Predicting Credit Rating Migration Consistently with Macroeconomic Conditions

(1)

Predicting Credit Rating Migration

Consistently with Macroeconomic Conditions

Master Thesis Semester 2.2 2011 MSc IE&B Supervisor: Robert Inklaar s1622358

(2)

Abstract

Credit risk has gained interest in the past decades, and particularly so in the past few years as previously steady global economic growth has become more turbulent. An important input in the evaluation of credit risk for a portfolio of assets is the credit rating of an asset. This study investigates whether macroeconomic variables drive transitions in credit ratings for the rated US Corporate Bond universe. To this end, credit rating migration is measured as a change in rating-based capital requirements. This is a highly innovative approach allowing for general-to-specific modeling. Four macroeconomic variables have been identified having a significant impact on credit rating migration: GDP, Yen/US$ exchange rate, housing price index, and short-term US$ interest rate.

(3)

Introduction

Credit Risk research has gained importance throughout the last 20 years, and exponentially so in the previous decade (Koopman and Lucas 2003). Early asset-level credit risk research was rooted in the Black-Scholes option-pricing model, presented by Robert C. Merton in 1974 (Merton 1974). Adaptations of the model followed, such as those by Vasicek (1987, 1991, 2002) and the Jarrow–Turnbull model (1995). Improvements in data availability and in statistical software capabilities have resulted in further publications on the subject of credit risk, such as by Caouette, Altman and Narayanan (1998) and Allen and Saunders (2003). Recently, interest has peaked due to changes in regulations for banks and insurance companies and the current subprime mortgage crisis (Qu 2008). This paper will contribute to the growing research on portfolio level credit risk, by developing a relationship between credit risk rating migration and multiple macroeconomic variables.

An increasing amount of credit risk research is dedicated to the study of credit risk ratings. Because assessing portfolio-level risk is highly complex, using credit risk ratings as an assessment of the amount of credit risk in an individual asset is appealing (Trück and Rachev 2005). Moreover, regulators of financial institutions are basing capital requirements on credit ratings. For a financial institution, the relationship between available capital and (rating based) capital requirements drives the financial stability of the company, as indicated by the company’s own debt rating. Credit rating migration in a portfolio of assets results in different, most often higher, capital requirements for the same portfolio of (rated) assets over time. To manage and understand its own credit quality, banks and insurance companies need to include the potential effects of credit rating migration in their financial planning exercises.

Macroeconomic variables are expected to impact credit rating migration patterns, which in turn impact the rating based capital requirements of a financial institution. This study establishes clear, statistically significant, relationships between changes in several macroeconomic variables and changes in rating based capital requirements. Because the same macroeconomic variables might also impact the company’s income and cost measures, the results of this study will help financial institutions in assessing the impact of the macro economy on its overall financial position.

The research question can be formulated as follows:

(6)

The results might be helpful for financial institutions to better plan their future capital requirements in line with their expectations for the macro economy. The relationships found could also be used by financial institutions for scenario analysis and stress testing. In addition, this paper adds to the existing knowledge base regarding the effects of individual macroeconomic variables on credit rating migration. Previous research on the effect of the macro economy on credit rating migration has focused on single variables that are supposed to encompass the entire macroeconomic situation. This is because credit ratings are taken as categorical variables. This study applies an innovative approach in which credit rating migration is modeled as a change in rating-based capital requirements, allowing for working with a continuous variable. This distinguishes this study from all previous studies on rating migration. The approach allows for general-to-specific modeling to be performed and reveals multiple statistically significant macroeconomic variables that impact credit rating migration patterns.

The structure of this paper will progress as follows:

 First, existing literature will be reviewed and summarized. At the end of the literature review an explanation of how other researchers investigated similar research questions will be given.

 Next, a hypothesis will be presented.

 After this, the research methodology applied, including an explanation of general-to-specific modeling, will be described.

 The model and all variables used in the analysis will be presented.

 Then the results and its analysis will be presented for regressions of individual rating categories, a portfolio of assets, and a panel dataset.

 A ‘specific’ model is presented and evaluated.

(7)

Literature Review

As mentioned, this paper is concerned with credit rating migrations and how they are affected by the macro economy. In order to review what is already known on this subject it is useful to define a few key terms. Credit risk, credit risk ratings and credit risk rating migration, are explained below.

Credit risk is also referred to as default risk and is defined as the risk of loss arising due to the borrowers failing to meet payment obligations. Probability of default is the predominant variable used to approximate credit risk (Qu 2008).

Credit risk ratings are unique for each credit rating agency and are not absolute. Standard and Poor’s website defines credit risk ratings as follows (Source: www.standardandpoors.com):

“Credit ratings are forward-looking opinions about credit risk. Standard & Poor’s credit ratings express the agency’s opinion about the ability and willingness of an issuer, such as a corporation or state or city government, to meet its financial obligations in full and on time.

Credit ratings can also speak to the credit quality of an individual debt issue, such as a corporate note, a municipal bond or a mortgage-backed security, and the relative likelihood that the issue may default.”

The first credit ratings were given in 1914 by Moody’s, which is one of the four major US rating agencies. Poor’s was second to develop credit ratings. Poor’s is now Standard and Poor’s, one of the four major rating agencies in the US (Altman and Kao 1992). Credit ratings have come a long way since then and are a key input in the new generations of credit risk models (Nickell et al 2000). The reason they have become important is due to the general difficulty encountered in assessing risk and the straightforwardness of using ratings (Trück and Rachev 2005).

Credit risk rating migration is simply the transition of credit ratings from one level to another. As mentioned previously credit ratings represent rating agencies’ opinion on the issuers. Rating agencies periodically review their ratings and if necessary, change their rating dependent on their assessment of whether the credit quality has improved or deteriorated (Altman and Kao 1992).

(8)

of the credit portfolios, according to Trück and Rachev (2005), there seems to be a lack of literature on modeling the relationship. This paper will aid in predicting credit rating migration and will thus aid in predicting value changes of investment portfolios.

Better predictability of credit rating migration and therefore also the value changes of investment portfolios are vital for the management of financial institutions. It can help them attain their individual targets as well as those determined by regulators. An example of an individual target is that the investment-grade bond funds held by financial institutions are not permitted to hold bonds below investment grade (Altman and Kao 1992). Bond values may, however, deteriorate due to shifts in the macro environment, if these deteriorations are anticipated; targets of the funds can be adjusted preemptively. As regulators also set targets for financial institutions, changes in regulation are important for financial institutions. These changes in regulation have led to an increased interest in credit rating migration.

