Master Master

Master Master

Master Master Thesis Thesis Thesis Thesis Title: Title: Title: Title:

The

The The The Influence Influence Influence Influence of of of of House House House House Prices Prices Prices Prices on on on on the the the the Credit Credit Credit Credit Risk Risk Risk Risk of of of of Commercial Commercial Commercial Commercial Banks Banks Banks Banks in

in in

in the the the the US US US US

Groningen, October 31^st2008

Author Author Author

Author: SupervisorSupervisorSupervisorSupervisor: SecondSecondSecondSecond SupervisorSupervisorSupervisorSupervisor

Xiaofei Guo Dr. Michael Koetter Dr. Padma Rao Sahib

Faculty of Economics & Business Faculty of Economics & Business Faculty of Economics & Business University of Groningen University of Groningen University of Groningen

The Netherlands The Netherlands The Netherlands

X.Guo.2@student.rug.nl m.koetter@rug.nl P.Rao.Sahib@rug.nl

Abstract Abstract Abstract Abstract

The subprime mortgage crisis led to severe economic problems throughout 2007 and 2008 in the U.S. The crisis can be attributed to a number of factors in both the housing market and the credit market. This paper examines the effect of combinations of seven factors (HPI, HOSR, PCDI, NOB, LOAN, NPL, and FFR) on the credit risk of commercial banks in the United States. An econometric model is specified and estimated in order to study the relationship between explanatory and dependent variables. The research is performed utilizing a panel data set that consists of repeated observations on 49 individual states of the United States for the period 1997-2007.

The regression results provide the answer to the main question of this paper, namely that house price index has a positive impact on the credit risk of commercial banks.

Instead of stopping at only the financial market in America, this crisis has already spread to influence economies in the whole world. Major banks and other financial institutions suffered from this crisis, with reporting a large amount of losses. The results of this research show compelling proof to suggest banks to pay more attention to the possible changes of house prices when they measure their credit risk in the future.

KEYKEY

KEYKEY WORDSWORDSWORDSWORDS: Credit risk, Subprime crisis, Commercial banks, House prices, LLA and The United States.

Content Content Content Content

LISTLIST

LISTLIST OFOFOFOF ABBREVIATIONSABBREVIATIONSABBREVIATIONSABBREVIATIONS 4 1.

1.

1.

1. INTRODUCTIONINTRODUCTIONINTRODUCTIONINTRODUCTION 5 2.2.

2.2. HYPOTHESESHYPOTHESESHYPOTHESESHYPOTHESES ANDANDANDAND THETHETHETHE ECONOMICECONOMICECONOMICECONOMIC MODELMODELMODELMODEL 7

2.1) HYPOTHESES 7

2.1.1) BANKING RISK 7

2.1.2) WHY ON STATE LEVEL 8

2.1.3) INFLUENTIAL FACTORS OF CREDIT RISKS 9

2.2) THE ECONOMIC MODEL 13.

3.

3.

3.

3. DATADATADATADATA SOURCESSOURCESSOURCESSOURCES ANDANDANDAND SAMPLESAMPLESAMPLESAMPLE DESCRIPTIONDESCRIPTIONDESCRIPTIONDESCRIPTION 14141414

3.1) DEPENDENT VARIABLES 14

3.2) INDEPENDENT VARIABLES 14

3.3) DUMMYVARIABLE: 15

3.4) CHOOSING A METHODOLOGY 16

4.

4.

4.

4. DATADATADATADATA ANALYSESANALYSESANALYSESANALYSES 16161616

4.1)DESCRIPTIVE STATISTICS 16

4.2) DIAGNOSTIC CHECKS 17

4.3) RESULTS OF DIAGNOSTIC CHECKS 19

5.5.

5.5. REGRESSIONREGRESSIONREGRESSIONREGRESSION RRRREEEESULTSULTSULTSULT 20202020 6.6.

6.

6. CONCLUSIONCONCLUSIONCONCLUSIONCONCLUSION 21212121 REFERENCES

REFERENCES REFERENCES

REFERENCES 23232323

APPENDIX APPENDIX APPENDIX

APPENDIX 26262626

List List List List of of of of Abbreviations Abbreviations Abbreviations Abbreviations

LLA Loan Loss Allowance

RoA Return on Assets

GAAP Generally Accepted Accounting Principles IBBEA Interstate Banking and Branching Efficiency Act OFHEO Office of Federal Housing Enterprise Oversight GMM Generalized Method of Moments

DW Durbin-Watson

LM Lagrange Multiplier Test

RESET Regression Specification Error Test HPI House Price Index per state in the U.S HOSR Homeownership Rate per state in the U.S NOB Number of banks per state in the U.S LOAN Total loans to total assets

NPL Non-performing loans to total assets

FFR FED Funds Rate

1.

1.

1.1. IntroductionIntroductionIntroductionIntroduction

The US economy is the world’s largest national economy. At the beginning of 2008, it was already in a turbulent position due to the subprime mortgage crisis in the U.S. On July 11^th, under the pressure of increasing foreclosures, declining housing prices and tighter credit standards, the largest mortgage lender in the U.S., the IndyMac Bank, went bankrupt. On September 14^th, Merrill Lynch was sold to Bank of America and on the same day, Lehman Brothers filed for bankruptcy. In March 2008, the ex-chairman of Federal Reserve, Mr. A. Greenspan, contended that the 2008 financial crisis in the United States probably has to be judged as the most severe since the end of World War II. Although subprime mortgage is not the only contributor to this financial crisis in the U.S, most economists regard this as a major reason.

The Subprime crisis is a serious economic problem in the whole banking system. This crisis started with the dramatic fluctuation of the American house prices (see Figure 1) and high default rates on subprime. Subprime is a high-cost lending (Souphala and Anthony, 2006) offered to the higher-risk borrowers. Higher-risk borrowers are a group of borrowers with worse credit history and lower income than prime borrowers.

Because of a lack of sufficient predictions on changes in house prices, defaults and foreclosure¹ rates raised dramatically. The effect of the subprime crisis on financial institutions was dramatic, as many banks suffered significant losses or changed to lower bank ratings and some even went bankrupt. Synchronously, this crisis is not simply a domestic problem within the U.S; instead the widespread dispersion of credit risk has severe influence on the world economy.

This paper focuses on examining the influence of house prices on the credit risk of commercial banks in the U.S. Severe home price decreases lead to a certain amount of owners holding negative equity, which means that the value of the property they have is lower than their mortgage debt. This resulted in higher-than-expected home foreclosure rates as well as tighter credit standards with respect to subprime mortgages (Brauneis and Stachowicz, 2007), and subsequently the subprime mortgage industry collapsed - now known as the subprime crisis.

