Z l ‘V | In . 4 an t .‘i :ESlSi MICHIGAN STAT ll will ill Ill/Ill 3 1293 MIL/'55 9432 I Ill/Illllll’l‘ 1" ”ill ' This is to certify that the dissertation entitled Interest Rate Risk and Thrift Capital presented by Robert Crile Wolf has been accepted towards fulfillment of the requirements for Ph . D . degree in Finance /) a. o a Major profesér , Date July 17,-1996fi MS U is an Affirmative Action/Equal Opportunity Institution 0-12771 LIBRARY Michigan State Unlverslty PLACE N RETURN BOX to remove this checkout from your mead. To AVOID FINES mum on or More data duo. DATE DUE DATE DUE DATE DUE MSU I. An Affirmative Adlai/Equal Oppommty Inflation Wanna-9.! Interest Rate Risk and Thrift Capital By Robert C rile Wolf A DISSERTATION Submitted to Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY Department of Finance 1996 ABSTRACT Interest Rate Risk and Thrift Capital By Robert Crile Wolf Accurate and meaningful measurement of interest rate risk (IR) is of critical importance to depository institutions. Recently, Congress has included IRR in the risk evaluation of depository institutions for risk based capital standards. In foresight of this regulation, the Office of Thrift Supervision developed the Market Value Model (MVM) to measure IRR. This study conducts two tests using the MVM IRR. The first test compares the ability of various IRR proxies to predict thrift failure. Besides the MVM IRR measure, this study evaluates two other IRR proxies. In logit regressions including an IRR variable and other control variables, the first test finds each IRR variable significant in predicting failure. Including all three IRR variables shows several variables significant, suggesting the information they provide is additive and multiple measures may provide a superior evaluation of thrift interest rate and default risk. When an interactive product between capital and each IRR measure is included in the previous regressions, each interactive term shows firms with higher capital levels are less sensitive to IRR. Including interactive dummy variables shows, for high capital levels, IR is positively correlated to thrifi health. The second test employs the various IRR proxies in predicting equity returns. First, this is done in two stages. The first stage uses the two index market model and tests whether depository institution stock prices are sensitive to interest rate changes. The second stage tests whether the interest sensitivity of equity is dependent on the IRR proxy. Using two stages, the results for the second test are unsuccessful as thrift equity is found unaffected by interest rate changes. Second, the test is run in a single stage. These results show the interest rate index and the interactive term between IR and the interest rate index insignificant. Possible explanations for insensitivity in both tests are time period and firm sample selection. ACKNOWLEDGEMENTS There are a countless number of people who have made contributions of varying degree. Some have contributed information, others data or technical advice. Without their efforts, I would not have been able to finish my dissertation. I wish to express my sincere appreciation to the members of my dissertation committee - Dr. Robert Rasche, Dr. James Wiggins, Dr. Alan Grunewald, and Dr. Anil Shivdasani, for their guidance, encouragement, and commitment of time and effort on my behalf. I would especially like to express my deepest gratitude to Dr. John O’Donnell for the support he has given me throughout my doctoral program. Finally, I would like to thank all my fiiends and family for their patience and encouragement. All fall short of the glory of God. Rom. 3 :23 (paraphrase) TABLE OF CONTENTS LIST OF TABLES ..................................... . .......................................................... viii CHAPTER 1 -INTRODUCTION .......................................................................... 1 CHAPTER 2 -REVIEW OF RELATED LITERATURE ........................................ 5 2.1 Literature on the Interest Sensitivity of Financial Institution Equity... 5 2.1.1 Theoretical Literature ........................................................... 5 2.1.2 Empirical Literature .............................................................. 7 2.2 Literature on the Failure of Financial Institutions .............................. 22 2.2.1 Failure Prediction with Accounting Data ............................... 23 2.2.2 Loss Prediction with Accounting Data .................................. 27 2.2.3 Problem Status with Accounting Data ................................... 27 2.2.4 Failure Prediction with Market Data ...................................... 28 CHAPTER 3 -THE MARKET VALUE MODEL ................................................. 30 3.1 Discounting Model ........................................................................... 32 3.2 Interest Rate Model .......................................................................... 33 3.3 Account Specific Present Value Methodologies ................................ 37 3.3.1 Asset Methodologies ............................................................ 38 3.3.2 Liability Methodologies ......................................................... 43 CHAPTER 4 -FAILURE PREDICTION MODEL ................................................ 4S 4. 1 Introduction .................................................................................... 45 4.2 Review of the Industry and Related Research .................. . ................. 48 4.2.1 Industry Review ................................................................. 48 4. 2. 2 Summary of Failure Prediction Studies ................................. 51 4.3 Methodology and Data ..................................................................... 54 4.3.] Interest Rate Risk and the Market Value Model .................... 54 4.3.2 Failure Prediction Methodology ............................................ 58 4.3.3 Data ...................................................................................... 61 4.4 Results ............................................................................................ 62 4.5 Interaction Term ............................................................................... 68 4.6 Summary .................................... . ..................................................... 77 CHAPTER 5 -TWO-INDEX MODEL ............ . ....................................................... 8O 5. 1 Introduction ......................... . .......................................................... 80 5.2 Methodology.................................................... ................................ 83 5.3 Summary ............................ _ ............................................................ 86 CHAPTER 6 -CONCLUSION ............. . ................................................................ 90 APPENDIX - MARKET VALUE MODEL FORMULAE..... .................................. 96 ‘7 LIST OF TABLES Table Title 11 12 13 14 15 16 17 18 19 20 21 22 23 Summary of major legislation affecting thrifis Summary of the failure prediction literature Summary of the literature on financial institutions in two-index rnodeb. MVM Methodologies Chart Price sensitivity table on December 31,1985, for fixed rate Federal National Mortgage Association (FNMA) mortgages. Failure Record for Thrifts in months 1 through 12. Failure Record for Thrifts in months 13 through 24. Summary Statistics Logit regression of year one failures onto base CAMEL and IRR variables. Logit regression of year two failures onto base CAMEL and IRR variables. Logit regression of year one failures onto base CAMEL and Dchg variables. Logit regression of year two failures onto base CAMEL and Dchg variables. Logit regression of year one failures onto base CAMEL and Achg vafiabk5. Logit regression of year two failures onto base CAMEL and Achg variables. Logit regression of year one failures onto base CAMEL and ISF variables. Logit regression of year two failures onto base CAMEL and ISF variables. Logit regression of year one failures onto base CAMEL and SHORT vanabkm. Logit regression of year two failures onto base CAMEL and SHORT vafiabkm. Logit regression of year one failures onto base CAMEL and IRR variables except MVPE / market is used for capital. Logit regression of year two failures onto base CAMEL and IRR variables except MVPE / market is used for capital. Logit regression of year one failures onto base CAMEL and IRR variables except MVPE / total assets is used for capital. Logit regression of year two failures onto base CAMEL and IRR variables except MVPE / total assets is used for capital. Logit regression of year one failures onto base CAMEL and IRR variables except exp (net worth / total assets) is used for capital. vi Page 107 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 24 25 26 27 28 29 3O 31 32 33 34 35 36 37 38 39 4O 41 42 Logit regression of year two failures onto base CAMEL and IRR variables except exp (net worth / total assets) is used for capital. Logit regression of year one failures onto base CAMEL and IRR variables except other real estate / total assets is used for asset quality. Logit regression of year two failures onto base CAMEL and IRR variables except other real estate / total assets is used for asset quality. Logit regression of year one failures onto base CAMEL and IRR variables except percentage change of median real estate is used for asset quality. Logit regression of year two failures onto base CAMEL and IRR variables except percentage change of median real estate is used for asset quality. Logit regression of year one failures onto base CAMEL and IRR variables except net income / total assets is used for earnings. Logit regression of year two failures onto base CAMEL and IRR variables except net income / total assets is used for earnings. Logit regression of year one failures onto base CAMEL and IRR variables except broker deposits / total assets is used for liquidity. Logit regression of year two failures onto base CAMEL and IRR variables except broker deposits / total assets is used for liquidity. Logit regression of year one failures onto base CAMEL and IR variables except cash and securities / total assets is used for liquidity. Logit regression of year two failures onto base CAMEL and IRR variables except cash and securities / total assets is used for liquidity. Logit regression of year one failures onto optimal CAMEL and IRR variables. Logit regression of year two failures onto optimal CAMEL and [RR variables. Logit regression of year one failures onto base and optimal CAMEL and all three IRR variables. Logit regression of year two failures onto base and optimal CAMEL and all three IRR variables. Logit regression of year one failures onto base CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Logit regression of year two failures onto base CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Logit regression of year one failures onto optimal CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Logit regression of year two failures onto optimal CAMEL variables, each IRR variable, and an interactive product between capital and IRR. vii 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 43 44 45 46 47 48 49 50 51 Logit regression of year one failures onto base CAMEL variables, each IRR variable, and a series of interactive terms which include sequential dummies for capital levels. Logit regression of year one failures onto base CAMEL variables, each IRR variable, and a series of interactive terms which include sequential dummies for capital levels. Logit regression of year one failures onto base or optimal CAMEL and [RR variables and a series of interactive terms which include sequential dummies for asset levels. Logit regression of year one failures onto base or optimal CAMEL and IRR variables and a series of interactive terms which include sequential dummies for asset levels. Logit regression of year one failures onto base CAMEL and [RR variables and a series of interactive terms which include sequential dummies for IRR levels. Logit regression of year one failures onto base CAMEL and IRR variables and a series of interactive terms which include sequential dummies for IRR levels. Logit regression of year one failures onto all CAMEL and IRR variables, a variable for hedging activities, and a size variable. Results of the two index model with various portfolios of S&Ls as the independent variable. Results of the single stage model. viii 152 153 154 155 156 157 158 159 160 CHAPTER 1 - INTRODUCTION Measurement of interest rate risk is of critical importance to depository institutions (DIS), as a result of their long-term mortgages and short-term deposits. During the early and mid 1980’s, interest rates were high and volatile which created an unfavorable environment for thrifis. Consequently, many thrifts failed because of excessive interest rate risk. Lately. risk based capital standards have received significant attention. Interest rate risk (IRR) is now included in the risk evaluation of the firm and used to compute adjustments to capital standards. Historically, IRR measurements have been based on simple balance sheet ratios. Recently, due to more frequent thrifi failure and emphasis on IR based capital standards, new measurement techniques, in particular the Market Value Model, have been developed and employed. As of yet, these methods have not been examined for their comparative accuracy or informational disclosure. This study tests the ability of the Market Value Model to outperform simple balance sheet ratios in measuring interest rate risk and predicting changes in thrifi capital. Previous tests use simple balance sheet ratios to proxy for interest rate risk. The proxy. SHORT, is the difference between the amount of assets and liabilities maturing in one year, standardized by equity value. Flannery and James (1984) find SHORT correlated to the interest rate sensitivity of equity values. Barth et a1. (1985), successfully use the ratio of interest sensitive funds to total liabilities (ISF) in failure prediction. Each of these measures has been used successfully in various studies. but none account sufficiently for the unique characteristics of a thrift's asset and liability portfolio. Regulators of banks and thrifis, realizing the potentially severe consequences of IR, have developed more comprehensive risk measures. These measures not only 1 2 implement a very detailed computation of the maturity and coupon of portfolio securities, but also analyze any option characteristics of these securities, such as prepayments. The Office of Thrift Supervision (OTS) developed the Market Value Model (MVM) especially for the purposes of measuring interest rate risk and determining risk based capital standards. The MVM measures IRR as the change in the market value of portfolio equity (MVPE) resulting from changes in interest rates. For regulatory purposes, the MVM uses 200 basis point parallel shifts upward and downward resulting in two IRR measures, referred to here as Uchg and Dchg, respectively. In conjunction with an increasingly comprehensive interest rate risk proxy. the OTS has required significant increases in the disclosure of financial information. Section H has been included in the Thrift Financial Report since 1985 and separates balance sheet accounts and account yields into seven maturity or repricing “buckets”. This research uses two models to test the effect of IRR on thrift capital and thrift equity returns. Bovenzi et a1. (1983) and Earth et al. (1985) use the first model that shows interest rate risk to be a significant predictor of financial institution failure. Theoretically, failure is when market net worth is equal to zero and therefore indicates capital depletion. In practice, failure depends on the book value of net worth as well as other factors. The failure prediction methodology shows which firm characteristics are related to firm failure and therefore which characteristics are associated with the depletion of net worth or capital value. Comparing various IRR measures in a failure prediction model will show which measure is a better indicator of probable failure. Consequently, 3 this comparison will show which IRR variable(s) would be the most useful with risk based capital standards. The second model is a two index market model that measures the effect on thrift stock price from a changing market and interest rate index. Several tests (Lynge and Zumwalt (1980), Sweeney and Warga (1986), and others) have shown a relationship between the returns of financial institutions‘ common stock and interest rate movements. Kane and Unal (1988), Kwan (1991) and Yourougou (1990) have extended this research by showing that the sensitivity of thrift capital to interest rate changes is not constant. Evidence that the interest sensitivity of stock price is related to maturity mismatch is found by Flannery and James (1984) and Scott (1986). Showing that equity value is related to maturity mismatch confirms the theory that thrift equity is interest rate sensitive. The IRR proxy that most effectively predicts stock price sensitivity should be the best measure of IR and thrift default risk. Therefore this proxy should be the best proxy for risk based capital standards. Stock price movements may be a better measure of changing capital values than failure because common stock prices measure market values, while failure partly depends upon book values. Further, stock price movement is measured more frequently and thus should be a more powerful indicator than failure. This study proceeds in the following order. The next chapter reviews the related literature. This review begins with a discussion of the literature analyzing the interest sensitivity of financial institution’s common stock. This includes a review of the relationship between equity values and an interest rate index, the relationship between the interest index and a maturity mismatch proxy, and the stationarity of the significance of the interest rate index. The review continues by recalling the literature on financial institution 4 failure, with attention given to accounting versus market predictive variables and problem status as the dependent variable. Further, several studies showing the significance of interest rate risk as a predictor of failure will be reviewed. The third chapter describes the Market Value Model as a measure of interest rate risk. This chapter details the discount, interest rate forecasting, and present value methods used in the model. The research methodology and data for the failure prediction model are explained in the fourth chapter. Chapter five presents the methodology and results for the two-index model. Concluding remarks are in Chapter six. CHAPTER 2 - REVIEW OF RELATED LITERATURE 2.1 Literature on Interest Sensitivity of Financial Institution Equity Risk based capital regulation presumes a relationship between interest rate risk and capital risk. Empirical tests have shown stock returns are sensitive to changes in interest rates, particularly stocks in the financial and utility industries. This section continues with a review of the theoretical explanations and empirical studies of this phenomenon. 2.1.1 Theoretical Literature The most prevalent of the theoretical explanations is the nominal contracting hypothesis.‘ This hypothesis divides firm assets into categories of real and nominal. Nominal assets are those whose cash flows are defined in nominal terms and do not adjust to changes in inflation. Cash flows of real assets do adjust to inflationary changes. This hypothesis maintains that common stock returns of firms holding nominal assets will be affected by unanticipated inflation and unanticipated changes in expected inflation. F ama (1975,1976), Fama & Gibbons (1982), and Nelson & Schwert (1977) suggest movement in the term structure of interest rates results primarily from changes in inflationary expectations. Therefore the nominal contracting hypothesis implies a relationship between stock returns and interest rate changes. Stockholders of firms with fewer nominal assets than nominal liabilities should benefit from unexpected increases in inflation. Merton (1973) and Long (1974) imply that if a risk averse investor is choosing between two portfolios giving the same distribution of future wealth but which have different covariances with interest rates, the investor will select the portfolio that gives 1 French et a1. (1983) introduce the nominal contracting hypothesis. U1 6 them a better hedge against unfavorable shifts in interest rates. Security prices will differ according to their interest rate sensitivity. Merton further shows that securities that are claims on real assets, such as industrial firm stocks, are less sensitive to unexpected changes in nominal interest rates than those which are claims on monetary assets, such as financial intermediary common stocks. For regulated industries, a regulatory lag implies utilities will face a delay between the time their costs rise and the time they are allowed to raise their output price. This will reduce the stock's price. According to Sweeney and Warga (1986). the equivalent of this for depository institutions was eliminated with the Garn-St. Germain Act, when interest rate ceilings were removed. Logue and Sweeney (1981) provide cross-country evidence that increases in inflation bring greater real economic instability. They argue that this is costly and reduces the real return on capital. Samuelson (1945) suggests interest rate sensitivity depends on the mismatch of balance sheet cash flows. In spite of the logical justifications for each of these theories, the nominal contracting hypothesis is most frequently cited as the explanation for the interest sensitivity of financial institutions‘ stock returns. 