Perhaps the most influential regulatory organization is the Basel Committee on Bank Supervision. The Basel committee does not regulate banks directly, but it sets international guidelines for banking supervision. It is upon these guidelines that national/local regulators associated with the Basel Committee base their regulations. In their new proposals of the Basel Capital Accord, banks are recommended to base their capital requirements on the credit risk in their loan portfolios (Basel Committee on Banking Supervision 2003). Capital requirements can also be referred to as capital reserves that are held to cover losses due to defaults of counterparties. Assessing this credit risk has to be done either using ratings from external rating agencies or internal ratings (Qu 2005, Trück and Rachev 2005, Koopman and Lucas 2003). A financial institution thus has a loan portfolio with a number of assets that all have their own rating. For each rating there is a different capital requirement that is mandated by a regulator. Aggregated, these requirements total the capital reserves that the financial institution has to hold. In the event that a credit rating migration occurs, some bonds will attain a new rating and thus also have a different capital requirement. The sum of these changes will affect the capital reserves that the financial institution is required to hold.

An influential regulatory institution, whose rating-based credit requirements are used later in this paper to quantify credit risk migration, is the National Association of Insurance Commissioners (NAIC). The NAIC is concerned with regulating insurance companies in the United States (Altman Kao 1992). The focus of regulatory institutions on credit ratings and rating migration naturally raises the interest of financial institution to better model them. This paper will help accomplish this.

(9)

Lonski (1992), Carty and Fons (1993) and Carty (1997) (Nickell et al 2000). Others have been done by academics some of which will be mentioned below.

Altman and Kao examined the aging effect in 1992; they find that as time increases since the issuance of the first rating, the propensity of ratings to change increases. They also uncover that for the majority of rating categories there is a greater propensity for downgrades than upgrades. Another study that looks into this issue and has similar findings was done by Helwege and Kleiman in 1996.

Nickell et al (2000) investigated industry effects on credit rating migration. They compare industrial firms to banking firms. They find some statistically significant differences but as far as recognizing a clear trend is concerned, their findings are inconclusive. The difficulty arises predominantly from comparing industries that have different average ratings. It is difficult to compare rating migration behavior if they did not have the same ratings to start with. Qu (2008) investigates the effect of different industries on expected default rates. He also finds varying statistically significant effects across industries but again with no clear trend. This subject warrants further research but goes beyond the scope of this paper.

(10)

How did other studies investigate the relationship between macroeconomic conditions and credit rating migrations?

Other studies have investigated the effect of the macro economy on credit rating migrations by working with transition matrices. A transition matrix has rows and columns with all of the individual rating categories including the default category (D) in an additional column. Matrices are measure transitions over a fixed time span for a specific time frame. The cells in the matrix present the percentage of bonds that have migrated from the original rating shown in the row to a new rating shown in the column. The majority of bonds retain their original rating but a small percentage is up or downgraded as can be seen in an example of a transition matrix presented below in Figure 1.

Figure 1: Transition Matrix 2010 Quarter 4

Rating Issuers AAA AA A BBB BB B CCC CC C D

AAA 35 88,57 11,43 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 AA 127 0,00 96,85 3,15 0,00 0,00 0,00 0,00 0,00 0,00 0,00 A 540 0,00 0,56 98,33 1,11 0,00 0,00 0,00 0,00 0,00 0,00 BBB 717 0,00 0,00 0,84 98,74 0,28 0,14 0,00 0,00 0,00 0,00 BB 471 0,00 0,00 0,00 1,70 96,60 1,70 0,00 0,00 0,00 0,00 B 843 0,00 0,00 0,12 0,00 1,42 97,39 0,95 0,00 0,00 0,12 CCC 91 0,00 0,00 0,00 0,00 0,00 8,79 81,32 3,30 0,00 6,59 CC 9 0,00 0,00 0,00 0,00 0,00 0,00 0,00 55,56 0,00 44,44 C 0 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00

Frequently, transition matrices are conditioned on the state of the macro economy. What this means is that two or more matrices are estimated controlling for in this case the state of the economy and then compared. The state of the macro economy is represented by a single categorical variable. The economy was either in an expansion or in a contraction state in Bangia et al (2002). Or as in Nickell et al (2000) an additional ‘normal’ state was included. In Nickell et al (2000) the categories reflecting the macroeconomic state of the economy is determined by GPD. GDP is a strong indicator of the macroeconomic condition but does not capture everything and thus may be deemed an oversimplification. Bangia et al (2002) used a broader indicator of the state of the macro economy to make their classifications. They used classifications created by the National Bureau of Economic Research (NBER). These classifications are based on broader set of variables and thus should capture the overall state of the macro economy better than just GDP but the overall method is still an oversimplification. In the investigation by Bangia et al, there are only two categories that determine the state of the macro economy. Therefore it is impossible to distinguish between the effects of individual variables that reflect the state of the macro economy on rating migrations.

(11)

(12)

Hypothesis

In line with intuition and with previous studies I predict that, as the state of the macro economy worsens, the occurrence of downward credit rating transitions will increase. In other words, if macroeconomic conditions deteriorate, so will the height of credit ratings. The opposite is, of course, also expected; when the macroeconomic conditions improve, so will the height of credit ratings.

This predicted relationship is depicted below:

Figure 2: Hypothesis

(13)

Methodology

This study will differentiate itself from previous studies in two major ways. The single independent variable approach described earlier will be swapped for an approach with multiple independent variables being tested simultaneously. In total, ten macroeconomic variables will be tested. These ten variables are expected to jointly capture the general state of the macroeconomic conditions. The second major distinguishing aspect is that quantitative effects of these macroeconomic variables will be measures, which was not clearly captured in the other studies.

The macro economy is very complex and is constantly evolving, and therefore, difficult to model. This is why in this study an attempt is made to model it by using general-to-specific modeling. General-to-specific modeling is also known as the London School of Economics approach to econometric modeling. The central idea of the general-to-specific modeling is to start with a wide or ‘general’ model with many variables and to remove variables that are not statistically significant. This results in a narrower or more ‘specific’ model. This type of modeling is advocated by and associated with David Hendry (Owen 2002). Roots can however be traced further back than Hendry’s work; namely to that of Denis Sargan in 1964 (Sargan 1964). Two other noteworthy authors that have spurred the use of general-to-specific modeling are Hoover and Perez (Owen 2002).

The strength of general-to-specific modeling is that it is useful in estimating models where theory cannot predict a complete/perfect model. In economics, as in most sciences, this is almost always the case, due to the nature of what is being modeled (Campos et al 2005). The macro economy is an example of something that can never be perfectly modeled, as it is dynamic and constantly evolving and is affected by millions of factors. Empirical modeling, by means of general-to-specific modeling is useful as there are multiple factors that are contenders for being regressed as macroeconomic condition encompassing variables.