While there is some previous research that has started to estimate banks’ risks by observing GDP, income, interest rates, insurance fund, inflation rate, unemployment, stock exchange and so on, not much attention has been paid to the influence of the fluctuation of house prices on credit risks of commercial banks. Based on previous research, dramatic changes of house prices is one of the major reasons leading to the current subprime crisis. And also house prices show apparent different amounts of growth across different states. (C. Wheaton & Gleb Nechayev, 2008) Therefore, in this paper, besides several factors which have been estimated in former researches, I also try to use the house price index per state to show the influence of subprime mortgages

1 Foreclosure is a legal proceeding. In the mortgage market, borrowers are required to secure their loans by

on the credit risk of commercial banks. By adding this factor, with my study, I intend to fill this gap in theory.

This study is relevant both for academic and practical concerns because it contributes to the on-going discussion about to what extent the subprime crisis influences the credit risk of commercial banks in America. Furthermore, the study provides insight into the factors on state level that contribute to the credit risk of banks. The study also uses variables which are only available on country level. For those indicators which are announced by banks, I calculate the average value and use them as indicators on state level. I note that this paper uses the ratio of loan loss allowance to total loan (net of unearned income) to measure the credit risk of commercial banks.

To guide this study, the following conceptual model was developed:

C C C

Conceptualonceptualonceptualonceptual ModelModelModelModel

Main Main

MainMain questionquestionquestionquestion: What is the influence of house prices on the credit risk of commercial banks in the U.S.?

The theoretical model used to approach this research question depends on a relevant study conducted by Hasan and D.Wall (2004). Their research focused on the determinants of loan loss allowance in the U.S banks. They also used a panel dataset covering the period from the fiscal year 1993 to 2000. In their academic paper, they estimated a regression model to investigate the influence of their explanatory variables on the level of loan loss allowance in American banks. However, depending on different data sets, the anticipated signs of some coefficients and the form of the model can be also different.

House Price Index

Homeownership rate

Market structure (Number of banks)

Total loan/Total assets

Non-performing loans (computed as loans 90+ days late)/Total assets

Fed funds rate (Short-term interest rate)

Loan loss allowance/Total loan (net of unearned income)

This paper continues with the hypotheses and the economic model. In the third section, additional information about data sources and sample description is provided. Section four briefly presents data analysis. Section five describes the regression results.

Consequently, section six will give the conclusion and suggestions for the future research of this study.

2.2.

2.2. HypothesesHypothesesHypothesesHypotheses andandandand thethethethe economiceconomiceconomiceconomic modelmodelmodelmodel

2.1 Hypotheses

2.1.1 Banking risk

The evaluation of bank risk in commercial banks is not a newly discovered topic.

There are a number of studies investigating bank risks by using different indicators. As stated clearly by Saunders & Millon Cornett (2007), financial institutions are facing various risks: interest rate risk, market risk, credit risk, off-balance-sheet risk, foreign exchange risk, country or sovereign, technology risk, operational risk, liquidity risk and insolvency risk. In this paper, the research question focuses on measu ring the credit risk of commercial banks in the United States. As defined by Saunder & Millon Cornett (2007), credit risk is the risk that banks fail to be paid in full, including principal and interest payments, on their loans. Long maturity loans such as real estate loans cause higher credit risks than banks that possess short maturity loans. Real estate loans are mainly mortgage loans and some revolving home equity loans (Saunder &

Millon Cornett, 2007). For the second quarter of 2006, Saunders and Millon Cornett (2007) find that, in the distribution of mortgage loans for American banks, the largest percentage of the real estate loan is residential mortgage. Moreover, commercial real estate loans are also the primary reason for the significant default and credit risk problems of banks in the early 1990s. More recently, as interest rates rose and the house prices decreased dramatically, this led to a severe increase in defaults on mortgage loans.

It is indicated that five proxies for evaluating bank risks are the following: (1) Return on Assets (RoA) (2) Loan to total assets (3) non-performing Assets (4) Net charge-offs (5) leverage. Except RoA, which has a negative influence on bank risk, the other four indicators have a positive effect on bank risk.(Krishnan, Ritchken & Thomson, 2006) Sullivan and Spong (2007) suggest that the main evaluation of bank risk is the standard deviation of operating return on equity. Some other studies assess bank risk by using loan-loss reserves as the dependent variable. Higher levels of reserves account for higher future banking risk and are therefore a possible indicator for banking risk.(Altunbas et. al, 2007) Furthermore, Kolari, et al (2002) measured credit risks of large U.S commercial banks by analyzing data of allowance for loan losses, provision for loan losses and net loan charge-offs. Next to these, the empirical contribution by Oshinsky & Olin (2005) suggests that, when loss provision, loan-loss allowance (LLA),

real estate owned and loan charge-offs increase, the probability of bank failure increases. Banks with more latent and severe loan problems need a large LLA when these problems are finally recognized. (Hasan and D.Wall, 2004) Therefore, based on these findings above, this paper uses loan loss allowance as the dependent variable to measure credit risks of commercial banks.

For banks or other financial institutions, there is always a certain amount of bad debt.

Bad debt is the portion of the receivables that can no longer be returned. The actual amount of uncollectible receivable is written off as an expense called bad debt expense.

In order to offset this part of losses, banks always keep a certain amount of money known as the loan loss allowance. According to Hasan and D. Wall (2004), banks in the U.S generally follow the standard GAAP (Generally Accepted Accounting Principles) to determine their amount of LLA. At the end of each accounting period, depending on the expectation of their net charge-offs, banks determine the possible value of the loan losses in its existing portfolio. Namely, if banks decrease their LLA, that maybe means they have an optimistic opinion on their loan portfolio.

As for my study on credit risk, because I do this research on state-level but not on individual bank level, using only LLA to measure credit risk cannot really show the truth: some states have more commercial banks or maybe more large commercial banks. This condition may make the absolute value of LLA higher than that in other states. However, it does not mean these banks have more credit risks. In order to avoid this bias, I apply the ratio of LLA to the total loan² to estimate the credit risk of commercial banks on state level.

In order to investigate the influence of house prices on the credit risks of commercial banks in the United States, this research is conducted on state level. The indicators which are announced by individual banks are computed as the sum value of the indicator divided by the number of commercial banks within that state, in order to use them also on state level. As for interest rates, every state has the same interest rates announced by Federal Reserve, so this research also has an interest rate: the country level-federal funds rate. In addition, although the long term interest rate-US treasure rate may also be relevant to my research question, it has the same trend as the federal funds rate. Involving both of them may lead to the multicollinearity problem with respect to the econometric model, so I only keep the federal funds rate in my model.