2.1.2 Empirical Literature The empirical literature on the sensitivity of the stock price of financial institutions to changes in interest rates is well developed. Stone (1974) suggests and implements the use of a two index model to value the common stock of financial institutions. The two index model is as follows: Rjt = Buj' + ijRmt+ 131le1 + an. (51) where, Rjt' the holding period return to the j‘h stock over the month ending at time t, R,“,- the holding period return on an equity index over the month ending at time t. R1,- the holding period return on a debt index over the month ending at time t, and Sjt‘ error term. Chance and Lane (1980) use the two index model to show financial institutions’ equity value unaffected by interest rates. Lynge and Zumwalt (1980) show both long and short interest rates are significant in predicting stock price separately in a two index model and together in a three index model. Giliberto (1985) suggests the significance test used by Lynge and Zumwalt (1980) and others was incorrect. Only Chance and Lane (1980) use the correct test and find interest rates insignificant. Sweeney and Warga (1986), testing the model using a variety of industries, find financial institutions and utilities are most sensitive to changes in interest rates. The following studies offer an explanation for the variability in size and significance of the interest sensitivity of an institution's equity. Kane and Unal (1988) show that interest sensitivity varies over time with different regimes and that the time varying regimes depend on the specific type of institution. Kwan (1991) and Yourougou (1990) support the conclusion of time varying interest sensitivity. Kane and Unal (1990) provide a framework for explaining time variability by showing that the on and off-balance sheet values of institutions vary oppositely with interest rates. Each institution's variability 8 is a weighted average of the two value sensitivities. Song (1994) uses an ARCH model to show significant sensitivity of financial institutions and compares results with previous time varying results. Flannery and James (1984) and Tarhan (1984) show interest rates significant in predicting equity returns for institutions and suggest the interest rate coefficient is a function of portfolio maturity mismatch. Only Flannery and James show the maturity mismatch correlated to interest rate sensitivity. Scott (1986) supports the conclusion of Flannery and James by showing that portfolios of different types of institutions with increasing maturity mismatch have an increasing sensitivity to interest rates. Following is a detailed review of the literature beginning with the interest sensitivity of common stocks, proceeding with the time variability of the sensitivity coefficient, and finishing with the correlation between sensitivity and maturity mismatch. Lynge and Zumwalt (1980) empirically test the interest rate sensitivity of commercial bank common stock returns. They accomplish this by estimating several multi-index models containing debt return indices. Two separate debt return measures are used: a long-term bond index and a short—term (one month) debt index. Each measure is used separately in a two index model, and then the two measures are used simultaneously in a three index model. Because of the construction of these debt indices, they will behave differently for a given change in long or short—term interest rates. Since short returns are dominated by the “interest” return on Treasury bills (at the end of each month only a few days remain until maturity, so the price is not very different from the face value), this index is highly positively correlated with the Treasury bill rates and is always positive. However, the long-term bond index tends to be dominated by price changes (because of 9 the long period remaining to maturity at the end of each month) and often takes on negative values. Thus, long index returns are negatively correlated with long-term interest rates. Therefore, controlling for a market influence, if stock prices decline (rise) when interest rates rise (decline), we would expect a negative coefficient for a short return index and a positive coefficient for a long return index. Lynge and Zumwalt estimate parameters for the 1969-1972 period to provide a basis of comparison with the results of Lloyd and Shick (1977). The estimated equity betas in the single index model for this period are all positive and 54 of 57 are significant at the 5% level. These beta values range from .358 to 1.265 and have an average value of .810. These results differ substantially from the equity betas calculated by Lloyd and Shick (1977) and appear to be more reasonable values for commercial bank equity betas; they also compare favorably with those reported by Value Line. The models were also estimated for the 1969-1975 period. The additional three years provide a highly volatile interest rate environment that may highlight the interest sensitivity of these stock returns. Lynge and Zumwalt find bank stocks exhibit a higher degree of interest rate sensitivity during the 1969-1975 period than during the 1969-1972 period. This is demonstrated by the substantial number of firms having significant coefficients for the debt indices during the longer sample period. These results lend support to the contention of Stone (1974) that a second factor reflecting interest rate movements will improve the performance of the market model, at least for commercial bank security returns. Lynge and Zumwalt report results even more supportive of the two index model than those reported by Lloyd and Shick (1977). 1f) Chance and Lane (1980) use monthly prices for the years 1972-1976 and conclude that extra market interest rate sensitivity is not present in the common stocks of financial institutions. This conclusion holds with respect to short, intermediate, and long-term interest rates, a conclusion contrary to that of previous findings. Giliberto (1985) reviews several studies using a multi—factor model to examine the interest rate sensitivity of a financial intermediary’s common stock. These reviewed studies use an alternate specification of the model in an attempt to estimate each factor’s influence. Giliberto’s note shows that the re—specification results in biased estimators. Hypothesis tests are flawed by failure to acknowledge the bias; this casts doubt upon the reported findings. The published evidence is contradictory. Lynge and Zumwalt (1977) and Lloyd and Schick (1977) report support for an interest-rate factor; Chance and Lane (1980) did not find a significant effect. However, only the latter used an unbiased estimator to examine the interest-rate sensitivity. Although all studies used the same basic model, each study introduced a misspecification that biases some, but not all, of the coefficient estimates. According to Giliberto, there is no reliable evidence that financial intermediaries’ stock prices are sensitive to interest rate changes. Sweeney and Warga (1986) examine whether firms are required to pay investors ex ante for bearing this risk of interest rate changes. The paper uses firll information maximum likelihood (FIML) estimation on groups of twenty-five individual firms, with both cross-equation constraints and within equation nonlinear constraints on the parameters as mandated by the APT model. Equally weighted portfolios were formed based on the two digit SIC. Utilities were broken down into three portfolios and they had 11 the largest interest rate sensitivity. The only other portfolio with a significant interest index was composed of banking, finance. and real estate firms. The following research provides additional evidence indicating financial institution stock price returns are sensitive to an interest rate index. Then these studies extend the literature by including tests of the significance of firm maturity mismatch in predicting interest rate sensitivity. Maturity mismatch is the difference between average asset maturity and average liability maturity and is frequently measured as the short-term assets minus the short-term liabilities standardized by firm value. Tarhan (1984) measures the unexpected movements in interest rates in two ways. In the first and more novel approach, interest rate movement is proxied by money supply announcements. There is evidence pointing out that money announcement surprises cause unanticipated movements in interest rates. The second approach uses unanticipated interest rates based on an ARIMA, autoregressive integrated moving average, model, similar to the interest rate movement used in other studies. Every quarter, the banks in Tarhan’s study are ranked according to the size of their gap (maturity mismatch) positions. On the basis of these rankings, three equally weighted portfolios are formed. The reaction of these portfolios to money supply announcements is investigated first. Three different tests are performed to analyze this issue. In the first test, the portfolio returns on the days following the money supply announcements are compared to see whether or not the response is related to the gap positions. The second test compares the returns of the portfolios using a seemingly unrelated regression technique. In the third test. the relation between the reaction of bank stocks and gap 12 positions of banks is investigated using a non-parametric approach. All of these tests are then repeated for the case where the residuals, obtained from an ARIMA model of daily interest rates, are used as the proxy variable for unanticipated interest rates. All the tests conducted in Tarhan’s study point out that there is no relationship between interest sensitivity and the maturity structure of bank balance sheets. Tarhan suggests several possible explanations for this result. For example, factors such as how a given change in interest rates affects the quality (probability of default) of bank assets may be a more important consideration to investors than the actual interest rate movement. Another possibility is that through the use of interest rate futures and interest rate swaps, banks may be well hedged against interest rate exposure. Finally, it can be argued that the gap measure employed in this study is an inadequate proxy of maturity mismatch. Tarhan shows that bank stock returns are affected by unanticipated changes in interest rates in a statistically significant manner. What is not supported by the results of Tarhan’s paper is the contention that the bank stock reaction to unanticipated interest rates is related to the gap positions of banks. This latter conclusion is in agreement with French, Ruback and Schwert (1983). F lannery and James (1984) test the effect of interest rate changes on the common stock returns of financial institutions. In the absence of a common or accepted motivation, F lannery and James are the first to suggest the nominal contracting hypothesis. Testing the nominal contracting hypothesis requires detailed balance sheet information, such as the information on the maturity structure of assets and liabilities provided by DI’s. They use weekly data from 67 commercial banks from 1976 through 1981. They regress 13 stock price returns onto a market index and an interest rate index to determine if banks have a significant interest rate sensitivity. The nominal contracting theory also suggests that common stock sensitivity to interest rate changes is dependent on the firms average maturity of assets and liabilities. To proxy for average maturity they subtract short-term liabilities from short-term assets and divide the result by equity value (SHORT). Regressing the coefficient of the interest rate index onto SHORT, they test for the significance of contract maturity. Their results confirm that commercial bank stock returns are very sensitive to interest rate changes regardless of the interest rate index employed. When regressing the coefficient of the interest rate index onto the maturity proxy, both the intercept and the coefficient of the maturity proxy are found to be significant. The significant maturity proxy coefficient implies that firm interest sensitivity is responsive to contract maturities. The significant intercept suggests that firms whose short-term assets are equal to their short-term liabilities are still sensitive to interest rate changes. An alternate explanation of the significant intercept is the relationship between maturity mismatch and interest rate sensitivity may be nonlinear. Flannery and James do a final comparison between the interest sensitivity for commercial banks and interest sensitivity for savings and loans. Savings and loans have a greater maturity mismatch and should be more sensitive to changing interest rates. This is confirmed and S&L sensitivity is estimated to be more than twice the sensitivity of commercial banks. The results from both of their tests support the nominal contracting theory. 14 Scott and Peterson (1986) investigate the extent to which portfolios of commercial banks, savings and loan associations. and life insurance companies’ equity values are affected by unexpected changes in interest rates. Their paper analyzes the effect of monthly changes in T-bond yields on the stock market returns of financial institutions for the years 1977 through 1984. The results of the equations show that unexpected T-bond yield changes have a significant effect on financial institutions’ returns. Thus, these results reinforce those of F lannery and James (1984) and others who find the interest rate changes have a significant effect on stock market valuations of financial institutions’ shares. Scott and Peterson show further that portfolios of different types of institutions with increasing maturity mismatch have an increasing sensitivity to interest rates. The remaining papers examine whether the risk sensitivities of financial firms are stationary. These papers provide possible explanations for the variable magnitude and significance of the interest rate index in predicting stock returns. Kane and Unal (1988) find that the riskiness of bank and savings and loan stocks declined in the late 1970’s but rose again later in the 1975-1985 period. They employ the Goldfeld-Quandt search routine as a way of developing policy analysis or event study benchmark models that can incorporate the effects of relevant movements in unspecified omitted variables. A switching—regression method provides a flexible way to identify changes in the systematic and unsystematic risks of asset portfolios. The strength of the technique is that the number of effective regimes, the parameter values in each regime, the switch dates at which one regime supersedes another, and the gradualness of each regime switch can all be estimated simultaneously. 15 Their results show savings and loans (S&Ls) experience significant shifts in market and unsystematic risk in early 1976, developing a significantly negative market beta and a greatly enhanced interest sensitivity. However, S&Ls show an abrupt second shift in 4/77, one combining increases in unsystematic risk and market beta with reduced interest sensitivity. In contrast to the 1979 shift for bank groups, the retum-generating process for their small sample of S&Ls shows no switch during the 1979-1980 era. S&Ls undergo their third shift in late 1981, with market and unsystematic risk doubling and interest sensitivity declining to insignificance. In the late-1983 shift. market risk declines and interest sensitivity rises (both significantly), while the fall in unsystematic risk fails to achieve statistical significance. With only eight extremely large S&Ls in their sample (and most of these headquartered in California), it is doubtfirl that the regression shifts they observe are representative of the S&L industry as a whole. Interest rate sensitivity varies over a far wider range of values for S&Ls than for banks. Moreover, three out of the four shifts in S&L interest rate sensitivity prove significant. During the 9/81-9/83 regime, although interest rates were highly volatile, S&L stocks show near-zero interest sensitivity. This may trace not only to a higher incidence of adjustable-rate mortgages and to mortgage prepayments in the last part of this era but also to the extent of hidden economic insolvency at sample S&Ls. Deep insolvencies could have forced the Federal Savings and Loan Insurance Corporation (FSLIC) to absorb the bulk of interest induced profits and losses on short firnded positions in long assets during this particularly troubled era. FSLIC’s own growing economic insolvency and decreased staffing reduced the threat of formal failure and allowed firms to continue in insolvency. l6 Kane and Unal conclude that their sample of deposit institution stocks became riskier investments in the wake of the many regulatory relaxations made in the 1980’s. They firrther suggest the information events of 1979 and 1982 substantially altered return- generating processes for deposit institutions' stock. Bank equity returns prove interest sensitive primarily during the 1979-1982 era, but S&L equity returns prove interest sensitive during the bulk of the observation period. Yourougou (1990) examines interest rate risk and the pricing of depository financial intermediary common stock. Like Sweeney and Warga (1986), Yourougou (1990) uses firms from all industries and applies an APT model to determine interest sensitivity. Yourougou examines the impact of interest rate risk on security prices in periods of relative interest rate stability (pre-October 1979) and during periods of great volatility of interest rates (post-October 1979). A sample of 219 weekly returns was obtained for each of 83 banks, 32 S&L associations, and 100 industrial firms. Yourougou’s results from the post-October 1979 period of more volatile interest rates show interest rate risk had a significant impact on common stock prices for financial intermediaries, but the industrial firms have been shown to have little or no interest rate sensitivity. In the pre-October 1979 period, when interest rates were stable and the interest-rate sensitivities of common stock were relatively low for most firms, the interest rate risk appeared to have had no effect on stock prices. Their evidence further indicates that the inability to detect such a pricing effect for the sample of industrial firms and financial institutions in the pre-October 1979 period was due to the low interest-rate 17 sensitivity, rather than the inadequate volatility of interest rates or the markets' failure to price interest-rate risk. Kwan (1991) reexamines interest rate sensitivity of commercial bank stock returns using a random coefficient model. Flannery and James (1984) and Kane and Unal (1988) suggest that the interest rate beta in the two index model appears to be nonstationary. Bank stock interest rate sensitivity is first modeled as a function of a bank's maturity profile. This relationship then substitutes into the traditional two index model. This method is unique in that it allows a time varying beta. Kwan’s findings suggest that commercial bank stock returns are related to unanticipated interest rate changes. In support of the hypothesis that an increase in the duration of net nominal assets corresponds to an increase in the interest sensitivity of bank stock returns, Kwan finds the coefficient of a maturity proxy positive and significant. The effect of unanticipated changes in short-term interest rates on bank stock returns, after accounting for the maturity proxy, is also found to be significant for a number of banks. When the long-term rate is used in constructing the interest rate index, the maturity proxy ceases to have explanatory power for interest rate sensitivity. However, bank stock returns are still sensitive to changes in the long-term interest rate, which is evidenced by a large number of significantly positive intercept terms. The intercept now captures the interest rate sensitivity caused by both the duration mismatch and the factors not related to balance sheet composition. A possible explanation for the above phenomenon is the maturity proxy is a noisy measure in determining duration mismatch between long term assets and 18 liabilities. Second. changes in the term structure of interest rates are not always a parallel shift. Kane and Unal (1990) acknowledge the accounting representation of a firm's net worth diverges from its economic value, by an amount they call hidden capital. Two sources of hidden