(14)

In the paper by Campos et al (2005), flaws of the conventional method of relying almost exclusively on theory are outlined. The first most notable flaw being that one must already know what relationships exist before performing statistical analysis, meaning nothing new can be discovered. Second being that there are multiple economic theories, which are at times contradicting, thus even models that are strongly supported by theory are disputable. More progressive research such as that relying on general-to-specific modeling is more likely to discover new relationships (Hendry 1995). It appears as if conventional and empirical researches are in direct contrast to each other. In actual fact, they are part of the same empirical cycle (de Groot 1961). In conventional research, you first develop theories and test them and then possibly go back to develop theories. In empirical research, you start by testing data and then develop theories and then can go back to testing.

There are multiple methods of performing general-to-specific modeling and due to current technological capabilities there are highly complex algorithms that can be used. In this paper, STATA 10C will be used to run regressions and also to remove insignificant variables. This will be done by running stepwise ordinary least squares (OLS) regressions with backwards elimination. What this means is that originally, all independent variables are included in the model and then those that are not statistically significant will be removed. The alternative would be to start with an empty model and to only add those variables that are significant.

Applying general-to-specific modeling by means of running stepwise OLS regressions is unique for studying credit rating migration behavior. As mentioned earlier, other studies have been subjected to using ordered probit and logit models. This is because they have chosen to model rating categories as categorical variables. In this case this is appropriate, because as the definition of credit ratings provided by Standard and Poor’s dictates, credit ratings express opinions and are not in fixed relation to each other. There is nothing wrong in applying ordered probit and logit models. They are the most appropriate method when rating categories are considered to be categorical variables. As will be demonstrated, however, research relying on these methods is only able to scratch the surface of how the macro economy affects credit rating migration. This is due to the ordered probit and logit models being more ‘simple,’ as they are not able to predict quantitative impacts of individual macroeconomic variables. This simplicity can also, however, be seen as its strength.

(15)

Data Description

Sample

The sample for this investigation is the United States and tax havens from 1981 until 2010. Quarterly transition matrices are obtained from Standard & Poor's CreditPro® - Corporate Ratings, which summarizes the migration of the number of issuers from one class to another. Standard and Poor’s has rated 8,779 firms that are included in this sample during this time period.

The reason that this investigation is focused on just the United States and tax havens is due to the data availability and the uniformity within the sample. Data is ample for the US, making it a prime candidate for investigation. Uniformity is important for this research because macroeconomic effects are potentially different for firms based in different countries. Potentially this investigation can be expanded to include a broader sample or be conducted for other countries or economic regions.

Tax Havens

In determining the sample, the issue arises as to what should be done with US firms that are based in tax havens like Bermuda and the Cayman Islands. These firms could either be excluded because they are not based in the US, or could be included. US firms based outside of the US are generally excluded because they are active in a different market and are therefore subject to different macroeconomic conditions. Tax havens are however, included in the sample because they do not have a substantive local presence in the country that they are based in. These firms are based in tax havens for tax reasons but are actually active in the US and are therefore exposed to the same macroeconomic conditions as the rest of the sample. In total, 181 of the 8,779 rated bonds included in the investigation originate from tax havens.

Withdrawn Ratings

(16)

Time Period

The time period is quite extensive, spanning 30 years. Data extending back from 1981 until 2011 is the maximum data that is available. The data included is recent; this differentiates this study from previous studies. The inclusion of recent data is particularly interesting due to the economic recession in the US that started in 2007. A longer time span is important for investigating cyclic patterns as it includes a greater number of business cycles. Therefore, observations for the effect of macroeconomic conditions are more frequent allowing for better generalization (Koopman and Lucas 2003).

Quarterly data is collected and used as it allows for greater degrees of freedom as opposed to the annual data, giving 120 measurements as opposed to 30. The quarterly data is however converted to annual changes because annual changes are more appropriate. These annual changes are composed of quarter-to-quarter changes therefore high degrees of freedom are retained due to the 117 measurements.

Variables

Endogenous Variable – Credit Rating Migration (CRM)

The endogenous variable is the credit rating migration (CRM). It has to be constructed in order to test multiple exogenous macroeconomic variables simultaneously and be able to run in a stepwise OLS regression. An exact description of the procedure is explained below. But first, a more detailed explanation of credit ratings, specifically those of Standard & Poor’s will be provided, and then Rating-based Capital Requirements will be better explained.

Credit ratings

Credit ratings express an opinion and are explicitly non-quantitative. In this study ratings are obtained from Standard and Poor’s, which is one of the major four independent credit rating agencies in the United States. Ratings from Standard and Poor’s are represented by letters, AAA being the highest rating and D being the lowest. It should be noted that rating categories AAA-BBB are commonly referred to as Investment Grade (IG) whereas the remaining rating categories BB-D are referred to as Below Investment Grade (BIG). The following overview from the Standard and Poor’s homepage clarifies the individual ratings:

‘AAA’—Extremely strong capacity to meet financial commitments. Highest Rating. ‘AA’—Very strong capacity to meet financial commitments.

‘A’—Strong capacity to meet financial commitments, but somewhat susceptible to adverse economic conditions and changes in circumstances.

‘BBB’—Adequate capacity to meet financial commitments, but more subject to adverse economic conditions.

(17)

business, financial and economic conditions.

‘B’—More vulnerable to adverse business, financial and economic conditions but currently has the capacity to meet financial commitments.

‘CCC’—Currently vulnerable and dependent on favorable business, financial and economic conditions to meet financial commitments.

‘CC’—Currently highly vulnerable.

‘C’—Currently highly vulnerable obligations and other defined circumstances. ‘D’—Payment default on financial commitments.

In this paper coarser data is opted for and only the major rating categories are used and not those created by an additional plus or minus sign. This is done in line with Nickell et al (2000) who claim that the additional categories add complexity, are less reliable and should therefore not be used. It would also lead to more volatility due to lack of data. Bonds with an additional plus or minus sign are not excluded from the investigation but are merely pooled in their respective major rating categories.

What is rated?

In this study, ratings are based on senior unsecured bonds, i.e., bonds that are not backed by collateral but have priority over other unsecured bonds in recovering funds should the firm go bankrupt. Most studies on ratings transitions use these types of bonds as a standard: Lucas and Lonski (1992), Carty and Fons (1993) and Carty (1997) (Nickell et al 2000). Presumably if an obligor has not issued senior unsecured bonds, ratings are derived by Standard and Poor’s from other types of bonds issued. This paper will use senior unsecured bond ratings, as this will allow results to be comparable to other studies. Rating-Based Capital Requirements (RBC)

(18)

Table 1: Rating-based Capital Requirements NAIC

Rating AAA AA A BBB BB B CCC CC C D

RBC % 0.40% 0.40% 0.40% 1.30% 4.60% 10.00% 23.00% 23.00% 23.00% 30.00%

Just to show that the Rating-based Capital Requirements of the NAIC are in line with those of other regulators, a test for correlation with rating specific Risk Weights for banks presented in the Basel II Accord is performed. A table with these Risk Weights is presented below in Table 2. The correlation is high at 0.884, giving support for the acceptability of using the RBC requirements of the NAIC.