2.1.2 Why on state level

The United States has 50 states and each one is on a different economic development level. According to Neely and Wheelock (1997), before the Riegle-Neal Interstate Banking and Branching Efficiency Act of 1994(IEBBA), since 1980 commercial bank failures and earnings both varied considerably across states. As a reason of this phenomenon, local economies played an important role. Depending on the regression

results of Neely and Wheelock’s research, state-level per capita income, differences in regulations and the varying market structures, all contribute to the differences in commercial bank earnings across states. In other words, banks have regulations at both the federal and state level and state-level economic activities strongly influence state bank earnings. Although the Riegle-Neal Interstate Banking and Branching Efficiency Act of 1994 allowed banks to engage in interstate branching starting in June 1997, (F.Shiers, 2002) the different performance of banks still exist because of other non- branching related bank regulations continue to differ at the state level. Moreover, the core of this research is to estimate the influence of house prices on the credit risk of commercial banks and all banks in one state will have the same value of house price (C.

Wheaton & Gleb Nechayev, 2008). Hence I will do this research on state-level rather than on bank-level or country-level. Nevertheless, for the variable federal funds rate, each state has the same values.

2.1.3 Influential Factors of Credit Risks

With regard to the mortgage lending industry, the subprime mortgage market surged from the mid 1990s, and the amount of total subprime loans increased dramatically from $65 billion in 1995 to $332 billion in 2003 (Chomsisengphet and Pennington- Cross, 2006). Furthermore, the adoption of the Depository Institutions Deregulation and Monetary Control Act in 1980 together with the Alternative Mortgage Transaction Parity Act in 1982 legalized subprime lending. From then on, financial institutions began to have the ability to charge high rates to subprime borrowers (Chomsisengphet and Pennington-Cross, 2006). Subprime lending is a high-cost loan to the borrowers who have bad credit histories and limited repayment ability. Because of subprime lending, low-income borrowers who could not afford a house before, had the opportunity to apply for a loan. Next to subprime lending, the dramatic fluctuation of house prices also impelled the subprime crisis to happen. Additionally, the former Fed chairman Greenspan commented that the recession in U.S housing might be attributed to an unsustained high house price. Between 2000 and 2005, house prices increased by 40%. Together with the relative lower growth of house supply, optimistic expectations of the housing market prompted lenders to rush to make loans - not only prime borrowers but also subprime borrowers (Brauneis and Stachowicz, 2007). In this way, banks hoped to expand their market share. Hence in recent years, the subprime loans became quite popular among both borrowers and lenders. After experiencing sustained decreasing house prices, the housing market slowdown surfaced in late 2006 (A.Moore and J.Brauneis, 2008). Since February 2007 foreclosure rates increased while housing prices declined and a large number of respondents tightened their credit standards (Brauneis & Stachowicz, 2007). Many financial institutions announced significant losses on subprime loans and subsequently increased their loan loss allowance and reduced related mortgage lending. Commercial banks were also struck hard by the fallout. They eliminated or diminished their mortgage origination business. From then on, the subprime crisis surged and still continues in 2008 (A.Moore and J.Brauneis, 2008).

From previous research it is found that interest rates are related to the fluctuation of credit risks of commercial banks (A.Moore and J.Brauneis, 2008). Lower interest rates enabled borrowers to afford larger mortgage loans and the sustained decreasing interest rates caused optimism among borrowers, that they will not have any problem on repaying while banks credit risks potentially increased. In that situation, not only borrowers but also lenders had too much confidence in the housing market and believed defaults were not going to happen. Even if it did occur, they believed that the increasing home value could always compensate their losses. But actually this attitude potentially increased the credit risk of banks (A.Moore and J.Brauneis, 2008). When more and more loans were originated through subprime lending, banks had much higher credit risks. The subprime nature of making potential unaffordable loans is not based on borrowers’ ability to repay, but on the assets they own (Chomsisengphet and Pennington-Cross, 2006). Consequently, when house prices decrease, the value of borrowers’ assets declines as well. Because a large number of subprime borrowers actually have no other assets but their home, banks credit risks increased as a result.

The income of borrowers is the major ability to repay their loans. The more shocks on income, the higher the probability that borrowers can not afford repayment (K.

Staikouras, 2005). However, the fluctuation of income cannot easily tell the trend of house purchases on its own. But the uncertainty in future income turns out to be one of the most relevant variables in the decision of homeownership (Diaz-Serrano, 2005).

R.Haurin and Leroy Gill's empirical findings (1987) proved that the consumption of housing falls when the uncertainty of income increases. The explanation was given by R.Haurin (1990), for most households, the house is the largest asset and their income is the most important source of wealth. The difference of state income probably leads to the bank earnings and failure rates vary across states. In Barakova et al's research (2003), when viewing housing as normal consumption, the values of houses should increase with income. Nevertheless, there are also other studies that do not support this.

Neely and C.Wheelock (1997) suspect that per capita income cannot explain the divergence of state level bank performance by itself. The reason is that it omits some economic shocks that may influence banks much more directly - such as oil or real estate crisis. From previous research, income should be related to the credit risks of commercial banks and in my studies, I choose to use per capita disposable income as an independent variable.

Subprime lending is a major contributor to the current rising credit risk of commercial banks, compounded by increasing interest rates, rising unemployment, economic slowdown and a cooling housing market (Wallace et al. 2008). In order to avoid the threat of deflation in the economy, the Federal Reserve reduced interest rates several times. Low interest rates from 2000 through mid-2004 spurred subprime mortgages.

However, since mid-2004, after experiencing historic low interest rates, the Federal Reserve has increased the federal funds rate 17 times to 5.25%, from only 1% in 2003.

The deceleration of housing prices, together with the increasing interest rate, pushed subprime borrowers into severe trouble and the subsequent bad performance on subprime loans enhanced banks’ credit risk (Brauneis & Stachowicz, 2007).

As subprime lending is a high-cost lending, the interest rates for subprime loans are also higher than those for prime loans. The difference is the subprime premium.

(Chomsisengphet & Pennington-Cross, 2006). Furthermore, most subprime loans are adjustable-rate mortgages in order to compensate lenders for higher risks. The future rate adjustments can lead to a significant increase in monthly payment (G Bhingarde et al. 2007). Once the interest rates of subprime loans go up, many borrowers can have unexpected trouble to repay and subsequently delinquencies increase (Neal, 2008).

Between 2000 and 2005, the delinquency rate increased from 0.57% to 0.81% (G Bhingarde et al. 2007). According to Chomsisengphet and Pennington-Cross’s research (2006), the possibility of delinquency is approximately six times higher for subprime loans than prime loans. As a measure of mortgage loans performance, high delinquency rates also influence the credit risk of banks.