capital exist: accountants' misvaluations of portfolio positions that accounting principles designate as on balance sheet items and the systematic neglect of off balance sheet sources of value. Their paper develops a model for estimating both types of hidden capital. The model makes direct use of accounting information on the bookable positions of a firm and separates bookable from unbookable sources of value. They use regression analysis to partition the total market value of a firm's stock into two components: recorded capital reserves and unrecorded net worth. Hidden capital is then allocated between values that are unbooked but bookable through asset turnover or write- downs on a historical cost balance sheet under Generally Accepted Accounting Principles (GAAP) and values which GAAP currently designates as an unbookable off balance sheet item. Kane and Unal estimate the net unbooked value of on balance sheet positions by estimating an intermediate variation ratio. This variable expresses the ratio of the market value to the book value of the collected components of a firm's bookable equity. Applying the valuation ratio to the value of accounting or book net worth assigns a market value to bookable assets and liabilities. Subtracting this estimate from market capitalization assigns a value to off balance sheet items. This statistically appraised value of unbookable equity expresses the net value of unbookable assets and liabilities. They call this regression l9 equation, the Statistical Market Value Accounting Model. SMVAM. Bank stock shares have shown weak and sporadic sensitivities to interest rates rather than strong and consistent ones, even though the value of a bank's individual asset and liability positions is inherently interest sensitive. Because GAAP gives bank managers options to realize unbooked gains and losses as "nonrecurring items" and authorities penalize low book value, individual banks have an incentive to sell assets with unrealized gains when the book value of their capital appears inadequate. In contrast, no reason exists to expect the valuation index to be a function of size, and estimated rank orderings of the valuation index against bank size class do vary over time. The model's coefficients describe the de facto deceptiveness of GAAP. Unless both unbookable net worth equals zero and the valuation index equals one, the accounting or book value of a bank's capital represents a biased estimate of the market value of stockholder equity. As an example, off balance sheet liabilities at the five largest US. banking firms totaled $1.16 trillion at year end 1986. This value is more than twice the $546 billion book value of these banks' assets (Forde 1987). In principle, each bank has a different unbooked equity and valuation index at each date. To estimate the model, it is necessary to restrict parameter variation either across time or across banks. To consider the full effect of the mutability of the 1975-1985 environment, they restrict SMVAM parameters across banks rather than across time. In applying the model cross sectionally, they restrict the valuation ratio applicable to each bank at a specific time to have the same valuation index across each bank class. To lessen the information loss from this restriction, it is necessary to focus on relatively homogenous 20 subsamples of banks. The estimates for unbookable capital and the valuation ratio are then separately regressed in the two index model. This shows the interest sensitivity of both booked and unbooked assets and suggests that firm sensitivity is a weighted average of the sensitivities of the two parts. Further, it provides an explanation for time varying interest sensitivity. When this model is applied to a sample of bank stocks, the interest and market sensitivities of bookable and unbookable values are often different in sign. In particular, increases in the value and sensitivity of hidden capital at the nation's 25 largest banks during the interest-rate spike of 1978-1982 cushioned a sharp decline in the valuation ratio for their net bookable assets. This is consistent with the hypothesis that during this period, increases in unbookable value resulted from Federal Deposit Insurance Corporation (FDIC) guarantees and enhancements in franchise values fed by technological change and relaxations of regulatory restrictions. Although they apply the model only to bank stock shares, in principle it could be used for any corporation. By permitting regime changes in the valuation models that reset market values each quarter, their methods provide new insight into changes in: (1) the market and interest sensitivity of a corporation's stock and (2) the impact of off balance sheet positions on a firm's stock price. Song (1994) applies autoregressive conditional heteroskedasticity (ARCH) modeling to the study of deposit institution stock returns. The questions Song investigates are whether and to what extent the market and interest rate risks of depository institutions have been changed in the 1980’s. The two-factor ARCH model allows Song to identify the dynamic pattern of the market and interest rate risks. Finally, comparing 21 the two-factor ARCH model and the switching regression model with the constant-beta model suggests that the two factor ARCH model provides a better measure for the average trend in the betas for the deposit institution stock returns. Song shows the interest rate risk for money center banks and savings and loans is positive; a positive value of interest rate risk means that the firm's market value declines when interest rates rise. One surprising finding is that there was no big difference in the interest rate risks between banks and S&Ls. The S&Ls in this sample are all large institutions and may be as efficient as large banks in hedging interest rate volatility. Kane & Unal (1988) suggest bank stock returns are interest sensitive primarily during the 1979-1982 era, but S&L stock returns are interest sensitive during the entire period of 1975-1985. Song finds interest rate risk for all three types of depository institutions showing only small increases at the end of 1979. This is consistent with the evidence found by Aharony, Saunders, and Swary (1986) that there was no significant change in market or interest rate risk (IRR) for banks and S&Ls around October 1979. At the end of 1982, all three types of depository institutions show a small increase in IRR, suggesting that deregulation allowed the institutions to take more unbalanced positions. Several events during 1982 are of significance. The repel of Regulation Q by the Depository Institutions Deregulation Committee, established under the DIDMCA, occurred in December 1982. In August, Mexico declared a moratorium on its foreign debt. In November, Congress passed the Garn St. Germain Act which among other purposes, authorized MMDA, Money Market Deposit Accounts, and SuperNow accounts. In 1982, the number of failed and problem institutions increased sharply. Kane 22 and Unal suggest the combined effect of these events seemed to increase the risk exposure of depository institutions. Song’s results suggest that both bank and S&L stock returns have a slightly higher and more volatile interest rate beta since 1982 then during the 1979-1982 period when the Fed altered its monetary policy. Significant changes in the risk exposure of depository institutions suggest the importance of modeling the time varying betas in the market model for deposit institution stock returns especially in the 1980’s, because of dramatic changes in banking regulations and the economic environment. In conclusion, this study casts doubt on the results obtained from conventional event studies that assume constant betas in the market model. 2.2 Literature on the Failure of Financial Institutions Another way to test the relationship between interest rate risk and the common stock price of financial institutions is through a failure prediction model. Tests of this relationship in failure prediction models regress thrift failure against interest rate risk as well as other relevant variables. Prediction models of thrift failure have three applications. First, the model can be used as an early warning system, a system that identifies operating thrifts with similar financial characteristics as failed thrifts. These operating thrifts are then considered for regulatory examination. A second use is in relation to federal deposit insurance. The variables and coefficient values relevant to failure, could be used in adjusting premiums for deposit insurance. Similarly, a third application relates to risk based capital standards. 23 Risk based capital regulation increases required capital levels for firms with more risk. F irm characteristics positively associated with the probability of failure could be used to adjust required capital levels. I first review tests that predict failure with either call report or examination data. The review continues with models, predominantly Sinkey (1975, 1978), that attempt to predict the Federal Deposit Insurance Corporation's problem bank list. Finally, studies that employ market prices along with call reports and /or examination data are reviewed. The literature is summarized in Table 1. 2.2.1 Failure Prediction with Accounting Data Martin (1977) uses bank call report data to determine bank failure. Martin compares discriminant analysis with a logit model and suggests their relative merits depend on the intended use of the results. If classification is the goal, the models‘ accuracy is similar. If the determination of a risk premium for deposit insurance is the goal, then the performance of the logit model is slightly superior. Collins (1982) predicts failure using credit union data. Collins compares the assumptions and predictive abilities of several statistical models. The assumptions underlying the linear model and multiple discriminant analysis are not consistent with the bankruptcy forecasting problem, but their performance is surprisingly good. The logit model, which a priori appears to match the requirements of the bankruptcy problem well, performs only slightly better than the previous models. The logit model provides a modest increase in the overall classification accuracy, and substantially reduces type I error. Since the purpose of most models is identifying failure, this is an important result. 24 Bovenzi, Marino, and McFadden (1983) review the present body of literature, then add a comparison of bank call report data and examination data. Examination data is found to improve the accuracy of classification, but the improvement diminishes as the time to failure increases. Examination data alone is found, at best, to be as good as call report ratios. A review of sampling techniques shows that for unweighted nonrepresentative samples the constant term of the model is biased in a direction that overestimates failure probabilities. The t-statistics of the estimates are also problematic and may distort the significance of the relation of the financial ratios to failure. Bovenzi et a1. (1983) use the difference between market rate assets and market rate liabilities divided by equity capital as an interest rate sensitivity variable. This variable improved the classification of the model, but was not as informative as the efficiency or credit risk variables. This may indicate that interest rate risk is not the most serious problem facing banks or it may indicate this interest rate sensitivity variable is not an accurate indicator of a bank’s interest rate risk. Korobow and Stuhr (1983) use bank call report data to predict failure but distinguish between different peer groups. In the group of banks with over $300 million in assets, 12 of the 13 marginal or weak banks had at least one foreign office. The inferior performance of the multinational banks tends to increase with asset size. Richardson and Davidson (1983) predict bankruptcy in firms listed on the American Stock Exchange. The classification ability of linear discriminant analysis is found to be sensitive to three non-linear transformations of normality: the log normal, the logit normal. and the inverse hyperbolic sine normal. The classification ability is also 25 affected by deviations from normality in the accounting data. These deviations cause small, but statistically significant differences in classification. Barth, Brumbaugh, Sauerhaft, and Wang (1985) use a logit model to predict the failure of S&Ls. They use data from the Thrift Financial Report beginning in 1982 and continuing through mid-1984 and they test the significance of twelve independent variables from categories similar to other tests. They find most of their variables significant; the most relevant to this study is the significance of their proxy for interest rate risk. Barth et a1. (1985) use three different variables for measuring interest rate risk: interest sensitive funds, jumbo certificates of deposit (CDs), and cost of advances.2 In analysis of variance tests between failed and non-failed thrifts, all three variables were significant at the five percent level, at all three time periods: failure, six months prior to failure, and twelve months prior to failure. In the multinomial regressions, the most explanatory model included interest rate risk only six months prior to failure. The most predictive of the three interest rate variables was interest sensitive funds. Rose and Kolari (1985) use univariate and multiple discriminant analysis to predict failure with bank call report data. Tests using linear and quadratic discriminant analysis show the predictive accuracy of linear models superior to that of quadratic models. Ratio analysis suggests the failure process for commercial banks is generally marked by liquidity 2 Interest sensitive funds = Interest-sensitive funds (savings accounts earning interest above regulated rate plus. Federal Home Loan Banks advances due in one year or less. and other borrowed money due in one year or less)/Total funds (non-interest-eaming demand and NOW accounts plus savings accounts earning interest below regular rate. and FHLB advances and other borrowed money) Jumbo CD/Total liabilities Cost of advances/Total liabilities 26 problems on the asset side of the balance sheet and increasing risk exposure in both loans and interest sensitive liabilities. Summarizing their results, troubled banks experience a squeeze on their profit margins due to the interaction of expense control difficulties, including loan losses, and rising interest costs associated with the acquisition of deposits. Pantalone and Platt (1987a) use Thrift Financial Reports from Boston area thrifts to predict failure with linear discriminant analysis. Testing both linear and quadratic models, they conclude the linear model had better overall classification results. They find broker originated savings are positively correlated to thrift failure. In another study, Pantalone and Platt (1987b) predict failure using a logit model with bank call reports. Healthy banks are found to have higher net income and equity ratios and relatively lower amounts of loans, particularly commercial and industrial loans. State economic variables are found insignificant in predicting a bank's state of health or failure in every model. Cole (1993) compares the determinants of failure prediction to the determinants of closure prediction. There is almost complete overlap in significant variables between the two models, but the weights of these variables are much different for the two tests. The variables are categorized into operating and agency risk. The agency risk variables suggest the existence of owner, manager, and regulator agency conflicts. Further, the evidence from these tests indicate regulator forbearance was a greater problem in the later 1980’s than in the earlier 1980’s. 27 Thomson (1992) uses a two-step logit to predict closure. Variables correlated to failure include loans to insiders, a dummy for bank holding companies, and a dummy for unit branching states. 2.2.2 Loss Prediction with Accounting Data Barth, Bartholomew, and Bradley (1990) provide a detailed examination of resolution costs of thrift institutions. They show that the model that most accurately predicts costs varies from the early and mid 1980’s to the late 1980’s. During the 1985- 1988 period, their model shows discount rate, off-book items, non-performing assets, and core deposits are highly significant with the expected sign. James (1991) predicts the losses realized in bank failures during the 1985-1988 period. He finds losses average 30% of bank assets. In a weighted least squares regression, seven different asset variables, book value of equity, and core deposits are significant in predicting realized bank losses. 2.2.3 Problem Status Prediction with Accounting Data Other tests have attempted to use problem status as the dependent variable. Sinkey (1975) predicts problem status with a quadratic discriminant analysis model using bank call report data. Using a chi-square test, the mean vector of the problem banks was found to be significantly different from that of healthy banks, but the two distributions substantially overlap. Sinkey (1978) uses bank call report and examination data with discriminant analysis to predict problem status. The net capital ratio, a capital ratio adjusted for substandard loans, is the most significant variable and the most important classification discriminator. Examination data failed to add to the classification ability of 28 the net capital ratio. Classifying large commercial banks by their net capital ratio produced a zero type I error. Altman (1977) tests the problem status prediction ability of quadratic discriminant analysis on savings and loan call report data. The dependent variable has three possible states: no problem, temporary problem, and serious problem. Two measurements of management efficiency, net operating income / gross operating income and its associated two-period trend, are the most important indicators of problem status. West (1985) employs a combination of factor analysis and logit to predict problem status in banks. Independent variables or factors are from call report and examination data. They find using factor scores as inputs in a multivariate logit estimation holds promise as an early warning system. The results of this test are robust to different geographic areas. 2.2.4 Failure Prediction with Market Data According to the eflicient market hypothesis, stock prices should incorporate all public information, including public accounting information. If this is the case, any abnormal negative return in the bank or thrift stock price may indicate financial trouble. Pettway and Sinkey (1980) use bank call report and market data to predict failure. Call report data show fewer investments and lower efficiency characterizes failed banks. The accounting information in a multiple discriminant analysis model detects all the failed banks in the sample. a type I error of 0. In all but one case the market filter detected the failed bank before the examination that led to problem status. The accounting filter generally led the market filter in predicting failure, but had a higher type 11 error. 