Table 2: Risk Weights Basel II Accord

Rating AAA AA A BBB BB B CCC CC C

Risk

(19)

Construction of the Credit Rating Migration (CRM) Variable

Credit rating migration is a variable that is constructed by having change in rating-based capital requirements model credit rating migration. Original credit rating migration data is extracted from Standard & Poor's CreditPro® - Corporate Ratings, in the form of transition matrices for every quarter in the time span of this investigation, which is a total of 120 transition matrices. The Rating-Based Capital requirements that are used are determined by the NAIC and are presented earlier in Table 1.

Quarterly Credit Rating Migration

Below is the transition matrix for all major rating categories in the fourth quarter of 2010. It will be used as an example to show how the credit rating migration variable is constructed.

Figure 1: Transition Matrix 2010 Quarter 4

Rating Issuers AAA AA A BBB BB B CCC CC C D

AAA 35 88,57 11,43 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 AA 127 0,00 96,85 3,15 0,00 0,00 0,00 0,00 0,00 0,00 0,00 A 540 0,00 0,56 98,33 1,11 0,00 0,00 0,00 0,00 0,00 0,00 BBB 717 0,00 0,00 0,84 98,74 0,28 0,14 0,00 0,00 0,00 0,00 BB 471 0,00 0,00 0,00 1,70 96,60 1,70 0,00 0,00 0,00 0,00 B 843 0,00 0,00 0,12 0,00 1,42 97,39 0,95 0,00 0,00 0,12 CCC 91 0,00 0,00 0,00 0,00 0,00 8,79 81,32 3,30 0,00 6,59 CC 9 0,00 0,00 0,00 0,00 0,00 0,00 0,00 55,56 0,00 44,44 C 0 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00

For the illustration, bonds will be used that were originally rated BB (See Figure 1, blue row). As can be seen in the ‘Issuers’ column, 471 bonds are given a rating of BB at the beginning of the quarter. For all of these 471 bonds the RBC requirement is 4.6% at the beginning of the quarter, as can be seen in Table 1.

Figure 1 displays the credit rating transitions that have occurred by the end of the quarter. Of the 471 firms that were originally rated BB, 8 bonds were upgraded to a BBB rating and another 8 bonds were downgraded to a B rating, and the rest retained their original rating. This can be seen from the 1.7% changes both up as well as down. These rating changes are thus 2 vectors because the change has a direction and a magnitude. These transitions will affect the RBC requirements. As for the upgraded bonds, the capital requirements will decrease. For the downgraded bonds, the RBC requirements will increase.

The total change in RBC requirements can be calculated by multiplying the vectors by the new RBC requirements. As a result of the migration at the end of the quarter RBC requirements would increase to 4.6357%.

(20)

Increase = ((1.7*1.3) + (96.6*4.6) + (1.7*10)) /4.6 – 1 = 0.7761% Table 3: Calculation of CRM Beginning of Quarter RBC requirement End of Quarter RBC requirement Increase 4.6% 4.6357% 0.7761%

The ‘Increase’ value represents the quarterly credit rating migration. Conversion to Annual Credit Rating Migration

Very few corporations experience rating changes in a single quarter. This is caused by the short time span, plus the fact that there are relatively few bonds in most rating categories. As a result, quarterly CRM data is quite volatile. To determine a relationship between CRM data and macroeconomic factors it makes sense to annualize the data series.

Quarterly changes can be converted to annual changes. The average annual increase is calculated by: adding 1 to all ‘increases’ and multiplying the new value for the current quarter and the three preceding quarters by each other and finally subtracting 1 again. In an equation it is written as shown below:

ACRM = (1+QCRM t-3)*(1+QCRM t-2)*(1+QCRM t-1)*(1+QCRM t) - 1

ACRM = Annual Credit Rating Migration QCRM = Quarterly Credit Rating Migration

(21)

Independent Variables

In this model a large amount of independent variables will be tested, all of which are macro-economic variables, these variables are supposed to jointly capture shifts in the general economic environment. For the purpose of the analysis the (absolute or relative) change for the following time series have been used.

SMP - Stock Market Price

The stock market price variable represents the composite equity price index of the S&P 500 index. The predictive power of equity prices on default rates has long been established. It was first seen in the Merton model, which was built upon the Black-Scholes model in 1974 (Merton 1974). Stock market prices are an indicator of the overall health of the stock market and therefore a good proxy for the entire economy. They also represent the joint expectations of all the investors.

GDP - Gross Domestic Product of the US

The gross domestic product of the US represents the total final production of goods and services and therefore is a proxy for the total demand/supply of the US economy (Qu 2008). It is commonly used as a proxy for the state of the macro economy as was done in Nickell et al (2000).

EXGDP - The exchange rate for the US dollar and the Great Britain Pound EXYEN - The exchange rate for the US dollar and the Japanese Yen

The United States is a large open economy, and exchange rates are expected to affect the firms. The direction in which the effect occurs is ambiguous because a low dollar value increases competitiveness of US firms and results in more exports, which improves the financial performance of the exporting firms. On the other hand, the reverse is true for the importing firms. A low dollar value means imports are more expensive and costs are high, decreasing the financial performance of the firms that rely on imports.

In addition to the effect on exports and imports, exchange rates also indicate the overall faith of money holders to hold US dollars (Flood and Garber 1983). If they anticipate US economic growth they will hold US dollars if they anticipate US economic decline they will sell their US dollars driving down the exchange rate. If there is an economic recession in Great Britain or in Japan this will also appear in their respective exchange rates and this will only say how the macroeconomic condition in the US is relative to Great Britain or Japan.

(22)

HPI - Housing Price Index

The housing price index will be measured by the Case-Shiller index, which is a commonly used housing price index in the US. It is, like GDP, a commonly used proxy for the general state of the economy. From 1981 until 1986 the Case-Shiller index did not exist yet and therefore the Federal Housing Finance Agency’s (FHFA) housing price index was used as a proxy in order to balance the dataset.

iShort - 1-year interest rate iLong - 10-year interest rate iSpread – Interest rate spread

For interest rate measures, a 1-year and a 10-year treasury at a constant maturity rate is used. In economic theory, interest rates represent the price of the capital; increases will cause some parties to be priced out of the market. Thus, if costs go up this will make it more difficult for borrowers to service their obligations and more difficult to obtain new capital and credit (Qu 2008). The difference in 10 and 1-year interest rates, also referred to as “interest rate spread,” is also regressed to determine if it is the difference that drives transitions.