In addition, besides some macroeconomic factors, there are also some bank level factors that influence LLA directly. In Hasan and D.Wall's study (2004), they explain the determinants of loan loss allowance and also provide an econometric model to estimate LLA in banks. Depending on their research results, variables as the ratio of nonperforming loans to total assets, the ratio of net charge-offs over year to total assets, the ratio of total loans to total assets and the ratio of net income before taxes and loan loss provisions to total assets are significantly related to the value of LLA.

Since this paper estimates bank's credit risks on state level, relevant other research indicates that the degree of branching restrictions also contributes to the level of credit risks that commercial banks possess. Between states more open to branching and other less open states, the difference in state's branching restrictions affects credit supply. The Interstate Banking and Branching Efficiency Act (IBBEA) (1997) makes banks expansion possible across state lines. However, the variation in state adoption rates of interstate branching after 1997 still influences the credit supply. Because, although IBBEA made states much more open, it allowed states set regulations by themselves on interstate branching. In Strahan's research (2008), he grouped states by no restrictions, moderate restrictions and the highest restrictions. The greater restrictions the states have the fewer interstate branches in those states. Based on this, in this paper I add another factor as a control variable-number of banks per state to measure this distinction among states.

There is some previous research that has started to estimate banks’ risks by observing both macroeconomic factors and bank level indicators. However, not much attention has been paid on the influence of the fluctuation of house prices and homeownership rate on credit risk of commercial banks. Based on previous research, dramatic changes of house prices is one of the major reasons leading to the current severe economic

problems in the United States. House prices show dramatic different amounts of growth in different states (C. Wheaton & Gleb Nechayev, 2008). Also, due to the previous explosive growth and current apparent decrease of home values, the homeownership level increased to a historical high level of 69.2% in 2004 (A.Moore and J.Brauneis, 2008). According to Chomsisengphet and Pennington-Cross’s research (2006), homeownership is a significant economic factor. Higher homeownership level may represent that financial institutions originate more loans to borrowers on the mortgage market. The national rate of homeownership fluctuated slightly from 62% to 64% between 1965 and 1995. However, since 1995 it increased 5% and in 2005, it went up to 69%. Along with the increasing homeownership rate, the total number of renters in the United States decreased for the first time since the Second World War (C.Wheaton & Gleb Nechayev, 2008). Furthermore, the homeownership rate also varies by geographical area. Therefore, in this paper, I estimate the credit risks of commercial banks by using variables announced by state, individual bank and the Federal Reserve. I use the house price index and homeownership rate per state to show the influence of subprime mortgages on the credit risk of commercial banks.

Hypothesis Hypothesis

HypothesisHypothesis 1111: As for factors announced by state, house price index, homeownership rate, per capita disposable income and the level of state's branching restriction may influence the credit risk of commercial banks.

Hypothesis 1a: House price index per state has a positive effect on the ratio of loan loss allowance/total loan in commercial banks. i.e. β₁>0.

Hypothesis 1b: Homeownership rate per state influences the ratio of loan loss allowance/total loan negatively. i.e. β₂<0.

Hypothesis 1c:The number of banks per state indicates that the degree of branching restrictions, which is expected to have a negative impact on the ratio of loan loss allowance/total loan in commercial banks. i.e. β₃<0.

Hypothesis 1d:Per capita disposable income per state influences the ratio of loan loss allowance/total loan positively. i.e. β₄>0.

Hypothesis Hypothesis

HypothesisHypothesis 2222: For factors announced by an individual bank, the ratio of total loan to total assets and non-performing loans to total assets may have impact on credit risk of commercial banks.

Hypothesis 2a: The ratio of total loan to total assets is expected to have a positive impact on the ratio of loan loss allowance/total loan. i.e. β ₅>0.

Hypothesis 2b:Non-performing loan/total assets is expected to have a positive impact on the ratio of loan loss allowance/total loan. i.e.β₆ >0.

Hypothesis Hypothesis

HypothesisHypothesis 3333: The factor announced by the Federal Reserve, short-term interest rate (federal funds rate) is related to the level of credit risk of commercial banks and it has a negative impact on the ratio of loan loss allowance to the total loans. i.e. β₇<0.

2.2 The economic model

The economic model that I will use is based on the methodology used by Hasan and D.Wall (2004) for their study about factors affecting the loan loss allowance of banks in the U.S. In order to avoid a multicollinearity problem, a correlation matrix (as shown in table 2) is applied to assess the correlation between variables. In the case of a high correlation, one variable is kept in the model. Consequently, the data set used in this study incorporate information about loan loss allowance/total loan, house price index, homeownership rate, number of banks, total loans/total assets, non-performing loans/total assets and federal funds rate with respect to 49 states (except Missouri because of missing data) in the United States for the period of 1997-2007. And for my research, I estimate the following regression model by using my panel data set:

Log(LLA )= β0+ β1log(HPI )+ β2log(HOSR )+ β3log(NOB )+_{i,t i,t i,t i,t}

β4log(PCDI_i,t)+ β5log( LOAN ) +β6log( NPL_{i,t i,t}) +β₇log(FFR )+_t

∑

e

(1)

Where i is the state index and t is the year index. I also include year dummy T . The_t dependent variable is credit risk measured as the loan loss allowance/total loan of commercial banks. Since I will do my research on state level, particularly aim to assess whether the house prices have a significant influence on the credit risk of commercial banks. Besides these explanatory variables in Hasan and D.Wall's equation, I add several other variables which may also influence the value of loan loss allowance/total loan of commercial banks. In the equation:

Factors announced by state:

HPI_i,t= House Price Index per state in the U.S

HOSR_i,t= Homeownership Rate per state in the U.S

NOB_i,t= Number of banks per state in the U.S (as a control variable)

PCDI_i,t= Per capita disposable income

Factors announced by individual bank:

LOAN_i,t= the ratio of total loans to total assets in state i at time t

NPL_{i ,t}= the ratio of non-performing loans to total assets in state i at time t

Factors announced by Federal Reserve:

FFR = FED Funds Rate (short-term interest rate) in the U.S_t

3.

3.

3.3. DataDataDataData SourcesSourcesSourcesSources andandandand SampleSampleSampleSample DescriptionDescriptionDescriptionDescription

The data will be estimated and analyzed for a panel of cross-state (in the U.S) observations over the period 1997 – 2007. In other words, it consists of a dataset among 49 states in the U.S.

3.1 Dependent variable

As stated in the Introduction, this research uses the ratio of LLA to total loan as the dependent variable. This dependent variable is reported by the Federal Reserve Bank of Chicago. The total sample coverage consists of 49 states in the United States for the period 1997-2007.

3.2 Independent variable³

Announced by state

In the model I have four explanatory variables which are announced by state. They are house price index, homeownership rate, number of banks and per capita disposable income. In this section, more information on these variables will be provided in detail.

