No. of bad predictions Model 1 14(70%) 6(30%) Model 2 16(80%) 4(20%) Model 3 16(80%) 4(20%) Table 7. ANN classification accuracy Model 1 Actual Predicted CMBS rating CMBS rating

AAA AA A BBB AAA 55 3 1 0 AA 0 22 1 0 A 1 5 11 0 BBB 0 0 0 19 Model 2 Actual Predicted CMBS rating CMBS rating

AAA AA A BBB AAA 59 0 0 0 AA 2 21 0 0 A 1 3 11 2 BBB 1 0 0 18 Model 3 Actual Predicted CMBS rating CMBS rating

AAA AA A BBB AAA 57 0 2 0 AA 1 20 2 0 A 1 3 12 1 BBB 1 0 0 18 Figure 7. CMBS rating variable contribution Geographical Diversity 13.483% Property Diversity 7.865% LTV ** 38.202% DSCR ** 23.596% Bond Tenure (Years) 6.742% Issued Amount (A$m) 10.112% Note: Table made from bar graph.