CPI – Consumer Price Index in the US

The total consumer price index for all urban consumers is a measure of inflation. It represents the price of a basket of goods purchased by an urban consumer. The effect of inflation is that it erodes the purchasing power of savings and therefore pressures those with savings to invest (Qu 2008). In the case of the US where over inflation is not a risk, it is beneficial for the economy if more funds enter the market as a result of pressures due to inflation. It can, however, also have negative effects on CRM if it is driven by rising prices such as increases in the price of oil.

Oil - Spot oil price

(23)

Table 4: Table of Variables/ Core economic data used

Variable

Name Abbreviation Frequency Source

Seasonally

Adjusted Description

Credit Rating

Migration CRM Quarterly

Standard and Poor's, CreditPro

Not

Applicable Constructed Variable Stock

Market Price SMP Quarterly Standard and Poor's Not

Applicable S&P 500 Index, index Gross Domestic Product US GDP Quarterly US Department of Commerce: Bureau of Economic Analysis Seasonally Adjusted Annual Rate

Total production of goods and services in US, Billions of US dollars

Exchange

Rate US/UK EXGBP Quarterly

Board of Governors of the Federal Reserve System

Not Applicable

Daily average of spot price on national markets, US dollars to 1 British Pound

Exchange Rate

Japan/US EXYEN Quarterly

Not Applicable

Daily average of spot price on national markets, Japanese Yen to 1 US dollar

Housing

Price Index HPI Quarterly

Standard and Poor's, Federal Housing Finance Agency

Seasonally Adjusted

1981-1986 FHFA housing price index, 1987-2010 Case-Shiller housing price index, index 1-Year

Interest Rate iSHORT Quarterly

Not Applicable

US Treasury Constant Maturity Rate, percentage

10-Year

Interest Rate iLONG Quarterly

Not Applicable

US Treasury Constant Maturity Rate, percentage

Interest Rate

Spread iSpread Quarterly

Not

Applicable Difference in iLong and iShort Consumer

Price Index CPI Quarterly

U.S. Department of Labor: Bureau of Labor Statistics

Seasonally

Adjusted Is a measure of inflation, index Spot Oil

Price Oil Quarterly

Dow Jones & Company

Not Adjusted

Daily average of spot price, US dollars per barrel

(24)

Conversion to Annual Changes

For the macro data to be compatible with the CRM variable and appropriate for entering the model, annual changes have to be calculated, otherwise it would not be meaningful to regress them against the CRM annual changes. For macro variables, it is necessary to asses on a case by case basis whether absolute change or relative change is more appropriate.

Absolute changes

For all interest rate related variables the annual absolute change is always used. These variables are: iShort, iLong and iSpread.

An example of how this is calculated is presented below: Absolute Annual Change iShort = iShortt – iShortt-4 Relative Changes

For the remaining seven variables the use of the relative annual difference is more appropriate.

These variables are: SMP, GDP, EXGBP, EXYEN, HPI, CPI, and Oil. An example of how this is calculated is presented below:

Relative Annual Change GDP = GDPt / GDPt-4 - 1

(25)

The Statistical Model

Below is the ‘general’ model that will be estimated:

CRM = β1 + β2 SMP + β3 GDP + β4 EXGBP + β5 EXYEN + β6 HPI + β7 iShort + β8 iLong + β9 iSpread + β10 CPI + β11 Oil + ε β1 = Constant

ε = Error term

Note: Variable acronyms are listed in Table 4, in this model all acronyms refer to annual changes.

Estimation Method

A Stepwise Ordinary Least Squares Regression is performed for all suitable rating categories where we begin with a full model and remove variables that have a p-value that is higher than 0.1 and are therefore not significant at the 10% level. All coefficients displayed in the results tables therefore have a significance level of at least 10%.

OLS Assumptions

There are four main assumptions that have to be met for the OLS regression to be appropriate and for coefficients to be Best Linear Unbiased Estimates (BLUE). The assumptions are listed below:

Endogeneity Multicollinearity Homoskedasticity Normality

(26)

Individual Rating Categories

Results

Table 5: Results of Robust Stepwise OLS regressions at 10%, Investment Grade AAA AA A BBB Sign SMP -0.068 -0.381 -0.12 - (3.20)** (3.49)** -1.78 GDP 0.536 -4.321 -1.645 - (2.87)** (3.89)** (2.63)** EXGBP 0.086 - (2.17)* EXYEN -0.083 -0.624 -0.905 -0.468 - (2.48)* (5.66)** (4.20)** (5.97)** HPI -0.29 0.375 1.114 0.927 + (3.10)** (3.22)** (2.68)** (5.15)** iShort -0.021 -0.068 -0.032 - (4.12)** (2.69)** (4.24)** iLong -0.011 0.053 + (3.39)** -1.7 iSpread -0.008 -0.035 -0.054 - (2.10)* (3.83)** (2.50)* CPI 1.92 + -1.73 Oil -0.06 + -1.78 Constant 0.009 0.071 0.585 0.199 + -0.76 (4.71)** (7.63)** (5.77)** Observations 117 117 117 117 Mean VIF 1.41 1.23 2.15 1.6 R-squared 0.33 0.4 0.38 0.42

Absolute value of t statistics in parentheses All coefficients are significant to at least 10%; *

(27)

Table 6: Results of Robust Stepwise OLS regressions at 10%, Below Investment Grade

BB B CCC CC C Sign SMP -0.077 -0.098 0.409 - -1.82 (2.36)* (4.56)** GDP -2.129 -2.814 -2.09 -4.838 -2.53 - (3.87)** (5.44)** (5.44)** (3.68)** (2.94)** EXGBP -0.124 -0.281 -1.253 - -1.68 (3.73)** (5.77)** EXYEN -0.198 -0.266 -0.196 -0.632 - (3.98)** (4.25)** (3.00)** (2.55)* HPI 0.607 0.761 0.944 0.794 + (5.24)** (5.35)** (7.03)** -1.66 iShort -0.041 -0.049 -0.056 -0.083 - (5.89)** (5.38)** (5.00)** (2.37)* iLong 0.041 0.044 0.06 0.113 + (4.53)** (3.99)** (4.74)** (2.87)** iSpread -0.032 -0.073 0.068 - (3.39)** (2.88)** (4.72)** CPI 2.039 -4.713 3.913 + (2.70)** (2.13)* (2.93)** Oil -0.05 0.054 0.2 + (2.18)* (2.66)** (2.64)** Constant 0.172 0.257 0.116 0.696 -0.018 + (7.14)** (9.63)** (4.38)** (9.63)** -0.38 Observations 117 117 117 117 117 Mean VIF 2.1 1.95 2.13 2.36 1.38 R-squared 0.54 0.57 0.46 0.38 0.26

(28)

Results Analysis

In analyzing the results we can look at Tables 5 and 6.