29 Pettway (1980) predicts failure using market information on banks with a modified Fama, Fisher, Jensen, and Roll (1969) approach. Negative stock returns are found as early as 38 weeks before the exam that led to problem list status and as early as two years before failure. Schick and Sherman (1980) use the Fama, Fisher, Jensen, and Roll (1969) method with examination and market data to detect decreases in stock price. Evidence clearly indicates that changes in condition are reflected in stock price. On average, detectable market decline predates the examination that led to problem status by an average of nine months. To summarize, the benchmark statistical model for failure prediction is a logit. The independent variables most frequently found correlated to failure and problem status are related to the following categories: capital level, asset quality, management efficiency, earnings ability, and liquidity (CAMEL). The most notable exception to this is an interest rate risk variable that was found to be significant in some studies. but only at marginal levels, possibly because these measures were inaccurate proxies. CHAPTER 3 - THE MARKET VALUE MODEL Interest rate volatility has been a threat to savings and loan capital since the industry was created. Measuring the portfolio interest rate risk of a savings and loan (S&L) is problematic due to the diversity of securities in the portfolio and the data necessary to compute security specific interest rate risk (ERR). Ignoring these diversity and data problems, numerous tests of interest rate sensitivity use a simple gap measure. Information on securities repricing in more and less than a year is available and creates a one year gap bucket. A gap bucket is the difference in dollar amount between the assets repricing within one year and liabilities repricing within one year. Continuing concern by the regulatory bodies of financial institutions about interest rate risk has led both the Office of Thrift Supervision (OTS) and the Federal Reserve Bank (FRB) to develop a more accurate proxy for this risk. The OTS Market Value Model (MVM) measures IRR as the percentage change in the market value of institution portfolio equity (MVPE) as a result of a hypothetical interest rate shock. To calculate the MVM, first compute the present value of firture cash flows for each asset and liability. Second, determine the MVPE by subtracting the present value of all the liabilities from the present value of all the assets. Third, adjust the discount rate with an interest rate shock. Then, use the new discount rate to recalculate the MVPE. The percentage change in the MVPE is the IRR proxy. To compute present values, the OTS model categorizes the assets and liabilities on the balance sheet into 30 categories of securities; each category is characterized by similar responses to changes in interest rates. The book value for each category is then broken 31 down into different time horizons or “buckets” by the maturity or repricing date of the security. The maturity for all securities is assumed to be the midpoint of each bucket unless otherwise stated. For each bucket, in each category, future cash flows are estimated and discounted to compute a present value. Cash flows for each bucket may include interest, principal, and prepayment. The discount rate is determined by internal rate of return or option adjusted spread, depending on the sensitivity of the cash flows to changes in interest rates. The exact present value methodology for each security is detailed below. After computing the MVPE based on the existing term structure, the MVPE is recomputed for a shocked term structure. The shocked term structures used for regulatory purposes are parallel shocks of 200 basis points (bp) both upward and downward. The percentage changes in the MVPE resulting from the term structure shocks provide measures of [RR for both upward and downward movements in the term structure. The data available for this dissertation is significantly different from the data available for previous failure prediction studies. The data is from the quarterly Thrift Financial Report (TFR) beginning in 198513 and ending in 1989212.3 The TFR includes standard financial forms and section H, a maturity and yield table. Section H reports seven balance sheet assets and six balance sheet liabilities and subdivides each category into seven maturity “buckets”. The buckets are 0-6 months, 6-12 months, 1-3 years, 3-5 years, 5-10 years, 10-20 years. and 20-30 years. For each bucket the corresponding yield or cost is also reported. 3 The MVM was designed in conjunction with the data requirements of the 1990 TFR. The data used in this study is quite similar, but not identical. Hence. the MVM has been modified slightly to accommodate the data. 32 The next section explains the two methods for determining a discount rate. Then, the model for the term structure of interest rates is explained. The final section defines each security category and provides a detailed explanation of the present value methodologies. 3.1 Discounting Methods The discount rate is the sum of the Treasury curve and a risk premium. The OTS MVM employs two methods for determining a discount rate depending on the interest sensitivity of the cash flows. If cash flows do not fluctuate with interest rates, the discount rate is determined by internal rate of return. If cash flows are interest sensitive, the OTS uses an option adjusted spread (OAS). A problem with the internal rate of return is that it does not account for uncertain cash flow. Thrift assets include many mortgage related securities with embedded options that affect their cash flows. For mortgage securities, these options may take the form of caps, floors. or prepayments. The value of these options is dependent on interest rates. The OAS methodology computes the risk premium for a representative security, with a known market price, for numerous simulated interest rate paths. The MVM uses the following present value (PV) equation: PV 1' 1&2 CF” (31) = 1m— ,, . N (1+5; +0AS )‘ 1' N <1 4 ) 1:1 i=1 where, N is the number of independent interest rate paths (200 paths are used for the MVM), n is the number of periods, ri is the annual spot rate at time i, OAS is the representative securities risk premium, and CFi is the interest sensitive cash flow for 33 period 1. The present value of the security is equal to the average present value for all 200 unique interest rate paths. By considering numerous interest rate paths, an accurate distribution of option values is included in the calculation. A constant OAS is added to each spot rate, causing a parallel shift, to reflect the additional risk of the representative security. The constant spread is determined by an iterative process, where an initial OAS is guessed and then repeatedly adjusted until the average present value is acceptably close to the market value.4 Using this OAS and the previous equation, the present value of any security similar to the representative security is calculated. The internal rate of return model is used for all categories of assets and liabilities whose cash flows are not interest rate sensitive. Again, a risk premium is computed for the representative security, but for only one interest rate path. The standard present value formula is as follows: PV=Z—€L7, (3.2) ,=1(1+r, +rp) where, n is the number of periods (years, months, etc), r, is the spot rate for period i (expressed as a per period rate), rp is the risk premium for the representative security, and C F1 is the interest sensitive cash flow for period i. The risk spread is again solved for using an iterative process, this time over only the current term structure. An initial value is guessed and this value is adjusted until the present value is sufficiently close to the market or par value of a representative security. The spread is used to value securities of 4 The OAS is estimated in two passes to reduce computer time. First. only twelve paths are used through five iterations. Second. the result from the twelve paths is used with the 200 paths and iterated to within a discount rate of .5%. 34 risk similar to the representative security in the internal rate of return present value equation. 3.2 Interest Rate Model The interest rate model is governed by two arbitrage conditions. First, the risk neutral short rate path distribution must reflect the current term structure. Second, as the term structure moves through time a dynamic arbitrage condition ensures that two equally risky portfolios have the same expected return. These two conditions are met using a parsimonious model for the current term structure developed by Nelson and Spiegel (1987) and a model for the path of short interest rates developed by Brennan and Schwartz (1980). The Nelson and Spiegel model is used for the single discount path with the internal rate of return and used to standardize the 200 paths of the OAS model. The Brennan and Schwartz model forecasts the 200 unique paths used in the OAS, such that each path is consistent with historical parameters of the short rate movement. Nelson and Spiegel (1987) introduce a parsimonious model that is diverse enough to represent the variety of shapes generally associated with yield curves: monotonic, humped, and S shaped. They let the instantaneous forward rate at maturity m be the solution to a second-order differential equation with real and unequal roots. The yield to maturity on a bill, denoted R(m), is the average of the forward rates, which integrates to, R(m)=,30 +(fll +fl:)[1—€XP(-m/ T)]/(m/ firflzeXM—m/ T). (3-3) 35 where, t is a time constant selected for best fit, and B0, [1,, and [32 are determined by initial conditions. The limiting value of R(m) as m gets large is 130 and as m gets small is ([30+Bl), which are necessarily the same for the forward rate function since R(m) is just an averaging of the forward rate. The time constant, t, represents how quickly the regressors decay. Small values of t correspond to rapid decay and provide a better fitting yield curve at short maturities, but are unable to fit excessive curvature at longer maturities. Correspondingly, large values of t correspond to slow decay in the regressors and provide a better fit over longer maturity ranges, but they will be unable to follow extreme curvature at short maturities. Using bills for their study, Nelson and Spiegel (1987) avoided some of the difficulties associated with coupon bonds, such as differential rates of taxation for coupon income and capital gains. Although the term structure is fitted using short term data, it is necessary to test the predictive power of the model on long term bonds. The actual and predicted bond prices are highly correlated at .963, but it is also clear that the predictions overestimate the actual prices. This suggests the fitted curves may flatten out too quickly. Similarily, estimating inverted yield curves, the models overshoot long-term discount rates and therefore underestimate the price of the bonds. A cubic polynomial fits the data on bill yields slightly better and it has the same numver of parameters as Nelson and Spiegel’s model. However, at longer maturities a cubic polynomial will shoot toward either plus infinity or minus infinity. The correlation between actual and predicted bond price is - .020, so the polynomial model has no predictive value. although it fits the sample data 36 very well. In this test the Nelson and Spiegel (1987) model is clearly superior to a cubic polynomial in predicting bond price. Using the Nelson & Spiegel model and Treasury prices for maturities of 3 months, 6 months, 1 year, 2 years, 3 years, 5 years, 7 years, 10 years, and 30 years from CitiBase, the model parameters are estimated at December 31. of each year, from 1985 to 1989. Model error is minimized over different values oft to obtain the best fitting coupon Treasury yield curve.5 For discounting, the zero coupon Treasury curve is used, so the above coupon yields are used to generate the zero coupon spot rates in the following equation: \ _/ 3|— 1 C —l, (3.4) where, rc is the yield to maturity of a coupon bond selling at par with maturity n and Ti is the zero coupon spot rate for i periods. The zero coupon spot rates are calculated sequentially starting with the first period. six months. Using the calculated zero coupon spot rates, the Nelson & Spiegel model is used to determine the optimal zero coupon yield curve over several values of t.6 In the present value equations, the zero coupon yield curve plus a risk spread is used as the discount rate. F orecasted term structures must be consistent with the probable movement of the short term market rate. Brennan and Schwartz (1980) developed a model to generate the 5 The following values of 1 led to the best fitting Treasury curves: 1:24 for 1985. 1:24 for 1986. F6 for 1987. t=6 for 1988. and t=6 for 1989. 6 The following values of I led to the best fitting zero coupon Treasury curves: FM for 1985. r=20 for 1986. r=4 for 1987. r=4 for 1988. and F1 for 1989. 37 path of short term interest rates with a mean reverting drift term and heteroskedastic variance. Chan et al. (1992) found this model to have the lowest Chi-square among the models they tested. A discrete time version of their model is as follows: rt+l'rt = Ot‘"l3rt+f3t+l~ (3.5) E[et+1] = 0, and E[52t+1] = 02’12’ where, rt is the one month Treasury rate at time t. Parameters are estimated using a one month Treasury series from January 1956, to December 1985. The Treasury series is from the Center for Research on Securities Prices. The ordinary least squares estimates/(t-statistics) are or = .007201/(7876), B = -.09932315/(4.858), and o = 1513277 Although the random short term interest rate paths generated by the Brennan and Schwartz (1980) model all start from current short term Treasury rates, there is no guarantee that the 200 n-month sequences of short rates used for discounting will, on average, exactly return the n-month zero coupon Treasury price estimated previously. The ability of the short rate distribution to explain current Treasury prices is a necessary requirement for model consistency. The generated distribution is adjusted to exactly price the currently observed zero coupon Treasuries. Beginning with month 1 and continuing to month 360, the number of basis points that causes the distributions average zero coupon Treasury value of the 200 paths to equal the value estimated by the Nelson and Spiegel 7 Weighted least squares estimates are similar. 38 (1987) model is added to (or subtracted from) all short rates at that month. The resulting short rate distribution is thus consistent with the current zero coupon Treasury curve. 3.3 Account-Specific Present Value Methodologies The OTS MVM divides the securities of savings and loan balance sheets into different categories, each with a slightly different present value methodology. Descriptions of discounting techniques used are in section 3.1. The following is a description of the methodology used for each thrift account. 3.3.1 Asset Methodologies To compute the present value of fixed rate mortgage (FRM) loans and securities for 1-4 family dwellings, the reported book value is multiplied by the appropriate multiplier from a price sensitivity table. Each multiplier converts one dollar of book value into a present value, which is a function of the maturity and coupon of the mortgage and the appropriate term structure scenario. Table 3 shows the FRM price sensitivity table for December 1985. The multiplier for each FRM is the average present value of future cash flows over the 200 interest rate paths. The cash flows are the sum of the standard payment calculation for amortized debt and the prepayment equation as stated below. The amortized debt payment given the present value is R = m (3.6) 39 where, R is the regular periodic annuity payment, PVAmn is the present value of an annuity of mn periods, r is the annual nominal interest rate, and m is the number of compounding periods. For each month, the outstanding loan balance must be computed based on the previous month’s prepayment. Each multiplier is computed using an OAS risk premium. The representative securities for FRMs are Federal National Mortgage Association (FNMA) 30 year, fixed rate, mortgage backed securities, with coupons ranging from 7% to 14%. The mortgage price data was obtained from the Bloomberg System. Prepayment models vary significantly in complexity. The historical data on which they are based is never sufficient to accurately predict the future. In spite of this, the OTS prepayment model for fixed rate mortgages adequately describes prepayment using only three factors: seasoning, seasonality, and refinancing. The equations for this model are as follows:8 CPRt = Rt * Zt "‘ St. (3.7) Rt = 0.1828 - .0892 Arctg (4.776 (-C/mt_3 + 1.083)), Zt = t/30 for t<30 months; 1 thereafter, and St = l + .20 sin [3.1415/2((month + t-3)/3-1)], where CPRt is the constant prepayment rate at month t, Rt is the prepayment resulting from the refinancing incentive, Zt is the seasoning variable, St is the monthly seasonality factor, C is the mortgage coupon, and mt is the current mortgage rate. The mortgage rate is the short rate at month t+60. 8 The prepayment model for the OTS MVM and the OTS Net Portfolio Value Model (NPVM) are similar except the MVM requires infonnaiton on previous prepayment rates. The equations used in this study are from the NPVM. 40 The present value for adjustable rate mortgages (ARM) is also computed as the product of the reported book value and a multiplier from a price sensitivity table. The price sensitivity table is broken down by index, maturity, current yield, and term structure. The adjustable coupon is set equal to a margin plus a market rate index. Indexes used are 1, 3, and 5 year Constant Maturity Treasuries (CMT) and correspond respectively to buckets with repricing at months 6 and 12, 24, and 48. Current yields range from 7% to 14%. The reset frequency is the length of time between resetting the variable rate coupon. The reset frequency matches the index with the first reset at half of the index, period caps and floors are 2%, and lifetime caps and floors are set at 6%. The margin is chosen as follows for each index: 1 year index, 275 bp; 3 year index, 240 bp; and 5 year index, 200 bp. The cash flows are the sum of the amortized payment and the prepayment using a variable index to compute the payment. The MVM uses a prepayment model identical to the fixed rate model except for one factor. The refinancing variable is specified: Rt=.4072-.2267Arctg(8.45((-Ct/mt-3)+1253)). (3.8) With these assumptions, a representative security is used to compute a representative OAS. The price quotations for the representative security, a one year CMT, are obtained from the Wall Street Journal. For ARM maturities greater than five years the balances are treated as FRM’s with a balloon payment. Principal and interest are computed as though the mortgage had a maturity of thirty years. The prepayment equation and the OAS are identical to those used for valuing a FRM. The balloon payments are paid at the actual maturity: 7.5, 15, or 25 years. 41 “Other mortgages” includes construction loans, permanent mortgages on multi- family residential property and nonresidential property, and land loans. The present value is found with discounted cash flows because cash flows are assumed interest insensitive. The cash flows are the sum of the amortized payment and a prepayment. The prepayment is identical to the prepayment of a fixed rate mortgage. The spread is unique for each mortgage type but constant over the forecast period. A discount spread is computed for each type of “other mortgage” and the weighted average spread is used to discount the aggregated cash flows. The weights are determined from firm specific balance sheet accounts. A maturity of ten years is assumed for the representative securities. The representative securities are the Barrons / Levy Commercial Mortgage index for permanent mortgages and prime plus 300 bp for land loans. Other mortgage securities repricing within one year are treated as adjustable rate securities. To compute the standardized discount spread. assume the maturity is 10 years, the margin is 300 bp, and the reset is 12 months except the first reset at month 7. The index is computed by adding a basis to the short rate. The basis is assumed to be half the difference between the short rate (1 month) and the mortgage rate (5 years). The equation for prepayments is the same as that for adjustable rate mortgages. Upon determining the spread, present values are computed with all of the above assumptions except maturity is taken as the weighted average of the 5 longer buckets. The category “second mortgages” includes all closed-end junior liens, revolving open-end liens, and junior lien loans on 1-4 family dwelling units. The present value is computed by discounting monthly cash flows by the Treasury curve plus a spread. Cash 42 flows are equal to the monthly amortization payment plus the prepayment. The prepayment is the fixed rate mortgage prepayment plus 2% annually. The spread is computed by valuing a security whose interest rate is equal to the current first mortgage rate plus a 100 bp premium, where price equals par, and whose maturity equals thirty years. 1 The shortest two buckets are considered to be adjustable rate second mortgages. The spread is based on the following assumptions: prepayment is identical to adjustable rate mortgages plus 2% annually, the reset equals the index except the first reset is at month 7, the margin is 300 bps, and the index equals the short Treasury rate plus a basis that is half the difference between the short rate and the mortgage rate. After computing the spread, all assumptions stay intact except maturity which is now the weighted average of the five long buckets. “Consumer loans” includes the following: loans on deposits, home improvement loans, education loans, auto loans, mobile home loans, other commercial loans including leases, revolving loans secured by 1-4 family dwelling units, and unsecured loans including credit cards. The cash flows used in the static discount cash flow model are constant monthly interest payments computed as the product of the reported rate and the account value. The category discount spread is the weighted average of individual security spreads. Security spreads are determined by setting the account value to par and using current market yields to establish the spread. These current yields are as follows: for loans on deposits, the 6 month certificate of deposit (CD) yield plus 100 bp; for home improvement loans, the first mortgage rate plus 200 bp; for education loans, the 3 month 43 T-bill yield plus 300 bp; and for the remaining consumer loans, the yields on Federal Reserve Release G. 19. Maturities for these accounts are as follows: loans on deposit, 2 years; home improvement loans. 