The house price index data (HPI _i,t ) is collected from the U.S Office of Federal Housing Enterprise Oversight (OFHEO), which is a quarterly broad measure of the movement of single-family house prices. The HPI is a weighted, repeat-sales index, which measures average price changes in repeat sales or refinancing on the same properties⁴. As my research focuses on annual changes instead of quarterly, the HPI in my dataset are the averages of every four quarters for 11 years from 1997 to 2007.

With regard to the homeownership rate, I obtained the data of homeownership rate (HOSR_i,t) per state from the U.S Census Bureau⁵. Number of banks per state per year

3 For a synthesis of the independent variables see table 4 in the appendix.

4 Takes the value of 1^stquarter 1980 as the basic value.

5 which is calculated by the following equation: Homeownership rate=owner households/total occupied

(NOB_i,t ) is collected from the Federal Reserve Bank of Chicago. Because the number of banks sometimes changes quarterly, I choose to use the number of banks in every first quarter of every year - in order to be consistent. Per capita disposable income

6(PCDI_i,t) is defined as the total disposable personal income divided by total midyear population estimates of the Census Bureau. These population estimates were released by the Census Bureau in December 2007. The PCDI data comes from the Bureau of Economic Analysis (BEA).

Announced by individual bank

I note that all the data of independent variables which are announced by individual banks come from the Federal Reserve Bank of Chicago. I have two variables of this kind. One is the ratio of total loans to total assets (LOAN_i,t), and the other one is the

ratio of non-performing loans to total assets (NLP_i,t). The total loan here is the net of unearned income⁷. And the total assets are the sum of all assets items equal to the sum of the total liabilities, the limited-life preferred stock and the equity capital. The non- performing loans here are computed as loans 90+ days late, including loans and lease financing receivables on which payment is due and unpaid for 90 days or more. Also includes all restructured loans and leases.

Announced by Federal Reserve

The Federal funds rate (FFR ) is a short-term interest rate (always overnight) at which_t unsecured loans are made between banks. It is a policy instrument of the Federal Reserve and is also released by the Federal Reserve.

3.3 Dummy variable

I also include time dummies which are represented by T . In this paper, I do the_t research on 49 states of the United States for 11 years from 1997 to 2007. However, when I tried to introduce state dummies into my model, the “near singular matrix”

problem occurred which might because the state dummy has a similar trend with the year dummy. According to my data set, the performance difference between my samples is larger from year to year. Furthermore, note that using too many dummies

6 It is measured in thousands of dollars.

(48 dummy variables for 49 states)⁸ may result in a statistic problem of low degrees of freedom and hence leads to a loss in the power of statistic tests. A model with a lot of dummies will also have other problems such as multicollinearity, heteroskedasticity and autocorrelation. In order to solve this problem and create a model with fewer dummy variables, I decide to use only year dummies in my model.

3.4 Choosing a methodology

According to certain econometric literature, there are several techniques which can be applied to analyze panel data, which are: pooled OLS, the fixed effect model, the random effect model and the dynamic models. Depending on the panel data I have, and also because of I do not have lagged variables, a dynamic model is not necessary. And I should choose a methodology between the fixed effect model and the random effect model. Whether it is better to use the fixed effects or the random effects model, according to previous research, can be tested by the Hausman specification test (see output 4). The result of the Hausman test shows p-value=0, which means that the random effects are correlated with the independent variables. In other words, the assumption of the random effect model is not satisfied. Furthermore, according to the previous studies, in generally, no matter whether or not we can measure all of the time- invariant variables possibly correlated with the other regressors, it is better to rely on the fixed effects model. Because the main assumption of the model is often violated leading to a biased random effect estimator. Hence, in this research paper, I decide to use the fixed effect model as the methodology.

44

44．DataDataDataData analysesanalysesanalysesanalyses

After estimating a regression equation, I will proceed by checking the assumptions made about the random variables in the model. The statistical inference consists of the following two steps: descriptive statistics and diagnostic checks with respect to the data. All this information will be presented in this section due to the fact that my data set is adjusted to meet the fundamental assumptions of the estimation technique to obtain more reliable results.

4.1 Descriptive statistics

Descriptive statistics simply provide the basic characteristics of the data, such as their mean (expected value, average value of the random variable in repeated samples) (Hill et al, 2001), median (middle value of the series), maximum and minimum value of the series,

the standard deviation (dispersion or spread of the series), skewness (how symmetric the residuals are around zero) and Kurtosis (the “peakedness” of the distribution) (Hill et al, 2001). Shown in table 1.

8 Omitting one dummy variable which is named a reference group in order to avoid the problem of exact

Table Table TableTable 1111

4.2 Diagnostic checks

Diagnostic checks consist of testing for multicollinearity, heteroskedasticity, nonstationarity, autocorrelation and model specification. They are used to test whether the relevant multiple regression assumptions hold.

Normal Distribution:

Since all the other tests are based on the assumption that all the errors are normally distributed, I first run the residual normality test. Consequently, the Jarqure-Bera χ²

critical value is calculated in Eviews. If the Jarqure-Bera value is higher than χ² , it means that residuals are normally distributed. (see figure 2)

Multicollinearity:

It is assumed that the independent variables are not too strongly collinear, indicating that there is no multicollinearity problem. Multicollinearity problem makes it difficult to isolate the effects of individual variables. According to Hill et. al (2001), the collinear relationships can be tested by applying sample correlation coefficients between pairs of explanatory variables. I will use a correlation matrix to show whether multicollinearity exists. The values range between 0 and +/- 1. Values close to 0 indicate a weak linear relationship, while value closer to +/- 1 show a strong positive (negative) linear relationship between two explanatory variables. When values are higher than +0.7 or lower than -0.7, it indicates that multicollinearity exists.

(see Table 2 in appendix)

LLA HPI HOSR NOB LOAN NPL PCDI FFR

Mean 0.018 275.4 69.73 159.25 0.644 0.002 26572.30 3.933

Median 0.014 244.3 70.60 100.00 0.659 0.002 26048.00 4.970

Maximum 0.727 720.2 81.30 866.00 0.955 0.020 45217.00 6.240

Minimum 0.006 130.7 50.20 5.00 0.217 0.00003 16733.00 1.130

Std. Dev. 0.032 105.4 5.19 161.92 0.095 0.003 4997.213 1.738 Skewness 20.450 1.606 -1.076 1.7961 -1.073 3.212 0.586 -0.503

Kurtosis 452.70 5.684 4.30 6.5066 6.566 15.630 3.220 1.743

Observations 539 539 539 539 539 539 539 539

Heteroskedasticity:

Heteroskedasticity exists when the variances for all observations are not the same. This problem is often encountered when using cross-sectional data (Hill et. al, 2001). I will use the White test to investigate the existence of heteroskedasticity. When the p-value of the White test is higher than 0.05, there is no heteroskedasticity. Furthermore, if there is a heteroskedasticity problem, a White estimator can be used to overcome it.