Some things that we would like to look at are the R-Squared values, the significance levels of the coefficients, the signs of the coefficients and the quantitative impact of the independent variables.

R-Squared Values

The R-squared values can be seen in Table 7 below:

Table 7: R-Squared values

Rating AAA AA A BBB BB B CCC CC C

R-squared 0.33 0.4 0.38 0.42 0.54 0.57 0.46 0.38 0.26

The R-squared values are a good initial indicator of the strength of the model and it tells if there indeed is a relationship between the macro economy and credit rating migrations or not. From the relatively high R-squared values of on average 0.416, one could even conclude that a very significant part of credit rating migrations can be modeled or explained by changes in the macroeconomic environment.

Which independent variables are best?

All independent variables are significant enough to be included in at least four models. Some variables appear significant in more models than others though. An overview of the frequency of inclusion in a model is presented in Table 8 below:

Table 8: Frequency of occurrence in models

Variable Frequency of Occurrence

EXYEN 8 HPI 8 GDP 8 iShort 7 SMP 6 iLong 6 iSpread 6 CPI 4 Oil 4 EXGBP 4

(29)

variables should be considered when predicting or anticipating credit rating migrations. These top 4 variables should be included in a ‘specific’ model opposed to the ‘general’ model, which would include all 10 variables. This is done later in this paper.

Signs of coefficients

The signs of the coefficients can be seen in the results tables 5 and 6 by the coefficients but also in the rightmost columns of both tables. The signs of coefficients are relatively consistent across all 9 regressions; this is advantageous as it suggests a greater predictive/explanatory power of the individual regressors than if they had not been consistent.

It is interesting to see in what direction the 7 statistically most significant variables effect credit rating migration.

EXYEN is negative, meaning that according to the model when the US dollar appreciates relative to the Japanese Yen, the credit ratings are expected to improve.

HPI is positive, which means that if the house prices increase credit ratings are predicted to deteriorate.

iShort is negative, indicating that if the short term interest rate increases credit ratings are expected to improve.

SMP and GDP are both negative and suggest that if they increase credit ratings are also expected to improve.

iLong is positive, which therefore suggests that long term interest rates have a different effect on credit ratings than short term interest rates. Namely, that when the long-term interest rate increases credit rating are expected to deteriorate.

iSpread is negative, indicating that as the difference between the short-term and long-term interest rates increases ratings are expected to increase.

The last coefficient that is worth mentioning is the constant. The constant is significant to a 1% level in seven out of nine regressions and is consistently positive with the exception of the single C regression. This indicates that factors outside of the model and probably not macroeconomic variables make credit ratings decrease. A likely explanation deduced from the literature review is the aging effect (Altman Kao 1992), which predicts that ratings are likely to deteriorate after issuance.

Quantitative Impacts of Macro Variables

(30)

understood as follows: for example, in the BB rating category, if GDP increases by 1%, the credit rating migration measured by the rating-based capital requirements are expected to decrease by 2.129%, keeping all other variables constant.

It is important to look at the average absolute value of the coefficients to determine which macro variables have greater impacts on credit rating migration modeled by rating-based capital requirements.

Below is a table of the average absolute values of coefficients:

Table 9: Average absolute values of coefficients for individual rating categories

Variable Average Absolute

value of coefficients CPI 3.146 GDP 2.613 HPI 0.727 EXGBP 0.436 EXYEN _0.422 SMP 0.192 Oil 0.091 iLong _0.054 iShort 0.050 iSpread 0.045

(31)

Portfolio Regression

Having looked at the effect of the macro economy on individual rating categories, for practitioners it is also interesting to look at what effect the macro economy would have on a portfolio of credit risk assets. In case of rating migration a model would be helpful to relate the macro economy to rating-based capital requirements. This could be conceived by blending the previous results. A disadvantage is that the regressions of rating categories such as AAA and C are less reliable due to a small number of rated assets in these categories. A cleaner model would be to perform a portfolio-level regression. The added advantage of the portfolio-level CRM variable is that it is based on more underlying rated corporations, and hence is more stable and likely more related to the macro economy. Constructing a portfolio is done by weighting every rating category. A stepwise OLS regression can then again be run. In Table 10, below, one can see a ‘typical’ portfolio composition in the column labeled: Typical Portfolio. However, different practitioners are likely to have different portfolio compositions. The portfolio composition examined in this study is the entire rated universe. Weights are determined by the average number of issuers in each rating category for every quarter from 1981 until 2010. These weights are listed in the column labeled: Issuers Portfolio. It should be noted that all rating categories are included in the model.

Table 10: Portfolio Composition

(32)

Results Portfolio Regression

Table 11: Results Issuers Portfolio

Robust Stepwise OLS Regression at 10%

Issuers SMP -0.110 (3.46)** GDP -2.093 (5.45)** iLong 0.037 (4.57)** EXYEN -0.220 (4.46)** HPI 0.692 (6.12)** iShort -0.045 (6.55)** Constant 0.214 (9.95)** Observations 117 Mean VIF 1.95 R-squared 0.58

Robust t statistics in parentheses

All coefficients are significant to at least 10%; * significant at 5%; ** significant at 1% Note: All VIF values are less than 5.

Analyzing Portfolio Regression Results

(33)

It is interesting to look at the quantitative impacts of the individual macro variables. It can be seen that the GDP has by far, the greatest quantitative impact on credit rating migration. The coefficients for HPI, EXYEN and SMP are considerably lower but when combined they can have a considerable impact. The impacts of iShort and iLong are statistically significant but only have a small quantitative impact.

Table 12: Quantitative impacts of macro variables for the portfolio regression

(34)

Panel Regression

A third regression is also worth running and analyzing. This regression is a panel regression. In a panel regression it is assumed that the effects of the macroeconomic variables are the same across all rating categories. The data set has been rearranged so that all the rating categories are under each other, the number of observations increases from 117 to 1,053.

Results Panel Regression

Table 13: Results Panel CRM Robust Stepwise OLS Regression at 10% CRM iSpread -0.016 (2.36)* GDP -2.137 (6.36)** EXGBP -0.186 (3.00)** EXYEN -0.375 (5.27)** HPI 0.562 (4.47)** iShort -0.04 (5.46)** iLong 0.035 (3.76)** Constant 0.24 (10.51)** Observations 1,053 Mean VIF 2.21 R-Squared 0.11

(35)

Analyzing Panel Regression Results

It is nice to see that again the top four performing macro variables are significant at a 1% level. These variables are: EXYEN, HPI, GDP and iShort. They also have the same sign as we have seen in the previous regressions, as do the other coefficients. Two other macro economic variables that were amongst the top seven seen in the individual rating regressions are again included, namely: iSpread and iLong. Surprising is, however, the appearance of the EXGBP as a significant variable. It was one of the worst three performing variables in the first set of regressions and was excluded from the model in the portfolio regression. The R-squared value of 0.11 is much lower than the values we have encountered earlier but this is not surprising as it is due to the assumption of equal effects of macro variables across all rating categories.