10 years; educational loans, 3 years; auto loans, 4 years; other personal loans, 2 years; and mobile homes, 10 years. The shortest consumer loan bucket is treated as though composed of adjustable rate loans. The short bucket weights of each type of consumer loan are assumed identical to the weights of fixed rate consumer loans and the maturity of the adjustable rate loans is considered equal to the weighted average maturity of the fixed rate loans. The spread is the same for fixed and adjustable rate consumer loans. “Commercial loans” includes secured and unsecured commercial loans and financing leases. Economic value is computed using the static discounted cash flow model. Maturity is the midpoint of each bucket. Cash flows are constant interest payments equal to the face amount multiplied by the reported rate of return. To calculate the spread, use the Federal Reserve Release E2, “Survey of Terms of Bank Lending”, rate to calculate cash flows. Then, set the face amount equal to one. Assume maturity to be 48 months for fixed and adjustable rate commercial loans. The shortest bucket is assumed composed of adjustable rate commercial loans. The margin and index used to compute the spread are the same as for adjustable rate consumer loans. Again, the maturity to compute the economic value is the weighted average of the fixed rate loans. The category “Investment securities” includes US Treasury securities, municipal bonds, investment grade corporate debt, and preferred stock. Cash flows are semiannual 44 interest payments generated by multiplying principal balances by the reported weighted average portfolio rates and the return of principal at maturity. Static discount cash flows are used with no additional risk spread to compute the present value. 3.3.2 Liability Methodology “Borrowings” includes Federal Home Loan Bank (FHLB) advances, redeemable preferred stock, and subordinated debentures, and others. The present value is equal to the cash flows discounted by the wholesale CD yield. Monthly cash flows are the sum of interest outflow and the return of balances at maturity. Monthly balances are assumed constant for each bucket until maturity, which is the bucket midpoint. For discounting cash flows beyond 6 months the spread between the 6 month CD and 6 month zero coupon Treasury is added to the zero coupon Treasury. Retail and wholesale CDs use the static discount cash flow method. Cash flows are equal to expiring balances and discount rates are wholesale CD yields. In addition, retail CDs have a non-interest cash flow for maintenance, and wholesale CDs have a one month interest penalty per year for early withdrawal. Except for the early withdrawal of wholesale CDs, retail and wholesale balances retire at maturity, which is equal to the bucket midpoint. “Core deposits” includes transaction accounts, money market deposit accounts, and passbook accounts. Cash flows are the sum of interest costs, non-interest costs, and attrition. All balances are assumed to expire after 300 months. Static discount cash flow calculates the present value of these cash flows. The discount rate is the wholesale CD yield as used for previous securities. 45 Various measures of interest rate risk, including SHORT and ISF, have been shown useful in various tests related to equity values. In spite of their success many previous authors have acknowledged their shortcomings. This is particularly true when considering the complexities of a thrift’s asset--liability portfolio. The OTS MVM should provide more and better information on the relationship between interest rate movements and thrift capital. CHAPTER 4 - THE FAILURE PREDICTION MODEL 4.1 Introduction Depository Institutions’ (DIS) asset maturities and liability maturities are significantly different, which leads to a high degree of interest rate sensitivity.9 During the early and mid 1980’s, the interest rate environment was unfavorable toward the banking and thrift industry. Many thrifts failed because of excessive interest rate risk. Lately, risk based capital standards have received significant attention from government regulators and Congress. Interest rate risk (IRR) is included in the risk evaluation of the firm and used to compute adjustments to the capital standards. Historically, methods used to measure IRR have been simple balance sheet ratios. Recently, due to more frequent thrift failure and increasing emphasis on IRR based capital standards, new measurement techniques, in particular the Market Value Model (MVM), have been developed and employed. This study compares the Market Value Model to proxies for IRR using simple balance sheet ratios in measuring interest rate risk and predicting thrift failure. Further, this study tests for an interactive effect between capital and IRR. Failure prediction studies show which characteristics of a firm are related to thrift failure. IRR may have a positive impact on thrift capital during periods of decreasing interest rates, but in general is considered detrimental to thrift solvency. Bovenzi et al. (1983), Barth et al. (1985), and Cole (1993) show an IR variable significant in predicting 9 Flannery and James (1984) and Sweeney and Warga (1986) and others show interest sensitivity of financral institutions. 46 47 thrift failure. The first objective of this study is to confirm these findings on the relationship between thrift failure and interest rate risk. A simple proxy for interest rate risk is the difference between the amount of assets and liabilities maturing in one year. standardized by the total assets (SHORT). Bovenzi et al. (1983) successfully use a similar proxy to predict failure. Barth et a1. (1985), successfully use the ratio of interest sensitive funds to total liabilities (ISF) in failure prediction. Both of these measures have been used successfully in various studies, but neither sufficiently account for the unique characteristics of a thrift’s asset and liability portfolio. Realizing the potentially severe consequences of IRR regulators of banks and thrifts have developed more comprehensive risk measures. These measures consider not only the maturity and coupon of portfolio securities. but also any Option characteristics such as prepayments. The Office of Thrift Supervision (OTS) developed the MVM especially for the purposes of measuring interest rate risk and determining risk based capital standards. ‘0 The IR measure generated by the MVM is theoretically and technically superior and should overshadow the balance sheet ratios. The second objective of this study is to determine the cumulative ability of all three IRR proxies to predict failure. Adding an interest rate risk component to risk based capital standards was a result of the Federal Deposit Insurance Corporation Improvement Act (FDICIA). (The final rule was effective January 1,1994, except for a few amendments.) According to the Federal Register: ‘0 The Federal Reserve Bank has developed an IR measure along similar lines as the MVM. but it is not as sensitive to the interest rate options inherent in mortgage securities. 48 Effective control of interest rate risk is critically important to the safe and sound operation of savings associations. To protect the insurance find and to create appropriate incentives for prudent risk management, thrift’s capital requirements must explicitly take account of interest rate risk exposure. The Federal Register continues by providing a general outline of the requirements of the regulation. The introduction of this regulation is based on the common result that capital level is negatively associated with propensity to fail. 1‘ There are several hypotheses that suggest a non-linear relationship between capital and propensity to fail. First, equity is an option on firm assets. Increasing leverage (decreasing capital) implies a greater sensitivity to firm risk, including IRR. Second, Brickley and James (1986) suggest that firms about to fail, become less sensitive to risk, as a result of potential government laxity. Next, high capital may signal superior management skills and management may alter IRR to their benefit. Also, Cole et a1. (1994) suggest high IRR firms will overestimate book net worth and be less sensitive to net worth levels. Finally, if capital levels are sufficiently high to endure interest rate cycles, the detrimental effects may be avoided completely. Several interactive variables are used to test for variable sensitivity of IRR and capital over other firm variables. The final objective of this study is to test the significance of several interactive terms. In summary, IRR is positively correlated with thrift failure. Comparing the different IRR proxies tested, the balance sheet ratio, SHORT, performs similarly to the MVM proxy. Using all three proxies, multiple IRR variables are found significant, suggesting more than one variable is appropriate to accurately measure IRR. Finally, 1' Sinkey (1978). West (1985). Altman (I977). Barth (1989). Benston (1985). and Pantalone and Platt (1987) show capital negatively related to failure. James (1991) shows capital negatively related to failure losses. 49 consistent with the idea of superior management, the interactive term shows high capital firms less detrimentally affected by interest rate risk. This chapter is organized as follows. The next section reviews the industry environment and the literature on using an IR variable as a predictor of thrift failure. The research methodology is explained in the third section including a brief explanation of the MVM. The fourth section reviews the results. Section five discusses the theoretical motivation and results for an interactive term. The final section interprets and summarizes all the results. 4.2 Review of the Industry and the Literature 4.2.1 Industry Review During the 1980’s, thrift failure occurred at unprecedented rates. Prior to the 1980’s, thrifts failed at a rate of three per year. According to Barth et a1. (1989), from 1980 through 1985 almost 600 thrifts failed. This list of failures includes liquidations, assisted and unassisted mergers, conservatorships, and management consignment programs.12 During the sample period of this test, 1986 through 1991, about 900 thrifts failed, almost 300 in 1990 and about 200 in 1989 and 1991.13 In an economy where thousands of thrifts fail, predicting which thrift specific variables forecast thrift failure is insufficient. External factors must also be evaluated to understand the causes of failure. The macroeconomic factors influencing the thrift failures 12 Liquidations are the sale of the failed thrift’s assets. Mergers occurred both with and without financial assistance. Conservatorships and management consignment programs are similar and represent a transfer of managerial responsibilities to regulatory authorities. ‘3 Closure for this test is represented by conservatorships. conservator-like programs. and receivership. Receivership is a change of ownership. 50 in the late 1980’s include unfavorable interest rate conditions and perverse risk incentives created by regulation. ‘4 In the late 1970’s, interest and inflation rates were at historical highs. To control these rates, the Fed changed monetary policy which created even higher rates and more volatility. Since banks and thrifts use predominantly short-term liabilities, this rate increase dramatically increased the cost of their short-term funding. Interest rate ceilings limited banks’ ability to pay market rate returns on deposits and other intermediaries with lower regulatory overhead were able to offer higher rates of return. This caused what is commonly referred to as disinterrnediation, as investors moved their finds from banks to non-bank intermediaries. Rising interest rates also had a detrimental effect on thrift assets. Banks, in particular thrifts, have long maturity assets which are very interest rate sensitive. The rising rates caused significant decreases in the market value of thrift assets. A secondary impact of these problems was negative net interest margins, as liability cost increased above asset return. Negative net worth often resulted, as the value of thrift assets decreased by more than the equity value. The perverse risk incentives were the result of inefficient regulation and took two forms. Table 1 describes the recent regulations that compounded the incentive problem. The first form, federal deposit insurance premiums were constant across firm type regardless of firm risk.15 Riskier firms, which have a higher probability of loss, paid the same premium as low risk firms. This encouraged firms to bear excessive risk because the 1“ Moral hazard as a cause of failure is discussed in Kane (1989); Barth. Bartholomew. and Labich (1989); Barth et al. (1989): Benston and Koehn (1989); Cole (l990a.l990b); Barth and Brumbaugh (1994) and McKenzie. Cole. and Brown (1992) and others. ‘5 Prior to F IRREA, thrift deposit insurance premiums were 12.5 basis points per year. After FIRREA these premiums rose to 23 basis points, but were still a flat rate. FDICIA introduced a variable rate deposit insurance premium. which was based on the firms capital level and its supervisory rating. 51 thrift owners would receive the increased upside potential of the extra risk and the insurance fiind would bear the increase in downside risk.16 Second, capital requirements were not risk adjusted. Firms were regulated according to maximum leverage, but the riskiness of the assets was not considered in determining acceptable amounts of leverage. Again, firms with riskier assets would provide a higher upward potential for equity holders and the greater downward risk would be transferred to the deposit insurance fund at no cost to the thrift owners. This gave thrift owners an incentive to over-invest in risky assets. These two risk incentives were particularly relevant when firms were insolvent or almost insolvent. ‘7 A healthy firm avoids excessive risk because of the detrimental impact it may have on firm capital. If capital levels are depleted, this disincentive is absent. In response to the severity of D1 risk in the US. and similar concerns across the globe, risk measures were included in various regulations. Depository insurance premiums were restructured to include firm riskiness. Globally, capital standards were set with consideration for the asset risk of the D1. The general premise of risk based capital standards is that as thrift risk increases the required capital level increases. Capital provides a buffer against failure, hence increasing the capital level counteracts increasing asset risk. Recently, these risk-based capital standards have been extended by FDICIA to include IRR. Hence. as firm IRR increases, the capital requirement increases. Although ‘6 Burnett. Rao. and Tinic (1991) show low risk firms subsidize high risk firms with a flat rate deposit insurance premium. '7 This problem is further compounded by a policy of forbearance which the regulators implemented throughout much of the 19805. See Kane and Yu (1994) or Cole (1993). 52 implementation of this concept has been long awaited, the actual details are actively debated. ‘8 4.2.2 Summary of Failure Prediction Studies In an effort to analyze firm characteristics associated with failure, researchers have repeatedly employed a failure prediction methodology. A summary of these studies is provided in Table 2. Generally, these studies use a binomial dependent variable, failure and health. The exact definition of failure is based on accounting numbers and may change over time depending on regulators’ fiinding and the health of the thrift industry. ‘9 Firm specific accounting ratios are the most common independent variables. The dependent variable for most tests has been problem status or failure. The meaning of failure has changed over time. Initially, when failures were few, it meant book value insolvency. Now as failures have become too numerous for regulators to promptly close, failure is more aptly interpreted as closure, because banks stay open well beyond book value insolvency. Some of the recent literature reflects this by renaming these studies as closure prediction.20 Alternately, other studies have predicted book value insolvency.21 Market value insolvency is the best dependent variable for most studies, because it eliminates accounting misvaluations and adjusts for regulatory behavior. For non-public firms, market value data is not available. An alternative continuous market ‘3 The OTS currently uses the Net Portfolio Value Model. NPVM. as a measure of IRR. The NPVM has many similarities to the model used in this study. O’Brien (1992. 1993) and Cordell and Gordon (1992) compare the OTS model to the FRB model. ‘9 The regulators closure rule may also be relaxed by increases in size or percentage conventional assets. by Federal vs. state charter. by stock vs. mutual ownership. or by southwest vs. other districts. See Cole (1993). 2° See Barth et a1. (1989). Thomson (1992). and Cole ( 1993). 2‘ Cole (1993) compares the determinants of failure prediction to the determinants of closure prediction. There is almost complete overlap in significant variables between the two models. but the weights of these variables are much different for the two tests. 53 measure of the degree of failure is the amount of loss thrifts incurred. Barth, Bartholomew and Bradley (1990) predict the resolution costs of thrift institutions. James (1991) predicts the losses realized in bank failures. The independent variables with the most significance in predicting failure and problem status are related to the categories implied by the CAMEL acronym: capital, asset quality, management efficiency, earnings ability, and liquidity.22 Rose and Kolari (1985) suggest liquidity and loan risk exposures are two probable indicators of commercial bank failure. Pantalone and Platt ( 1987b) find that higher earnings and equity ratios are a signal of health. Altman (1977) successfully uses management efficiency as a determinant of savings and loan problem status. Sinkey (1978) shows that the net capital ratio is the most significant classification variable. The efficient market hypothesis suggests all accounting information is included in the market price. Pettway (1980) tests the predictive ability of market price and finds lower stock price returns two years before failure. Additionally, IRR has been shown to be a significant predictor of failure. Bovenzi et a1. (1983) use the difference between market rate assets and market rate liabilities divided by equity capital as an interest rate sensitivity variable. The variable improved the classification of the model, but the efficiency and credit risk variables were more informative. Bovenzi suggests this lack of significance may indicate that interest rate risk is not the most critical problem facing banks or it may indicate this interest rate sensitivity variable is not an accurate indicator of a bank’s interest rate risk. Barth et a1. (1985) find the interest sensitive fiinds variable (ISF) the most predictive of the three interest rate variables they test. 22 More recently. the thrift regulators have used a similar acronym. MACRO. 4.3. Methodology & Data 4.3.1 IRR & MVM The intent of this research is to compare different IRR measures in predicting thrift interest rate risk and thrift failure. IRR measures the value change of interest sensitive securities resulting from interest rate changes. Generally, interest sensitive securities are those with specific contracted cash flows.23 A bond with contracted coupon and maturity cash flows is a good example. When interest rates, and correspondingly the discount rate, increase (decrease), cash flows are discounted at a higher (lower) rate, and the bond value decreases (increases). The value change is greater when contracted cash flow duration is longer; hence, long-term bonds have more IRR than short-term bonds. A depository institution (DI) may be viewed as a portfolio of interest sensitive securities. The assets are mostly long-term loans and the liabilities are mostly short-term deposits. When rates move up, the value of long-term assets decreases more than the value of short-term liabilities. Since net worth is the difference between assets and liabilities, the interest rate increase causes a decrease in net worth. When rates move down, the net worth increases. The larger the maturity mismatch between assets and liabilities, the greater the effect on net worth. In support of this relationship, Samuelson (1945) suggests interest rate sensitivity depends on the mismatch of balance sheet cash flows. Thrifts’ primary assets are thirty year home mortgages which subjects thrifts to a larger maturity mismatch than other DIs. Although IRR will increase net worth when 23 The nominal contracting hypothesis suggests securities with contracted cash flows are interest sensitive because of the rigidity of their cash flows (French et a1. 1983). Flannery and James (1984) are the first to apply this theory to financial institutions. 