(see output 3 in appendix)

Autocorrelation:

Autocorrelation is a natural problem with time series data. It exists when the error terms are correlated. Both Durbin-Watson (DW) statistic and the LM test can be used to test autocorrelation⁹. When DW 2, it indicates that there is no autocorrelation in the≈ model. Since I do not have lags or differences in my test equation, in this research paper the Durbin-Watson test can be used for testing the autocorrelation problem. In addition, for the LM test, when the p-value of LM test is lower than 0.05 (significant at 5%), autocorrelation exists. In order to adjust a model with autocorrelation problem, we can add AR(1) or AR(2) in our model.

Nonstationary:

Stationary series are stable series, whose mean and variance do not change much as time passes. In econometrics, nonstationary series can lead to trouble with estimation and interpretation of the results. I will perform a Unit Root Test (an Augmented Dickey-Fuller test) to test nonstationary. (see output 1 in appendix) When the p-value of the ADF test statistic is smaller than 1% critical value, the null hypothesis will be rejected and it means the tested variable has no unit root. Hence, the model passes the ADF test.

Model Specification:

Whether the model is adequate or we still need to improve, depends on testing for model misspecification. The Ramsey RESET Test (Regression Specification Error Test) can be applied for testing this problem. (see output 2 in appendix) If the p-value

>0.05 (significant at 5%), it indicates that there is no model misspecification. (Hill et.

al, 2001)

4.3. Results of diagnostic checks

Table:

Table:

Table:Table: TheTheTheThe resultsresultsresultsresults ofofofof thethethethe validityvalidityvalidityvalidity checkscheckscheckschecks

NoteNote

NoteNote: For the complete tables, see output1-output4 in appendix.

In order to check whether there is a model misspecification problem or not in a particular model, I rely on the Ramsey RESET Test. To detect the multicollinearity problem, a correlation matrix between independent variables is shown. Concerning the assumption of homoskedasticity of the error terms, I use the White test to see if the error terms fulfill this assumption. Furthermore, I also include the result of the Durbin-Watson test to assess whether there is an autocorrelation problem or not. ADF test results are presented to assess the stationarity of the residuals.

When I look at the nature of my data set and run all the tests without any adjustments, it is shown that none of the assumptions are fulfilled except that the errors are normally distributed (see Figure 1) and the variables are stationary (see output 1). In this case, the results I obtain will be highly biased and not strongly reliable. In order to improve my model and gain reliable results, several options with regard to the transformation of the variables have been attempted. Since I have eight continuous variables (LLA, HOSR, HPI, PCDI, NOB, FFR, LOAN, and NPL) and the rest are dummy variables, the transformation of the variables can only occur within the first group. All the variables are transformed into the natural log-series and hence I finally have the log- log model.

As shown in the above table, the p-value of the Ramsey RESET Test is 0.74 which is higher than 0.05 (significant at 5%), it means that there is no model misspecification anymore. Based on the correlation matrix (see table 2), in order to avoid the multicollinearity problem, I keep seven independent variables and delete others who create multicollinearity. However, among these seven variables, the correlation value between the house price index and per capita disposable income is 0.77. Although this value is quite high, at least one macroeconomic factor is included in my model to make the model more reliable and reasonable. Furthermore, according to Hill et. al (2001), a commonly used rule of thumb is that a correlation coefficient between two explanatory variables greater than 0.8 or 0.9 in absolute value indicates a strong linear association and a potentially harmful collinear relationship. The value 0.77 is still lower than 0.8, so I keep PCDI in my model.

According to the result of the White Test, the p-value of White Test is 0.000 and hence there is a heteroskedasticity problem in the model, so a White’s estimator should be used to overcome it. Adding AR (1) in the equation leads to DW≈2, which means the assumption of no autocorrelation is satisfied.

Durbin-Watson Durbin-Watson Durbin-WatsonDurbin-Watson statistic

statistic statisticstatistic

Hausman Hausman Hausman Hausman testtesttesttest (p-value) (p-value) (p-value) (p-value)

White White White White TestTestTestTest (p-value) (p-value) (p-value) (p-value)

Ramsey Ramsey Ramsey Ramsey TestTestTestTest (p-value) (p-value) (p-value) (p-value)

2.109 0.000 0.000 0.735

5.

5.

5.5. RegressionRegressionRegressionRegression resultsresultsresultsresults

Table Table TableTable 3333

Note:

Note:

Note:Note: “***” indicate significance at the 1% level.

The results are obtained by using Eviews 5.0.

As can be seen clearly from the table above, the results of estimating the model for the commercial banks in the United States show that all the coefficients on the explanatory variables are statistically significant. Moreover, the coefficient for the control variable NOB also appears to be plausible. Table 3 shows the impact of the seven independent variables on the ratio of loan loss allowance/total loan of commercial banks. From the coefficient and the corresponding p-value of the variable HPI, house price index has a positive and significant impact on the ratio of loan loss allowance/total loan as anticipated in the hypothesis 1a. When HPI increases by 1 unit,

LLA/total loan increases by 0.041 units. Hypothesis 1b is also proved by the results, homeownership rate has a negative impact on this dependent variable. The ratio of LLA/total loan decreases by 0.298 units, if HOSR increases by 1 unit. Moreover, both the number of banks and the federal funds rate have negative impact on the value of LLA/total loan, which means that hypothesis 1c and hypothesis 3 are accepted. The value of LLA/total loan increases by approximately 0.321 units and 0.084 units, if the federal funds rate and the number of banks decrease by 1 unit respectively.

As shown in table 3, per capita disposable income and non-performing loan/total assets have a positive impact on the dependent variable as predicted in hypothesis 1d

Variable

Coefficient

(p-value) Std. Error t-Statistic

LOG(FFR)

-0.321***

(0.000) 0.011 -28.25887

LOG(HOSR)

-0.298***

(0.000) 0.053 -5.642593

LOG(HPI)

0.041***

(0.002) 0.013 3.076638

LOG(LOAN)

-0.344***

(0.000) 0.028 -12.06619

LOG(NOB)

-0.084***

(0.000) 0.003 -28.93210

LOG(NPL)

0.230***

(0.000) 0.004 62.99903

LOG(PCDI)

0.196***

(0.000) 0.026 7.394878

C

-3.093***

(0.000) 0.378 -8.172978

Adjusted R-squared 0.795 Durbin-Watson stat 2.11

and hypothesis 2b. The value of LLA/total loan increases by around 0.196 and 0.230 units, if the per capita disposable income and the value of non-performing loan/total assets increase by 1 unit respectively.