In terms of the quantitative impacts, a similar trend can be observed as in the individual rating and portfolio regressions. GDP again has the greatest quantitative impact, followed by HPI and then both the exchange rates. Last are again all three interest rate variables, having quantitatively small impacts.

Table 14: Quantitative impacts of macro variables for the panel regression

(36)

Specific Model

After analyzing the results presented in previous sections we can conclude that the reduced form of the ‘general’ model tested originally should include the following four variables: GDP, EXEN, HPI and iShort. These four variables occurred most frequently in the individual rating regressions and remained statistically significant in both the portfolio and panel regressions. SMP, iSpread and iLong appeared a respectable 6 times each in the individual rating category regressions. However, SMP and iSpread were not consistently present in the portfolio and panel regressions. iLong was present in both the portfolio and panel regression so could arguably be included in the specific model. Nevertheless it was excluded because it has a high degree of multicollinearity with iShort.

The resulting specific model would thus look as follows:

CRM = β1 + β2 GDP + β3 EXYEN + β4 HPI + β5 iShort + ε β1 = Constant

ε = Error term

Note: Variable acronyms are listed in Table 4, the acronyms refer to annual changes.

(37)

Results Specific Model Regression

Table 15: Results for the specific model:

Issuers Portfolio Robust Stepwise OLS Regression at 10%

Issuers GDP -2.338 (4.83)** EXYEN -0.163 (3.39)** HPI 0.522 (3.95)** iShort -0.020 (4.71)** Constant 0.225 (8.50)** Observations 117 Mean VIF 1.31 R-squared 0.44

All coefficients are significant to at least 10%; * significant at 5%; ** significant at 1% Note: All VIF values are less than 2.

Analyzing Specific Model Regression Results

The specific model holds up well when evaluated based on the R-Squared value. The R-Squared value is 0.44, which means that a significant part of credit rating migration can indeed be predicted by the four selected macroeconomic variables.

(38)

Analysis of 4 variables GDP

GDP has by far the largest quantitative impact on rating-based capital requirements and therefore also on credit rating migration. As commented earlier it has a negative impact, indicating that as GDP rises ratings are expected to improve as well. Both the size of the impact and the direction are as was to be expected from literature. Nickell et al (2000) only look at the effect of business cycles by means of examining the effect of GDP growth and they find that rating downgrades are more likely when GDP growth deteriorates (Allen and Saunders 2003). Bangia et al (2002) also find that as there are downturns in the business cycle, ratings are expected to deteriorate (Allen and Saunders 2003). Bangia et al (2002) take more than just GDP to determine what the state of the macro economy is but GDP is the most influential component.

Seeing as the majority of research regarding the effect of the macro economy on credit rating migration has focused on the effect of GDP it is interesting to see how the specific model holds up against a model with only GDP. The outcome of a robust stepwise regression at 10%, like those run earlier for just the effect of GDP, is presented below:

Table 16: Results for GDP Issuers Portfolio Robust Stepwise OLS Regression at 10%

Issuers GDP -2.222 (7.08)** Constant 0.252 (12.08)** Observations 117 R-squared 0.28

Robust t statistics in parentheses ** significant at 1%

(39)

EXYEN

The significance level of the exchange rate for the US dollar and the Yen is surprising. It is significant to a 1% level in both the portfolio and the panel regressions as well as in 6 out of the 8 times it is present in the individual rating categories. The sign of the coefficient is consistently negative and the size of the impact is small but definitely noteworthy. When the US dollar appreciates relative to the Yen credit ratings are expected to improve and there would be lower rating-based capital requirements.

The question is begged whether or not this finding is in line with previous research. I have not encountered research relating exchange rates directly to credit rating migration so this is potentially a newly uncovered relationship. What has been examined in literature is the relationship between exchange rates and expected default frequency. Seeing as credit ratings are essentially a measure of the probability of default, these finding are open to comparison. Qu (2008) finds that at the industry level exchange rates will have varying effects, due to the complexity of exchange rate dynamics and certain industries being more dependent on exports and imports than others. Based on Qu’s research, one would therefore at a national level, not necessarily expect to find a consistently negative effect. The exchange rate is, however, also an indicator of the faith of money holders to hold US dollars (Flood and Garber 1983). It would be expected that the coefficient is negative because when the faith in the US economy is higher, one would expect higher ratings or less downwards rating migration.

What is most striking about the high significance level of the EXYEN, however, is the lack of significance of the EXGBP. If exchange rates are a significant factor in determining credit rating migration, one would expect that, in the absence of high multicolinearity, both exchange rate variables to be significant. The relationship between exchange rates and credit rating migration therefore merits further research.

HPI

(40)

iShort

The impact of the short-term interest rate is relatively small. The sign is negative, meaning that as the short-term interest rate goes up downward rating migrations decrease. This is in contrast to what Qu (2008) writes about regarding the effect of the short-term interest rate on expected default frequency. He writes that when interest rates increase, the cost of capital increases and therefore deteriorates the liquidity of firms. Observing the opposite effect of what would be expected from theory does, however, not come as a complete surprise. This is because the Federal Bank artificially decreases the short-term interest rate to stimulate the economy during recessions (Economic Report of the President 1990). It is likely that the observed negative relationship between the short-term interest rate and rating-based capital requirements is explained by the actions of the Federal Bank.

Specific Model Graph

In order to get a more intuitive feel for how well the specific model models performs it is useful to take a look at a graph displaying both what is predicted by the specific model and what is observed over time (Issuers Portfolio). This is displayed below:

(41)

Conclusion

From this investigation we can conclude that the macroeconomic environment has a significant impact on the behavior of credit ratings and therefore also on credit risk as a whole. When analyzing credit risk at an asset or issuer level, microeconomic factor are usually deemed most relevant. As this study shows, at a portfolio level macroeconomic factors should be considered as the core variables that impact credit ratings over time. These findings are not new as they were already suggested by Bangia et al (2002), Nickell et al (2000), and Trück and Rachev (2005), among others. What is new in this study is the examination of individual macroeconomic factors. It is shown that the (US) GDP, the US dollar/Japanese Yen exchange rate, the US housing price index, and the short-term US$ interest rates have significant effects on credit rating migrations for US Corporate Bonds, as measured by changes in rating-based capital requirements. The added value of these findings is that it expands the academic research knowledge, which might spur further investigation into the subject. Based on this research, practitioners better know what macroeconomic variables have to be considered when managing credit risk. New is also the suggested innovative approach to translate a rating migration matrix into one variable using rating-based changes in capital requirements. This allowed this investigation to work with a continuous variable and perform stepwise OLS regressions, which distinguishes this paper from previous research on credit rating migration. Policy makers and regulators of financial institutions can use the results of this study to analyze how macroeconomic variables, which they may or may not be able to influence, impact the stability of banks and insurance companies.