55 interest rates decrease, it increases the variance of its capital level and increases the probability of insolvency. Measures of portfolio IRR usually include information on the maturity mismatch between assets and liabilities. SHORT adheres directly to this guideline, but ISF is based only on the liability side of the balance sheet. Thrift asset characteristics vary more than thrift liability characteristics, but the omission of the assets in the ISF measure should be detrimental to its performance. Regardless of their simplicity, both SHORT and ISF have been shown significantly related to failure and provide a benchmark for evaluating more complex IRR measures. Saunders (1994) suggests IRR is composed of both a market value effect and a net interest income effect. The regulator's model should account well for the market value effect, while the balance sheet ratios are orientated around the income effect. The second objective of this study is to compare the ability of the different IRR variables to predict failure. The IR measure generated by the MVM is more comprehensive and should provide more predictive ability. Since balance sheet ratios measure predominantly the income effect, they may provide additional information for firms or periods where the income effect dominates the market value effect. There are five IRR variables tested: three generated by the MVM and two simple balance sheet ratios. They are defined as follows: SHORT = (short-term liabilities - short-term assets) / total book assets24 2“ Short term assets include those assets repricing or maturing in less than one year according to section H of the TFR. Short term liabilities include liabilities repricing or maturing in less than one year. A similar proxy standardized by market equity is used by Bovenzi et al. (1983) and F lannery and James (1984a). Market equity is not available for most thrifts and book equity may be very small or take negative values biasing the variable results. 56 ISF = interest sensitive funds / total funds25 Uchg = (MVPE-MVPE“) / Market Value of Assets Dchg= (MVPEd-MVPE) / Market Value of Assets Achg= (MVPE- MVPE...) / Market Value of Assets where MVPE,- Market Value of Portfolio Equity under interest rate scenario i u - 200 bp. upward shock d - 200 bp. downward shock a - actual shock using following December 31 term structure. Note that each IRR is designed such that it is predominantly positive and increasing with more risk. 4.3.2 Failure Prediction My methodology is similar to that of previous failure prediction studies. The logit model is used over the entire population of thrifts. The binary dependent variable, failure, is proxied by firms that entered receiverships, conservatorships, or similar programs and has a value of one. The independent variables selected are those which have been found significant in previous studies. The independent variables are generally represented by the CAMEL acronym, with the addition of an IRR proxy and interactive variables between capital and IRR. The failure prediction model is as follows: Failj=d>(Xj, B, 81') j = 1, 2, ..... N, (4.2) 25 Interest sensitive funds include short term FHLB advances. short term other borrowings, and deposit balances less than $ 100.000. 57 where Failj is equal to one if an institution failed and zero otherwise, (I) is the logit maximum-likelihood operator, X,- is a vector of independent variables representing the C AMEL acronym and an IRR proxy, [3 is a vector of parameter estimates for the independent variables. 81- is a normally distributed random distribution term with zero mean and unit variance, and N is the number of observations. The results from these regressions are in Tables 10 through 35. Each table shows the results of six regressions; one for the pooled sample and one for each of the five years. Variables representing the CAMEL acronym found significant in previous studies are included to reduce specification error. Since accounting variables are integrally related, an accounting based IRR proxy may inadvertently proxy for other firm risks. This would bias the results of the proxy and the importance of IRR. Including variables representing the CAMEL acronym will reduce the probability of misspecification. Each risk represented by the CAMEL acronym is proxied by several variables. To determine the information each variable adds. it is substituted into a base set of CAMEL variables. The base variables representing the CAMEL acronym are capital (net worth / total assets),26 asset quality (repossessed assets / total assets), management efficiency (non- interest expense / total assets), earnings ability (net income / total assets), and liquidity (amount eligible for regulatory liquidity / total assets).27 The likelihood ratio index (LRI) measures regression performance. An increase in LRI resulting from a substitution of a 2‘5 An alternate specification of this proxy is exp (net worth / total assets). This alternate specification is less skewed and more closely fits the normality assumption of the logit model. 27 Amount Eligible for Regulatory Liquidity- includes Insured or Guaranteed by an Agency or Instrument of the United States (A070), Accrued Interest Receivable (A090), Cash and Demand Deposits (A360), US. Government and Agency Securities (A370). Common and Preferred Stock (A382). Other Investments (A3 84). and Accrued Interest Receivable (A390). 58 CAMEL variable implies an improved specification of that variable. The optimal CAMEL variables are those that maximized the LRI over each different variable. The various specifications of each CAMEL and IRR are listed below along with their expected sign. The specifications and expected signs are as follows: Independent Variables Capital- net worth / total assets - MVPE / market assets28 - MVPE / total assets - exp (net worth / total assets) Asset Quality- repossessed assets / total assets - other assets / total assets - percentage change in median real estate prices Management Efficiency- non interest expense / total assets Earnings Ability- net income / total assets - interest income / interest expense Liquidity- amount eligible for regulatory liquidity / total assets - brokered deposits / total assets - cash and securities / total assets Interest Rate Risk- Uchg - Dchg - Achg - ISF 23 Market assets is the estimated market value of assets as computed by the MVM. Expected Sign H H (+) (-) (+) (+) (t) (+) 59 - SHORT (+) This study will estimate parameters using year-end financial statement data from 1985 through 1989. Failed thrifts are from either the following calendar year or the subsequent year, months thirteen to twenty-four. For example, if ten thrifts fail in 1986, estimation procedures use the data from the ten failed firms and the nonfailed firms from, December 31, 1985. This tests failure zero to twelve months into the future, year one failures. To evaluate the significance of the independent variables with a longer time until failure, the above regression is run with failures from months thirteen to twenty-four, year two failures. For example, independent variables from December 31, 1985, will be used to predict the failures in 1987. Failure data is from 1986 to 1991. Year one and year two failures are predicted from each of the five years individually, but a year by year analysis may be significantly affected by temporary phenomena and may not show consistent results. The five years of Thrift Financial Report (TFR) data are pooled to form a single data base of approximately 15,000 observations. Outliers were eliminated for impossible or excess values. Excess values are those more than four standard deviations from the mean. About 300 observations are eliminated. Net worth / total assets is restricted to values greater than or equal to -1. MVPE / market assets is restricted to values greater than -1. MVPE / total assets is restricted to values greater than -1. The repossessed assets ratio is restricted to values greater than or equal to 0 and less than or equal to 1. Other real estate / total assets is restricted to values greater than or equal to O and less than 1. Non-interest expense / total assets is restricted to values greater than or equal to 0. Net income / total assets is 60 restricted greater than -1. Interest income / interest expense is restricted to values greater than or equal to 0. Amount available for regulatory liquidity is restricted to values greater than or equal to 0 and less than 1. Uchg is restricted to values greater than -. 1. Achg is restricted to values greater than -.89. SHORT is restricted to values between -.8 and 1. ISF is restricted to values greater than 0 and less than or equal to 1. Using a single IRR variable in each regression provides information on the individual performance of the variable. Including all three IRR variables will provide more information on their cumulative performance. The results from regressions including all three IRR variables are in Tables 38 and 39. 4.3.3 Data The data for independent variables is from the Thrift Financial Report from year- end, 1985 through 1989.29 Section H of this report provides the maturity and yield data necessary to compute the MVM.30 The failure data is compiled from lists provided by the Federal Deposit Insurance Corporation (FDIC) and OTS. 4.4 Results This section provides results for two tests. The first verifies the previous result that IRR is positively correlated to thrift failure. This result is verified over year by year data and over the appended five year data set. Further, various specifications for the 29 Observations are eliminated based on eligitimate or excessive values for any variable. Eligitimate variables are those which are non-sensical based on the possible values. Excessive variables are those which exceed approximately four standard deviations from the mean of the entire sample. 30 The MVM also required data on the one month Treasury which was obtained from CRSP. on the Treasury yield curve which was obtained from CitiBase. and for category market securities which were obtained from many different sources. 61 C AMEL variables are compared over both the single and five year data sets. Based on the regressions using the pooled sample, optimal CAMEL and IRR variables are identified. Then this Optimal set of variable specifications is tested. Next. a failure prediction model with all three IRR variables is tested to confirm which IRR proxy is the most significant and to analyze IRR performance using multiple proxies. This test is only on the five year data base. Each of these failure prediction tests uses failures in the twelve months following the TFR data (year one) and failure in the thirteen to twenty-four months following the TFR data (year two). Table 8 shows the median, mean, and standard deviation of the variables tested. For the first test of the failure prediction model, a base set of CAMEL variables is used as a standard of comparison. In each regression, a single variable is changed. The results are then compared to the results for the base model to determine the best specification of each variable. Each table includes results for six regressions. Only the results for the pooled sample are discussed. In Table 9, the year one results for the base CAMEL and base IRR variable are shown. Four C AMEL variables, CAEL, and Uchg are of the correct sign and significant. The management variable is of the correct sign but only of marginal significance. The management variable is seldom significant in these regressions. The likelihood ratio index (LRI) for this regression is .273. The year two results for the base CAMEL and base IRR variables are shown in Table 10. These results show all five CAMEL variables and Uchg of the correct sign and significant when predicting failure. The LRI for this regression is .168. 62 The results for the base CAMEL variables with the Dchg IRR variable are shown in Tables 11 and 12 for year one and two respectively. For year one failures, the four CAMEL variables. C AEL, and Dchg are of the correct sign and significant. For year two failures, all five C AMEL variables and the IRR are correctly signed and significant. Compared to Uchg, the variable Dchg slightly decreases the LRI for year one and increases the LRI for year two. The IRR variable, Achg, is included with the base CAMEL variables in the next two tables. Table 13 has year one results. The four CAMEL variables, CAEL, and Achg are significant and the LRI is slightly higher than for Uchg. Table 14 shows the year two results that the five CAMEL variables and Achg are significant and that the LRI is again slightly higher than that for Uchg. However, Achg is not an ex ante measure, but uses ex post interest rates. This result shows a small improvement in performance when actual interest rate changes are used instead of a hypothetical parallel shock. ISF, interest sensitive funds divided by total assets, replaces the previous [RR in a regression with the base CAMEL variables. These results are in Tables 15 and 16. For year one failures, the regression including ISF shows significant CAMEL and IRR variables except for the management variable. The LRI decreases slightly, hence Uchg is still a more informative IRR variable. With year two failures, the CAMEL variables are all significant, but ISF has an incorrect sign. The LRI’s for Uchg and ISF are identical. The final IRR variable tested, SHORT, the difference between short term liabilities and short term assets divided by total assets, is included with the base CAMEL variables. The results on Tables 17 and 18 show the four CAMEL variables, CAEL, and the five 63 CAMEL variables significant for year one and year two failures respectively. SHORT is significant in both years. The LRI for SHORT is identical to the LRI for Uchg in both years. Table 19 has results for using an alternate capital specification. MVPE divided by market assets. All five CAMEL variables and the IRR are significant and the LRI is .278 compared to .273 with the base capital variable. Table 20 shows the corresponding results for year two failures. All five CAMEL variables and the IRR are significant and the LRI is again improved over the base model from .168 to .170. Another specification of the capital variable, MVPE over total assets, is evaluated next. All five CAMEL variables and the Uchg IRR variable are significant in both years. The LRI improves in both years to .288 for year one and to .174 for year two failures. These results are shown in Tables 21 and 22. The most informative capital specification is exp (net worth / total assets). This capital specification reduces the impact of extreme values and creates a more normal distribution. When included with the other base CAMEL variables and Uchg, all variables are significant and the LRI is .291, the highest LRI of any single alternate specification. For year two failures, all CAMEL variables and Uchg are significant and the LRI rises to .177. These results are shown in Tables 23 and 24. An alternate specification for asset quality is other real estate1 divided by total assets. For year one failures, the inclusion of this variable with the other base CAMEL variables and Uchg shows Uchg and the three variables, CEL, significant and a decrease in LRI to .270. This result is in Table 25. For year two failures, the inclusion of the 3‘ Other real estate is the physical assets owned by the financial institution as overhead. 64 alternate asset variable shows Uchg and all five CAB/[EL variables significant and a decrease in LRI to .156. This result is in Table 26. Percentage change in real estate value is the best specification of asset quality. Table 27 shows the four CAMEL variables, CAEL, and Uchg significant with an LRI of .277. Table 28 shows the five CAMEL variables and Uchg significant with an LRI of .172 for year two failures. Table 29 has results using interest income over interest expense for the earnings variable. The four CAMEL variables, CAML, and Uchg are significant in predicting year one failures, but the LRI is very low at .258. Table 30 shows CAML, the four CAMEL variables, and Uchg significant for year two failures, again with a very low LRI at .166. An alternate liquidity specification is brokered deposits over total assets. Results are in Tables 31 and 32. When included with the other base CAMEL variables and the base IRR proxy, CAEL, the four CAMEL variables and Uchg are significant. When included in the regression predicting year two failures, all base CAMEL variables and Uchg are significant. The LRIs for year one and year two failures are .254 and .143 respectively. In Table 33, the base CAMEL and Uchg variable are included with another alternate liquidity variable, cash and securities over total assets. The results of this regression predicting year one failures shows the four CAMEL variables, CAEL, and Uchg significant. The results of this regression, predicting year two failures, shows the five CAMEL variables and Uchg significant. This result is in Table 34. The LRIs for this alternate variable are .254 and .140 for year one and year two failures respectively. 65 The results of the previous regressions suggest the optimal CAMEL specification for year one and year two failures includes the following variable specifications: exp (net worth / total assets), percentage change in median real estate value, non-interest expense divided by total assets, net income over total assets, and amount eligible for regulatory liquidity divided by total assets. For the two year failures, ISF has the wrong sign and SHORT is slightly more significant then Uchg. Uchg is the optimal one year IRR proxy and indistinguishable from SHORT in performance for year two. Uchg will be used as the optimal year two IRR proxy. Table 35 shows the results of year one failures using the pooled data set. This regression has the highest LRI of .297 for any six variable, CAMELI, specification. The four CAMEL variables, CAEL, are significant as well as Uchg. Table 36 shows the result of year two failures using the pooled data set. For this regression, all five CAMEL variables and ISF are significant. Evaluating the sign and significance of the independent variables for individual years shows that significance varies widely and even signs are somewhat volatile. For example, the sign of the IRR variable is frequently wrong in 1985 and 1986, but generally insignificant. This may suggest an inferior proxy or changing sample properties. Proxies may vary in their ability to measure failure causing risks; better proxies may solve the problem. Also, economic factors could affect the significance of accounting ratios. Including economic variables or interactive terms between economic factors and accounting ratios in the regression may increase parameter consistency. If certain economic factors are more significant in a given year, then an associated risk measure may be more significant. For example, when interest rates are high and volatile IRR is a 66 greater threat to thrift health. A complication is that economic variables probably have a longer cycle than one year. Alternatively, parameter significance may vary because of an arbitrary closure rule. This may be corrected by using an accurate market value estimate of equity for the closure rule or by using failure loss or net income as the dependent variable. The second test evaluates the comparative and additive abilities Of three IRR proxies: Uchg, ISF, and SHORT. For each regression either the base or optimal CAMEL variables are included with all three IRR proxies. If several of the IRR proxies are significant, this suggests that interest rate risk may best be explained by multiple measures. Table 37 shows the results of predicting year one failures with the base and optimal CAMEL variables along with the three IRR proxies. In the regression with base CAMEL variables, the four variables, CAEL, are significant. Uchg and SHORT are significant and ISF is marginally significant. In the regression with the Optimal CAMEL variables, the four CAMEL variables, C AEL, are significant. Uchg is the only significant IRR variable. ISF is marginally significant and SHORT is insignificant. With the base CAMEL variables, several IRR measures are significant and provide information on the IRR of the firm. With the Optimal CAMEL variables, ISF is marginally significant and may provide additional IRR information over that provided by the Uchg measure. Table 38 shows the results for base and Optimal CAMEL variables with the three IRR proxies for predicting year two failures. All of the base CAlvflEL variables are significant. Uchg is marginally significant. SHORT is insignificant and ISF has the wrong sign. All the Optimal CAMEL variables are also significant. The results for the IRR variables with the optimal CAMEL variables are Uchg is clearly significant and SHORT 67 and ISF are insignificant. These results suggest that for year two failures Uchg provides the most information on IR. 4.5 Interaction Term Risk based capital standards require higher capital levels as a result of higher IRR levels. Higher capital levels have been found by Sinkey (1978) and others to be negatively related to failure. Increasing capital standards for firms with high IRR effectively neutralizes the IRR and returns the probability of failure to acceptable levels. Alternatively, firms with lower capital are more sensitive to IRR. The following theories analyze this association. Further, the theories are applied to a potential interactive effect between IRR and asset size. The first theory of interaction between capital level and the effect of IRR is the leverage theory. Low capital is a result of high leverage. The equity of leveraged firms is an option on the value Of firm assets and the equity has a greater volatility than the volatility of the assets. This suggests that as capital increases IRR should have a lower impact on the probability of failure. Similarly, as IRR decreases, the sensitivity of thrift failure to capital level should decrease. Brickley and James (1986) suggest that government guarantees may reduce the co- movement Of S&L stock returns with their underlying assets. During periods of financial distress, regulators may not fail a technically insolvent thrift in hopes it will recover without government expenditure. This regulatory leniency provides a free Option to insolvent thrifts. Hence the market value is buoyed up by the value Of the free option 68 provided by regulators and market value sensitivity of vulnerable Dls to risk is reduced. Vulnerable DIs include poorly capitalized firms and firms with excess IRR. According to Brickley and James, both of these firm characteristics should be associated with lower sensitivity to risk. Hence, low capital firms should be less sensitive to IRR. Further, as IRR increases and the probability of failure increases, the value of government leniency increases which would make capital less beneficial for high risk firms. As the probability of government intervention increases with higher risk levels, the value of capital is supplanted by government laxity. Therefore for high IRR firms, the coefficient for capital should be negative, but less significant for low IRR firms. Finally, as firm size increases, government laxity becomes more significant and more probable; hence, IRR or other risk becomes less influential. Alternatively, high capital levels may be a signal Of management quality. If higher capital is a signal of superior management, then thrifts with higher capital levels should be more resilient to the detrimental effects of IRR. A well-managed firm may hedge risk if they predict an upward movement in interest rates and, alternatively, leverage risk if they predict a downward movement in interest rates. Thus a well-managed firm, that correctly predicts interest rates, may reduce its sensitivity to IRR and even use IRR to its benefit. As IRR or firm risk increases, the value of good management increases. As the firm’s position becomes more precarious, the value of good decisions is more important. For higher risk levels the capital variable should be more significant. Finally, larger firms can afford better and more management. Their ability to deal with IRR should be superior. 