What is interesting and somewhat puzzling is that hypothesis 2a should be rejected.

The effect of the ratio of total loan/total assets on the dependent variable is significant and unexpectedly negative. The regression result shows that the value of LLA/total loan goes down when the value of total loan/total assets goes up. I will give two possible reasons of this unexpected result in the next section.

6.

6.

6.6. ConclusionConclusionConclusionConclusion

This paper examines the effect of combinations of seven factors (HPI, HOSR, PCDI, NOB, LOAN, NPL, and FFR) on the credit risk of commercial banks in the United States. In this study, the credit risk of commercial banks is measured by the ratio of loan loss allowance to total loan. One econometric model is specified and estimated in order to study the relationship between explanatory and dependent variables. The research is performed utilizing a panel data set consists of repeated observations on 49 individual states of the United States for the period 1997-2007. Furthermore, in this paper, besides the other five factors which have been estimated in previous research, I include the house price index and homeownership rate in my model. These two are also quite related to the current subprime crisis in the U.S.

The regression results provide the answer to the main question of this paper. House price index has a positive impact on the credit risk of commercial banks. The possible explanation for this result is when house prices increase, borrowers seldom have trouble with paying back the money to the banks. Even if some of them fail to return the money, banks still have their houses which values are increasing. Banks expect no losses but actually they may suffer from potentially higher credit risks because of this optimistic expectation. Shown as the regression results, all the hypotheses are accepted except hypothesis 2a. An unexpected coefficient sign shows that the ratio of total loan to total assets has a negative impact on the credit risks of commercial banks.

According to experience in practice, this is not plausible. With regard to the estimation results, the two following reasons probably can explain why this result runs counter to the expectation. First, the total loan of commercial banks consists of not only the real estate loans but also other kinds of loans, like agricultural loans, commercial and industrial loans, loans to individuals and lease financing receivables.

The diversity of loans can lower the credit risks of commercial banks by dispersing credit risks. So when banks have larger amounts of loans, perhaps because they have more kinds of loans, namely that the diversity of loans is larger. Second, the total loans in commercial banks also consist of good loans and bad loans. When the percentage of good loans increases, the credit risk of commercial banks decrease as it is shown in my result. Unfortunately, I do not have relevant data which can show the different percentages of good/bad loans of the total loans. Hence, this could be a

limitation of this research and could be a trigger for future studies.

An additional limitation of this research is with regard to the dependent variable.

Since there are several other proxies that can measure bank risks, such as Return on Assets (RoA), Loan to total assets, Non-performing Assets, Net charge-offs and Leverage. However, in this research I use only loan loss allowance to total loan. In addition, although the credit rating of borrowers should be also a significant factor which can influence the credit risks of commercial banks, it is difficult to obtain the data. Results from this estimation should be taken with caution.

The result of this research, together with the ongoing subprime crisis in the U.S, has a number of implications and suggestions for managers of commercial banks. When managers attempt to measure the credit risk of their banks and determinate the value of loan loss allowance to total loan, they should pay attention to all the seven influential factors in my model. All of them are statistically significant. Among these seven factors, FFR, HOSR, LOAN and NOB have negative impact on the ratio of LLA to total loan, while HPI, NPL and PCDI influence the value of LLA to total loan positively. Together with the severe subprime crisis in the U.S, the fluctuation of house prices should be regarded as an important factor that can influence the credit risk of commercial banks. Especially when the housing market turns down, banks may suffer more credit risks.

This research also leaves space for relevant future research. Because of the severe subprime crisis and financial crisis occurring in the U.S from 2007, trust is also an important factor that has impact on the credit risk of banks, including not only the trust between consumers and banks but also among banks. Although it is difficult to measure trust quantitatively, it is interesting to include it as another explanatory variable.

References:

References:

References:References:

Altunbas, Yener; Carbo, Santiago; Gardener, Edward P. M.; Molyneux, Philip.

"Examining the Relationships between Capital, Risk and Efficiency in European Banking"European Financial Management, (2007), Vol. 13 Issue 1, p49-70.

Anthony Saunders; Marcia Million Cornett. "Financial Institutions Management (A risk management approach)"(2007).

Bhingarde, Nikhil G.; Khasseria, Harpreet S.; Yellavalli, Balaji. "The Subprime Mortgage Market: Current State and the Road Ahead." Bank Accounting & Finance (08943958), (2007), Vol. 20 Issue 6, p3-10.

Brauneis, Michael; Stachowicz, Steven. “Subprime Mortgage Lending: New and Evolving Risks, Regulatory Requirements” Bank Accounting & Finance(08943958),

(2007), Vol. 20 Issue 6, p28-34..

Chomsisengphet, Souphala; Pennington-Cross, Anthony. "The Evolution of the Subprime Mortgage Market." Review(00149187), (2006), Vol. 88 Issue 1, p31-56.

Cole, Rebel A; Jeffery W. Gunther; "Predicting bank failures: A comparison of on- and off-site Monitoring Systems."Preview Journal of Financial Services Research, (1998), Vol. 13 Issue 2, p103.

Donald R. Haurin; "Income Variability, Homeownership, and Housing Demand"

Journal of Housing Economics,(1991) p60-74.

Hasan, Iftekhar; Wall, Larry D. "Determinants of the Loan Loss Allowance: Some Cross-Country Comparisons" Financial Review, (2004), Vol. 39 Issue 1, p129-152.

Haurin, Donald R.; Gill, H. Leroy. "Effects of Income Variability on the Demand for Owner-Occupied Housing" Journal of Urban Economics, (1987), Vol. 22 Issue 2, p1336.

Hsiao, Cheng. "Analysis of Panel Data." Cambridge University Press, (1986).

Irina Barakova; Raphael W. Bostic; Paul S. Calem; Susan M. Wachter; "Does credit quality matter for homeownership?"Journal of Housing Economics, (2003), p318-336.

Jack Johnston; John Dinardo; "Econometric Methods" fourth edition (1997).

Kolari, Jame; Glennon, Dennis; Hwan Shin; Caputo, Michele. "Predicting large US commercial bank failures." Journal of Economics & Business, (2002), Vol. 54 Issue 4, p361.

Krishnan, C. N. V.; Ritchken, P. H.; Thomson, J. B. "On Credit-Spread Slopes and Predicting Bank Risk." Journal of Money, Credit & Banking, (2006), Vol. 38 Issue 6, p1545-1574.

Luis Diaz-Serrano; "Labor income uncertainty, skewness and homeownership: A panel data study for Germany and Spain"Journal of Urban Economics(2005) p156-176.