Limitations of this study go hand in hand with some of the suggestions for future research mentioned below. Particularly, a deeper (more underlying assets per rating class) and longer dataset would have resulted in more robust findings, especially for the rating categories: AAA, CCC, CC and C. In these categories the average number of issuers throughout the time span of the investigation was less than 100. Another limitation of this investigation, which is closely linked to the previously mentioned limitation, is that no out of sample testing has been performed. By splitting the dataset this could have done. In that case one would have been able to see whether the results of regressions performed on the first part of the sample held when tested on the other part of the sample. This was not done; as it was preferred to keep the dataset used for the regressions as large as possible. It would also be interesting and useful to perform out of sample testing by comparing the results based on S&P ratings to a universe of bonds rated by other rating agencies, like Moody’s.

(42)

significant here, and find ways to better predict them with the aim of better predicting future credit rating migrations.

(43)

References

Allen, L., Saunders, A., 2003. A survey of cyclical effects in credit risk measurement models. Technical report, BIS Working Paper 126.

Altman, E.I., Kao, D.L. 1992. The implications of corporate bond ratings drift. Financial

Analysts Journal, Pp 64-75.

Bangia, A., Diebold F.X., Kronimus, A., Schagen, C., Schuermann, T., 2002, Ratings migration and the business cycle, with application to credit portfolio stress testing, Journal of Fixed Income, Pp 445-474 Basel Committee on Bank Supervision (2003, January). The new Basel capital accord.

Report, Bank of International Settlements, Basel.

Campos, J., Ericsson, N. R., Hendry, D.F., 2005. General-to-specific Modeling: An Overview and Selected Bibliography, Board of Governors of the Federal Reserve System International Finance Discussion

Papers, Number 838, Pp 1-92

Caouette, J., E. Altman, Narayanan, P., 1998. Managing credit risk. The next great

financial challenge. New York: Wiley.

Carty, L.V., Fons, J., 1993. Measuring Changes in Credit Quality. Moody's special report, Moody's Investors Service, New York.

Carty, L.V., 1997. Moody's Rating Migration and Credit Quality Correlation, 1920-1996. Special

comment, Moody's Investors Service, New York.

Credit Suisse Financial Products, 1997. CreditRisk+: Technical manual. Discussion paper, CSFP, New York.

Economic Report of the President, 1990. Washington, D.C.: U.S. Government Printing Office. Flood, R., Garber, P., 1983. A Model of Stochastic Process Switching, Econometrica, 51, Pp 537-551. Friedson, M., Garman, C., Wu, S., 1997. Real Interest Rates and the Default Rates on

High-yield Bonds, Journal of Fixed Income, Pp27-34

Greene, W.H., 2000. Econometric Analysis (fifth edition), Prentice Hall, Pp 736-740. Groot, A.D. de, 1961. Methodologie. Den Haag: Mouton.

Helwege, J., Kleiman, P., 1996. Understanding aggregate default rates of high yield bonds. Federal

Reserve Bank of New York Current Issues In Economics and Finance 2 (6), Pp1-6.

Hendry, D. F., 1995 .Econometrics and Business Cycle Empirics, Economic Journal, 105, 433, Pp 1622-1636.

Hoover, K.D., Perez, S.J., 1999. Data mining reconsidered: encompassing and the general-to-specific approach to specification search. Econometrics Journal, 2, Pp 167-191.

Jarrow, R.A., Turnbull, S., 1995. Pricing Derivatives on Financial Securities Subject to Credit Risk.

(44)

Jonsson, J.G., Fridson, M.S., 1996. Forecasting default rates on high-yield bonds. The Journal of

Fixed Income, Pp 69-77.

JP Morgan, 1997. Creditmetrics - Technical Document. Discussion paper, JP Morgan, New York.

Keynes, J. M., 1938. ‘Letter of 4 July 1938 from Keynes to Harrod on Tinbergen’s work,’ in The Collected Writings of John Maynard Keynes XIV The General Theory and After, Part II Defense and Development, London and Basingstoke: MacMillan, St, Martins Press, 1973, Pp 295-297.

Keynes, J. M., 1939.Professor Tinbergen.’s Method. Economic Journal, 49, 195, Pp 558-568. Keynes, J. M., 1940. Comment. Economic Journal, 50, 197, Pp 154-156.

Koopman, S.J., Lucas, A., 2003. Business and Default Cycles for Credit Risk, Tinbergen Institute

Discussion Paper, Pp 1-20.

Koopmans, T. C., 1947. Measurement Without Theory. Review of Economics and Statistics (formerly the

Review of Economic Statistics), 29, 3, Pp161-172.

Leamer, E. E., 1978. Specification Searches: Ad Hoc Inference with Non-experimental Data, JohnWiley, New York.

Lucas, D.J., Lonski, J.G., 1992. Changes in corporate credit quality 1970-1990. Journal of Fixed

Income, Pp 7-14.

Merton, Robert C,. 1974. On the Pricing of Corporate Debt: The Risk Structure of Interest Rates. Journal

of Finance 29, Pp 449–470

Nickell, P., W. Perraudin, Varotto, S., 2000. Stability of rating transitions. Journal

of Banking and Finance 24 (1–2), Pp 203–227.

O'Brien, Robert M., 2007. A Caution Regarding Rules of Thumb for Variance Inflation Factors, Quality

and Quantity 41(5), Pp 673-690.

Owen, P.D., 2002. General-to-Specific Modelling Using PcGets, University of Otago Economics

Discussion Papers, No. 0213, Pp 1-23

Qu, Y., 2008, Macro Economic Factors and Probability of Default, European Journal of Economics,

Finance and Administrative Sciences, Issue 13, Pp 192-215

Sargan, J.D., 1964. Wages and prices in the United Kingdom: a study in econometric methodology. in P.E. Hart, G. Mills and J.K. Whitaker (eds), Econometric Analysis for National Economic Planning. London: Butterworth.

Thadewald, T., Büning, H., 2007, Jarque--Bera Test and its Competitors for Testing Normality -- A Power Comparison, Journal of Applied Statistics, Volume 35, Number 1, Pp 87-105

Trück, S., Rachev S.T., 2005, Credit Portfolio Risk and PD Confidence Sets through the Business Cycle.

Journal of Credit Risk, Volume 3, Number 2, Pp 1-34

(45)

Appendix

Graphs of Changes Macro Variables

(46)

(47)

(48)

(49)

Predicting Credit Rating Migration Consistently with Macroeconomic Conditions