69 For example, large firms should have less costly hedging tools. Therefore, larger firms should be less sensitive to IRR. Cole et al. (1994) suggest, following periods of rising interest rates, the book value of net worth more severely overestimates the value of net worth for firms with higher IRR. The greater the IRR the greater will be the effect Of interest changes. Book value net worth is negatively related to failure. High IRR firms will show a less negative relationship to net worth. Alternatively, high capital levels may increase the long-run stability of the firm. Since interest rates mean revert, a detrimental upward shift in rates is usually followed by a beneficial decrease in rates. If a firm has sufficient capital to endure the cycle, the effect of IRR may be temporary, hence nonexistent at the cycle’s end. Therefore, firms with high capital levels should be less sensitive to IRR. Also, large firms would have increased access to capital and liquid assets and could endure adverse business cycles easier than smaller firms with fewer options. For larger firms, increased capital access or increased liquidity access may serve as a temporary substitute for high capital level. Larger firms should be less sensitive to IR. There are four specifications for interactive effects. The first interactive term is the product of [RR and capital level. The capital level is specified as exp(net worth / total assets) to eliminate negative capital values and to give the interactive term a more manageable distribution and interpretation. The second specification is the product of Uchg with sequential capital level dummies. The sequential capital dummies represent the following capital levels: C1 < 0, 0 3 C2 < .02, .02 3 C3 < .05, and .05 5 C4. The next 70 interactive effect is measured by the product of capital with sequential dummies for Uchg. The dummies for Uchg are for the following levels: IRR < .02, .02 s IRR < .04, and .04 s IRR. The final specification is the product of Uchg and asset level dummies. The asset levels are less than 100 million, between 100 million and 1 billion. and greater than 1 billion. To test the interactive effects, the failure prediction model is as follows: Failj=<1>(Y-, l3, Sj) j = l, 2, N, (4.3) where F ailj equals one if an institution fails and zero otherwise, (I) is the logit maximum- likelihood operator, Yj is a vector of independent variables representing the CAMEL acronym and interactive terms, 13 is a vector of parameter estimates for the independent variables, a,- is a normally distributed random distribution term with zero mean and unit variance, and N is the number of Observations. The expected sign on the product between Uchg and capital level is negative. The other expected signs are as follows. Capital Level Dummies (increasing) Theory IRR sign Trend Leverage (+) decreasing Regulatory Leniency (+) increasing Management Proxy (+/-) decreasing Overestimated Net Worth (+) not applicable Cyclical Interest Rates (+/0) decreasing Theory Leverage Regulatory Leniency Management Proxy Overestimated Net Worth Cyclical Interest Rates Asset Level Dummies (increasing) 71 IR Level Dummies (increasing) Capital sign (-) (-) (-) (-) (-) Theory IRR sign Leverage (+) Regulatory Leniency (+) Management Proxy (+) Overestimated Net Worth (+) Cyclical Interest Rates (+) Trend not applicable increasing decreasing increasing decreasing Trend not applicable decreasing decreasing not applicable decreasing Tables 39 through 43 show the results for regressions including the base CAMEL variables, an IRR variable, and the product of IRR and capital. The results for year one failures are on Table 39. There is a separate regression for each Of the three IRR variables: Uchg, ISF, and SHORT. For all three regressions, the IRR and the three C AMEL variables, AEL, are significant. Management is marginally significant with the Uchg variable but not significant Otherwise. Using SHORT, the capital variable is significant. With ISF or Uchg, the capital variable is not significant. The IR variable and 72 the interaction term are significant with Uchg and ISF. IRR is expected to be positively correlated with failure, as more risk increases the probability of low capital levels. The interaction term shows the sensitivity Of failure to IRR is a decreasing function of capital. Surprisingly, with SHORT the [RR and interaction term have the wrong sign and this LRI is the highest of the three. Table 40 uses the base CAMEL variables and shows the results for each regression predicting year two failures. The four CAMEL variables, AMEL, are significant in each of the three regressions. For year two failures, the capital variable performs similarily to the capital variable for year one failures, over each IRR. Also, the sign and significance of the IRR variables is the same for year one and year two failures. The Uchg and ISF variables and their corresponding interactive terms are significant with the correct sign, but SHORT and the corresponding interactive variable are the wrong sign. The LRI of the SHORT regression is .219, the highest of the three IRRs. Tables 41 and 42 present the results for the regressions including the interactive term, but use the optimal CAMEL variables in place Of the base CAMEL variables. Using the optimal CAMEL variables increased the LRI for each regression in both years, but did not reduce the superiority of SHORT with a LRI of .352 for year one and .229 for year two. Predicting year one failures shows the four CAMEL variables, CAEL, significant in each regression. With the optimal C AMEL variables, the performance of the IRR and interactive variables is unexpected. Each variable in all three regressions is insignificant. Predicting year two failures the results are identical except the management variable is also significant. 73 As an alternative to the interactive product, sequential dummy variables multiplied by Uchg are used to distinguish non-stationary IRR sensitivity over different capital levels. This result is in Table 43. A different dummy variable is used for each of the groups: less than 0%, 0% to 2%, 2% to 5%, and greater than 5%. Base and Optimal CAMEL variables, Uchg. and the product of the dummies and Uchg are used to predict year one failures. The three CAMEL variables, CEL, are significant in the regressions with either CAMEL proxies. The management variable is marginally significant in both regressions and the asset quality variable is significant only with the optimal CAMEL variables. Uchg and each Uchg dummy are significant for both CAMEL specifications. The hypothesis that the parameters for each dummy variable are equal is rejected by a likelihood ratio test. The test statistics are 984.6 for the base CAMEL variables and 857.4 for the Optimal C AMEL variables. For three degrees of freedom, using the Chi-squared distribution these statistics are significant at the one percent level. The results for these variables suggest that IRR is most detrimental for low capital level firms. Both high capital level groups show a negative correlation between failure and IRR suggesting that for high capital firms IR is beneficial. The LRI for both regressions is above .400, the best of any regressions. The AIC, which accounts for the number of parameters in the regression, is the most favorable Of all regressions. Table 44 shows that the parameters for the three lower capital levels are significantly different than the parameter for the high capital level. Table 45 considers the sensitivity to IRR over different size categories. Included in the regressions are the base or Optimal CAMEL variables along with Uchg and asset level dummies representing the size categories: less than $100,000,000, between 74 $100,000,000 and $1 billion, and greater than $1 billion. Predicting year one failures, both regressions show the four CAMEL variables, CAEL, significant. The Uchg variable is significant in both regressions. The Uchg dummy representing the smallest firms is marginally significant with the base CAMEL variables and significant with the optimal C AMEL variables. The Uchg dummy for intermediate size firms is marginally significant for both CAMEL specifications. The hypothesis that the parameters for each dummy variable are equal is not rejected by a likelihood ratio test. The test statistics are 2.3 for base CAMEL variables and 3.2 for Optimal CAMEL variables. For two degrees of freedom, using the Chi-squared distribution these statistics are not significant. This is an unexpected result. Larger firms should have the resources to better manage IRR. The ‘too big too fail’ hypothesis suggests large firms have a lower propensity to fail. My results do not support this hypothesis. A possible explanation for this is prior to this period more large insolvent thrifts were kept Operating than small insolvent thrifts, due to the greater costs of failing large firms; possibly, this period is characterized by a reversal of this trend and an increase in large thrift failures. Table 46 shows variables similar to Table 45, except the size dummies confirm the results of the likelihood ratio test that the . IRR response of the largest and smallest groups of thrifts are not significantly different. The results using optimal CAMEL variables are similar. In Tables 47 and 48, a series of IRR dummy variables test for a non-constant sensitivity of failure to capital, over different levels of IRR. Three different capital specifications are tested: net worth / total assets, MVPE / market, and exp (net worth / total assets). Using net worth / total assets. failure has a decreasing sensitivity to capital 75 as [RR increases. Table 48 shows that each IRR group has significantly different capital sensitivity, but only when using the net worth / total assets specification. The hypothesis that the parameters for each dummy variable are equal is rejected by a likelihood ratio test only for the capital specification, net worth / total assets. The test statistics are 6.3 for net worth / total assets, 1.7 for MVPE / market assets, and 0.4 for exp(net worth / total assets). For two degrees of freedom, using the Chi-squared distribution the statistic for net worth / total assets is significant at the five percent level. The regression results further show the three CAMEL variables, AEL, significant as well as the Uchg variable. Table 49 includes all specifications of all CAMEL and IR variables as well as a hedging variable and a size variable. All capital variables are significant with exp (net worth / total assets) most significant. Repossessed assets and percentage change in real estate are significant with the best performance by the latter. Management has the correct sign. Net income / total assets is the only consistently significant earnings variable and amount eligible for regulatory liquidity is the only consistently significant liquidity variable. Surprisingly, the hedging variable is positive and significant in all three regressions suggesting those firms with the most hedging devices had a higher probability of failure. Size is also shown positively correlated to failure. Uchg is significant in all three regressions and clearly outperforms balance sheet ratios. 4.6 Summary In contrast to previous failure prediction studies, the performance of the capital variable is inconsistent in this study. The inconsistency may result from technically 76 insolvent firms that are not failing. Further, several interactive terms include capital level as a factor and may reduce the information provided by the capital variable. The CAMEL variables performed as expected in the majority of regressions regardless Of the [RR variable. In support Of our predictions, IRR is positively correlated with failure for all IRR measures. An argument to the contrary is, during interest rate declines, a firm with high interest rate risk should benefit. This argument finds some support when regressions are run for specific years with decreasing rates, but not over the full five year sample. The theoretically and technically superior MVM performs similarly to the balance sheet ratio SHORT when the IRR proxies are used individually or together. Although SHORT includes information on thrift asset maturity, it does not distinguish between different long term maturities and different security characteristics. The performance of ISF is good for year one failures; this is surprising considering that it is a liability side ratio. The performance of the MVM is probably detrimentally affected by a failure rule that depends on accounting versus market values and a policy of forbearance. Using a consistent and timely market measure of insolvency should improve the comparative performance of Uchg. Using three IRR variables provides information on the additivity of the IRR information. Predicting year one failures, Uchg, ISF, and SHORT are positive and significant. This suggests multiple measures of IRR are helpful in determining risk exposure and may be useful in risk based capital standards. The t-statistics in this regression show SHORT most significant. 77 The results with the interactive variables appear to provide support for risk based capital standards. As the level of capital increases, the detrimental effect of IRR is reduced. This suggests a dual benefit of higher capital. The first benefit of capital is as a buffer against equity loss. The second benefit is that as capital increases, the sensitivity to IRR decreases. For the two highest capital levels, IR is negatively associated with failure. The relationship between high capital and [RR may result from more efficient management. If this is the case, requiring higher capital levels will not improve management effectiveness and may not reduce the sensitivity of failure to IRR. Another implication of the results Of the interactive variables is that for high capital levels, IRR increases the probability of health. Hence, a high capital firm has a lower probability of failure with higher IRR. Brickley and James suggest during financial distress government intervention may buoy up capital values. During the sample period for this test, the regulators may have tightened their closure rule, eliminating any previous value from regulatory laxity. This would explain the IRR sensitivity Of the two lower capital levels, but not the beneficial response of the higher capital levels. Similarly, the leverage argument can explain higher sensitivity for low capital levels, but not the beneficial sensitivity for highly capitalized thrifts. Management quality is the only hypothesis which justifies the current result on the interaction between capital and IRR. Different size firms are shown to have similar sensitivity to failure from IRR. This is a surprising result and not suggested by any of the hypothesis stated here. A possible explanation is a reversal of the regulatory leniency / too big to fail hypotheses. This time period may be characterized by regulator responsibility. The prompt response of 78 regulators may have removed the market value gains resulting from leniency. This may be particularly true for large firms where lenient regulator behavior was more probable. This is supported by a proxy for firm size shows significant and positive correlation to thrift failure. Interest rate risk is shown to impact the sensitivity of failure to capital level when capital is measured with book values. Two hypotheses may support this result: overestimated net worth and regulatory leniency. The regulatory leniency is rejected in this time period in other tests, hence the overestimation of net worth hypothesis may best explain why firms with high IRR levels are less sensitive to capital. Further, this hypothesis is based on book value measurement Of capital which suggests the insignificant results with non-book measures of capital do not contradict the hypothesis. CHAPTER 5 - TWO INDEX MODEL Depository institution (DI) stock returns have been shown to be sensitive to an interest rate index. The primary explanation for this result is the nominal contracting theory. Inherent in this theory is the suggestion that maturity mismatch increases interest rate sensitivity. Thrifts have significant maturity mismatch and should show a significant interest rate index and subsequently a significant interest rate risk (IRR) proxy. Using a two index model, both coefficients are shown in the literature to be significant in predicting DI stock returns. 5.1 Introduction The most prevalent of the theoretical explanations suggesting interest sensitivity for D1 stock returns is the nominal contracting hypothesis.32 The nominal contracting effect is compounded by the maturity mismatch problem. The impact on a nominal long maturity financial contract is greater than the impact on a short maturity contract. Thrift assets are Often long-term fixed rate contracts and their liabilities are short-term liabilities. When interest rates rise, the values of both contracts go down, but the long-term contract will lose much larger percentages of its market value. Thrift market value of equity is the difference between the market value of assets and liabilities. Hence when interest rates rise, asset value decreases more than liability value and market equity decreases. 32 French et al. (1983) originate this hypothesis. F lannery and James (1984) are the first to use the hypothesis in connection with the interest rate sensitivity of financial institutions. A more thorough discussion of the relevance of this hypothesis is found in the Flannery and James paper. 79 80 Maturity mismatch measures the difference between asset maturity and liability maturity. Flannery and James (1984) and Others have used SHORT, market rate assets minus market rate liabilities standardized by market equity, as a maturity mismatch proxy.33 Another balance sheet measure Of maturity mismatch is interest sensitive funds (ISF), those funds that will reprice within one year standardized by total funds.34 Uchg is the Office of Thrift Supervision’s measure Of maturity mismatch and considers maturities for all asset and liability cash flows including those which are interest rate options. It is technically and theoretically superior to balance sheet ratios and should provide the most accurate measure of maturity mismatch.35 5.2 Methodology The two index market model examines the relationship Of both a market and an interest rate index with stock price returns. The two index model is as follows: R11: 1301+ ijRmt + [31er1+ 811- (5.1) where, Ry, - the excess holding period return to the j‘h stock or portfolio over the month ending at time t, R... - the excess holding period return on the S&P 500 over the month ending at time t, 33 Bovenzi (1983) and Tarhan (1984) also use this measure. 