Michelle Clark Neely; David C. Wheelock. "Why does bank performance vary across states?"Review (00149187), (1997), Vol. 79 Issue 2, p27.

Moore, Matthew A.; Brauneis, Michael J. "U.S. Subprime Crisis: Risk Management's Next Steps." Bank Accounting & Finance (08943958), (2008), Vol. 21 Issue 3, p18- 48.

Neal, Penny. "The subprime mortgage crisis: lessons for regulators"Policy, (2008), Vol.

24 Issue 2, p19-25, 7p.

Robert Oshinsky; Virginia Olin. "Troubled banks: why don't they all fail?" working paper (2005).

Shiers, Alden F. "Branch banking, economic diversity and bank risk"Quarterly Review of Economics & Finance, (2002), Vol. 42 Issue 3, p587, 12p.

Staikouras, Sotiris K. "Multinational Banks, Credit Risk, and Financial Crises"

Emerging Markets Finance & Trade, (2005), Vol. 41 Issue 2, p82-106.

Sullivan, Richard J.; Spong, Kenneth R. "Manager wealth concentration, ownership structure, and risk in commercial banks" Journal of Financial Intermediation, (2007), Vol. 16 Issue 2, p229-248.

Tara Rice; Philip E. Strahan; "Does credit supply affect small-firm finance"

http://ssm.com/abstract=1107562.

Wallace, Jeffrey; Avis, Mary Ashli; Smith, Stephen C. "The Credit Crunch: A Domino Effect."Business Perspectives, (2008), Vol. 19 Issue 2, p58-63.

Wheaton,William C.; Nechayev, Gleb. "The 1998-2005 Housing "Bubble" and the Current "Correction": What's Different This Time?" Journal of Real Estate Research, (2008), Vol. 30 Issue 1, p1-26.

Yaffee, Robert A. "A Primer for Panel Data Analysis." University of New York, Information Technology Services, (2003).

APPENDIX APPENDIX APPENDIXAPPENDIX

Figure Figure FigureFigure 1:1:1:1:

Source:

Source:

Source:Source: OfficeOfficeOfficeOffice ofofofof FederalFederalFederalFederal HousingHousingHousingHousing EnterpriseEnterpriseEnterpriseEnterprise OversightOversightOversightOversight

Figure Figure

FigureFigure 2:2:2:2:NormalNormalNormalNormal DistributionDistributionDistributionDistribution

0 20 40 60 80 100 120

-0.5 -0.0 0.5 1.0 1.5 2.0 2.5 3.0

Series: Residuals Sample 1 539 Observations 539 Mean -1.09e-15 Median -0.017773 Maximum 2.921734 Minimum -0.772540 Std. Dev. 0.270307 Skewness 2.400814 Kurtosis 28.05735 Jarque-Bera 14618.72 Probability 0.000000

Modified:1 Modified:1 Modified:1

Modified:1 539//539//539//539// a=@qchisqa=@qchisqa=@qchisqa=@qchisq (.95,8)(.95,8)(.95,8)(.95,8) 15.50715.50715.50715.507

Table Table

TableTable 2:2:2:2:CCCCorrelationorrelationorrelationorrelation matrixmatrixmatrixmatrix

Output Output

OutputOutput 1:1:1:1:UnitUnitUnitUnit RootRootRootRoot TestTestTestTest (stationary)(stationary)(stationary)(stationary)

FFR HOSR HPI LOAN NCO NOB NPL PCDI RETN USTR

FFR 1.00 -0.11 -0.13 0.10 -0.06 0.04 0.02 -0.21 -0.03 0.79

HOSR -0.11 1.00 -0.21 0.25 -0.02 0.01 0.06 -0.14 -0.02 -0.16

HPI -0.13 -0.21 1.00 -0.23 -0.01 -0.23 -0.12 0.77 -0.05 -0.40

LOAN 0.10 0.25 -0.23 1.00 0.18 -0.02 0.34 -0.20 0.32 0.04

NCO -0.06 -0.02 -0.01 0.18 1.00 -0.19 0.79 -0.05 0.84 0.03

NOB 0.04 0.01 -0.23 -0.02 -0.19 1.00 -0.16 -0.06 -0.25 0.06

NPL 0.02 0.06 -0.12 0.34 0.79 -0.16 1.00 -0.13 0.85 0.09

PCDI -0.21 -0.14 0.77 -0.20 -0.05 -0.06 -0.13 1.00 -0.06 -0.52

RETN -0.03 -0.02 -0.05 0.32 0.84 -0.25 0.85 -0.06 1.00 0.06

USTR 0.79 -0.16 -0.40 0.04 0.03 0.06 0.09 -0.52 0.06 1.00

Null Hypothesis: FFR has a unit root Exogenous: Constant

Lag Length: 0 (Fixed)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -9.576417 0.0000 Test critical values: 1% level -3.442276

5% level -2.866693

10% level -2.569575

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation Dependent Variable: D(FFR)

Method: Least Squares Date: 10/25/08 Time: 01:17 Sample (adjusted): 2 539

Included observations: 538 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

FFR(-1) -0.291844 0.030475 -9.576417 0.0000

C 1.146333 0.130965 8.752949 0.0000

R-squared 0.146099 Mean dependent var -0.000818

Adjusted R-squared 0.144506 S.D. dependent var 1.327560 S.E. of regression 1.227899 Akaike info criterion 3.252197 Sum squared resid 808.1464 Schwarz criterion 3.268137

Log likelihood -872.8409 F-statistic 91.70777

Durbin-Watson stat 1.191790 Prob(F-statistic) 0.000000

Null Hypothesis: HOSR has a unit root Exogenous: Constant

Lag Length: 0 (Fixed)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -5.867569 0.0000 Test critical values: 1% level -3.442276

5% level -2.866693

10% level -2.569575

*MacKinnon (1996) one-sided p-values.

Augmented Dickey-Fuller Test Equation Dependent Variable: D(HOSR)

Method: Least Squares Date: 10/25/08 Time: 01:21 Sample (adjusted): 2 539

Included observations: 538 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

HOSR(-1) -0.120906 0.020606 -5.867569 0.0000

C 8.440958 1.440644 5.859157 0.0000

R-squared 0.060355 Mean dependent var 0.011152 Adjusted R-squared 0.058602 S.D. dependent var 2.553820 S.E. of regression 2.477861 Akaike info criterion 4.656379 Sum squared resid 3290.930 Schwarz criterion 4.672319

Log likelihood -1250.566 F-statistic 34.42837

Durbin-Watson stat 1.871518 Prob(F-statistic) 0.000000

Null Hypothesis: HPI has a unit root Exogenous: Constant

Lag Length: 0 (Fixed)