34 Barth et al. (1985) use this measure of maturity mismatch to predict failure. 35 Maturity mismatch is the difference between asset maturity and liability maturity and is used as an IRR proxy. This discussion uses maturity mismatch as a synonym for IRR proxy. 81 R1. - the excess holding period return on an index Of constant maturity default- free bonds over the month ending at time t,36 and e,~.- error term. The relationship between stock price and an interest rate index is well established, including a relationship between interest sensitivity and a maturity mismatch factor. Maturity mismatch is a measure Of interest rate risk (IRR). The intent of this research is to extend our understanding of the dependence of the interest rate coefficient on an IRR factor and to determine which IRR factor is most predictive. There are two tests of this relationship; the first is a two stage test. The first stage runs the two index model, which verifies the relationship between the stock of financial institutions and interest rate movement. The second stage regresses the interest rate index coefficient onto the IRR measure as follows: 131i = a0 + (1le + Vj, (5.2) where, IRR]- - the j’h portfolio’s or S&L’s average IRR over the test period”, SHORT - (market rate liabilities minus market rate assets) / total assets”, ISF - interest sensitive funds / total funds, 36 These interest rates are the unexpected changes in interest rates Obtained by using the residual of an AR(3) model. F lannery and James (1984), Tarhan (1984), and Others use similar methodology. Previous studies have orthogonalized the market or the interest rate index. but this biases t-statistics. Giliberto (1985) discussed this issue in depth and concludes the unorthogonalized model provides unbiased results. These tests use unorthogonalized indices. 37 Average IRRs are the average of lRRs over the months that data was available for the portfolio or stock. For example: the average IRR for a firm trading from June 1986 to September 1989 would be the average IRR over that specific trading period. 33 Short is standardized by total assets instead of market equity. which is more common in the literature because the market equity data is not available for the entire sample. Also. book equity takes negative and very small values and would make the variable distribution volatile. 82 Uchg - (MVPEu-MVPE) / estimated market assets, and Vj - error term. Savings and loans (S&Ls) are grouped into portfolios by similar characteristics to test interest sensitivity Of specific groups of firms. Portfolios are formed by total assets, net worth percentage, and thrift assets as a percentage of holding company assets. The second test combines the two stages into a single regression as follows: R11: 1301+ BmiRmt+ 1321(R11*mRi)+ BuRn + 51:» (53) where, R). - the excess holding period return to the j‘h stock over the month ending at time t, Rm - the excess holding period return on the S&P 500 over the month ending at time t, Rn - the excess holding period return on an index of constant maturity default- free bonds over the month ending at time t, IRR,- - the jlh S&L’s average IRR over the test period, SHORT - (market rate liabilities minus market rate assets) / total assets, ISF - interest sensitive fiinds / total funds, Uchg - (MVPEu-MVPE) / estimated market assets, and Ejt‘ error term. The interest rate index is the return on a portfolio Of bonds. When interest rates rise, bond prices decrease and the return index is low. Rising interest rates cause a similar 83 decrease in S&L asset values leading to low S&L equity returns. Hence, the return on the interest rate index is expected to be positively correlated to the return on S&L common stocks. In the second stage of this test. interest rate sensitivity is regressed onto an IRR proxy. As IRR or maturity mismatch increases the responses to changes in interest rates, interest rate sensitivity, should be larger. Hence, the coefficient of the IRR proxy is expected to be positive. The expected signs for the coefficients for portfolios or individual stocks are: Independent Vam Expected Sigg Two Index Model (Stage one) Market Index (+) Interest Rate Index (+) IRR Model (Stage two) IRR Proxy (+) Single Stage Model Market Index (+) Interest Rate Index (+) Interactive Term - (R1.*IRR,-) (+) The return data for the these tests are monthly returns from the Center for Research on Security Prices (CRSP). All firms are on the New York Stock Exchange, American Stock Exchange, or National Association of Securities Dealers Automated Quotation system with at least 3 years Of consecutive return data from 1986 through 1990. The data for the market, interest rate. and risk-free indices are from Ibbotsen and 84 Associate’s Encorr. The balance sheet data for the IRR measures are from the Thrift Financial Report, December, 1985-1989. 5.3 Results Preliminary tests on the interest sensitivity of D1 stock prices are not supportive Of previous results or the corresponding theory. Results are in Table 50. The tests used the two index model to determine the interest sensitivity of equity portfolios of savings and loans. The savings and loan portfolios are found not sensitive to an interest rate index. These results are consistent over short, long, and mortgage rate indices. The coefficient of the long index is insignificant and the short index coefficient is significant with the wrong sign. The market index is significant with the expected sign and magnitude. Table 51 shows the results of the single stage regression. The market index is consistently significant with a coefficient estimate of slightly less than .6. The interest rate index is insignificant as in the two stage model. Finally, for both interest indices and all three IRR proxies the interaction term between the interest index and the IRR proxy is insignificant. The literature frequently suggests hedging as an explanation for a lack of sensitivity. If the S&Ls have hedged their interest rate risk, they will not respond to movements in the interest rate index. Table 49 shows results for a hedging variable in a failure prediction test. This test shows hedging securities positively correlated to failure. This result is appropriate when interest rates are decreasing. The period of this study is characterized by mildly fluctuating rates with no large increases. Hedging benefits the firm 85 during periods of increasing rates. Although hedging instruments were legal over the sample period, even the larger thrifts did not use substantial amounts of hedging instruments. Hedging securities may have provided limited portfolio diversification. Along similar lines. Off-balance sheet items may have played a role in reducing interest sensitivity. Kane and Unal (1990) show on and off-balance sheet items have Opposite responses to changes in interest rates. If the sample period is characterized by higher amounts of Off balance sheet items or greater sensitivity to these items, then DI’s may not be sensitive to interest rate changes. Data on off-balance sheet items are only available for the final year Of the test period. Swaps are included in Off-balance sheet items and are a good example Of an instrument which would reduce interest rate risk. For example, an S&L may exchange assets with fixed interest payments for assets with variable interest payments. However, Kane and Unal (1990) show even through various changes in interest rate sensitivity, the Off-balance sheet effect never dominates the on- balance sheet effect. Brickley and James (1986) show risk sensitivity is muted by a lag in regulatory intervention. As a firm approaches failure, its capital value is inflated and its interest rate sensitivity may be reduced if regulators actions are delayed or neglected. This problem should not apply to well capitalized firms that are in no danger of failing. A portfolio of firms with capital levels above 5% (HIGHCAP) and a portfolio of intermediate capital levels between 2% and 5% (MEDCAP) showed no interest sensitivity. Hence delayed intervention does not explain the lack Of interest sensitivity. 86 Non-synchronous trading may be a problem when dealing with small firms as in the S&L industry. This phenomenon has a tendency to affect the magnitude and significance Of the market model coefficients. However, in most tests the market beta was of the correct magnitude and significance. Regardless, to correct for this potential problem the monthly returns were compounded into quarterly returns. The returns were still unresponsive to interest rate changes. Secondly, the firms were separated by total assets, above and below one billion dollars, under the assumption that larger firms trade more frequently and are less likely to suffer a bias from non-synchronous trading. This also failed to produce positive sensitivity to the interest rate index in either the high or low asset portfolio (HIGHAST and LOWAST respectively). Since the stock price may be from a holding company which through non-bank subsidiaries diversifies its interest rate sensitivity. Moody’s Banking and Finance Manual was used to verify the percentage of holding company assets which were related to a savings and loan. A portfolio of firms with savings and loan assets representing at least 85% of their holding company assets (HCRTNS) shows no sensitivity to interest rates.39 Finally, many of the previous studies on the two index model use a sample period from the late 1970’s to early 1980’s when interest rates were exceptionally volatile and high. The sample period in this study, 1985 through 1989, is characterized by moderate interest rates and interest rate movements. Moderate interest rate behavior may be more easily managed by thrift managers. The market may not incorporate the impact of these moderate interest rate movements into stock price because of industry wide management 39 Flannery and James (1984) and Tarhan (1984) use the same methodology to show thrifts sensitive to interest rates. 87 ability or because the minimal impact Of interest rates in this moderate environment may be lost amidst stronger economic influences or the complexities of computing corporate cash flows from the cash flows Of its underlying assets. The results using an Ibbotsen portfolio of S&Ls contrasts the results from the CRSP data. The Ibbotsen quarterly portfolio started in March, 1986 and the test period concluded in December, 1990. The results of the two index model using this portfolio show a positive and significant interest rate beta for both long and short rate indices. This is consistent with theory, but contradictory to every test run with the CRSP data. Apparently, firm selection is critical in evaluating interest rate sensitivity. Although all logical firm characteristics are controlled for, our sample did not demonstrate interest rate sensitivity. CHAPTER 6 - CONCLUSION Using either failure prediction or the two index market model, previous studies have shown interest rate risk (IRR) to be positively correlated to the variability of capital level. Failure prediction tests have shown IRR positively correlated to thrift failure. The two-index market model has shown thrift stock price sensitive to changes in interest rates. Further, the sensitivity is shown proportional to maturity mismatch, which is a measure of IRR. The results for the failure prediction tests are verified here, but the results for the two index market model are not supported by this study. The failure prediction model shows which variables are positively related to thrift failure. Various CAMEL (capital, asset quality, management ability, earnings strength, and liquidity) specifications are shown to have the expected correlation to failure. Each of the IRR variables tested is shown to be positively correlated to failure. When multiple IRR variables are included, several are shown significant. Including interaction terms shows that thrift sensitivity to IRR and capital varies over firm characteristics. Various specifications of the CAMEL acronym are tested to see which specification performs best. The better performance may be a sample specific result, but may indicate a superior variable measure. The capital variable is the most critical and is specified by four different proxies. The performance of all three alternate proxies is superior to the base proxy. The base proxy, net worth / total assets, has a high standard deviation that may detrimentally affect its performance. The alternate specification, exp (net worth / total assets), is designed specifically to reduce the standard deviation and it performs as well as any of the capital variables. Asset quality is proxied best by the 88 89 percentage change in median real estate values. Unlike the other CAMEL and IRR variables, this variable is based on the regional economy. The management variable is marginally significant over the failure prediction tests. Interest income over interest expense is the best earnings variable. The best liquidity variable is the amount eligible for regulatory liquidity over total assets. This variable includes Treasury securities and may also proxy for asset safety. The inclusion of the CAMEL variables is to reduce specification error. Omitting the CAMEL variables, an IRR variable may inadvertently proxy for other firm characteristics. Including the various CAMEL variables as non-IRR measures, reduces the risk of specification error and improves the reliability of the test results. When included with the base CAMEL variables, each of the IRR variables tested is significant in predicting failure. The Achg yields the highest LRI, but Achg is an ex post proxy for IRR using the year end term structure following the failures. Uchg performs as well as the other IRRs, but does not definitively outperform any of them. In the tests including the base CAMEL variables, the extra effort and data required for the Uchg and Dchg variables does not yield a superior performance. Although SHORT includes information on thrift asset maturity, it does not distinguish between different long term maturities and different security characteristics. The performance of ISF is surprising considering that it is a liability side ratio. In spite Of these imperfections, these balance sheet ratios performed well. The performance of the MVM is probably detrimentally affected by a failure rule that depends on accounting versus market values. It may be that 90 the failure rule is actually in error. This test would support the use of any of these [RR measures for risk based capital standards. When Uchg, SHORT, and ISF were included in the regression with the base or optimal CAMEL variables, Uchg is significant, ISF is marginally significant, and SHORT is significant depending on which CAMEL variables are used. These regressions suggest that the information these IRR variables provide is additive. The implications for risk based capital is that it may be advantageous to include more than one proxy to determine thrift IRR. Further, in support of Uchg, it is consistently the most significant IRR proxy when all three are included and as the regression explains more variation its significance increases, while that of the balance sheet ratios remains constant or decreases. Interactive terms are included as a product of two variables or as the product of a variable and sequential dummies for a second variable. The product of IRR and capital level shows how the sensitivity of failure to capital changes over different IRR levels. For higher [RR levels, this variable suggests that capital is a more significant deterrent to failure. Alternatively, it may be interpreted that for higher capital levels IRR has a beneficial effect. This is an unusual result, but is supported by other interactive results. The result with Uchg and sequential capital level dummies suggests, as capital level increases, IRR decreases in detrimental effect. Then for the two highest capital levels IRR becomes an increasing beneficial factor. This may imply a cause effect relationship between capital and IRR that would provide support for risk based capital standards. Alternatively, the correlation between capital and IRR may be driven by a common third 91 variable. The only presented theory consistent with this result is that capital is a proxy for management ability. Higher capital levels imply better risk management skills. The interactive term between Uchg and the dummies for total assets shows an increasing sensitivity to IRR as size increases. This is not explained by any theory presented in this research, but there is a plausible explanation. Usually, regulators exercise greater forbearance with large firms because of the extra costs resolution will incur. This sample period is characterized by a large number of failures that were insolvent for several years. It may be that the large firms were backlogged and then failed during this sample period. In other regressions, size is shown positively correlated to thrift failure which supports this idea. A final interactive term between capital and IRR dummies shows that as IRR increases, sensitivity to capital level decreases when capital level is measured by net worth / total assets. This supports the contention of Cole et a1. (1993) that higher IRR leads to inflated book value measures of capital level and would make the stated capital level less significant in preventing failure. There are several implications of these results for risk based capital regulation. First, the significance of multiple IRR proxies in measuring IRR suggests capital standards may depend on a fiinction of several IRR proxies. Second, variable sensitivity to IRR over asset size may be a sample period bias, but if not, larger firms should have higher capital standards for the same amount of IRR as smaller firms. Finally, the unusual beneficial aspect of IR for high capital firms conversely suggests that low IRR is detrimental for 92 high capital firms. This implies that risk based capital should require more capital for firms with high [RR and more IRR for firms with high capital. These results suggest additional research in the area of optimal firm risk levels. For highly capitalized firms, low IRR is positively associated with failure. This may indicate the firm risk level is too low to make adequate profits. Including all types of thrift risk: capital risk, IRR, asset risk, and others, in a single variable may provide information on whether risk is definitively bad or if some amount of risk is Optimal. Over time DIs have failed at very different rates, suggesting macroeconomic factors influencing failure rates. These macroeconomic factors may influence the optimal risk level. The relationship between macroeconomic factors and firm specific factors needs to be determined. To gain a more detailed understanding of this relationship, a continuous dependent variable such as net income may be more effective than the binary variable, failure. The two-index model uses a market and an interest rate index to show a correlation between DI stock return and interest rate changes. Past studies with this methodology have shown support for this relationship. In this study, using thrift stock returns and various interest rate indices, the relationship between stock returns and interest rate changes is not confirmed. This is contradictory to previous studies and to the results of the failure prediction test in this study. The most probable explanation for this unusual result is the lack of volatility in interest rates during the sample period. Also, the interest rates are at moderate levels which reduces their impact on firm value. 93 The intent of this study is to evaluate [RR proxies. This study shows all the IRR proxies tested to be good measures of actual risk. The MVM, a more technical and theoretically superior measure, showed moderate improvement over simple balance sheet ratios. Finally, the sensitivity of failure to [RR is shown to vary over other firm characteristics, such as capital level and asset size. APPENDIX 94 Appendix: (Equations are from the OTS MVM except where noted.) Prepayment rate- Prepayment Equation: cprm = seasoning} seasonality."‘refinancing“J ( A, 1) n - path index. t - month index Seasoning depends on the age or time since issue of the mortgage. It assumes new mortgages have a slower prepayment rate. seasoning = t / 30 for t < 30 months (A2) 1 fort 2 30 months Seasonality depends on the month of the year (month = 12) and the number of months into the simulation. seasonality = 1+.2000*Sin{ 1.571 *[(month+t-3)/3]-1} (A.3) Refinancing depends on the coupon of the mortgage and the simulated mortgage rate. The refinancing equation is security specific and listed following the appropriate security. The annual prepayment rate is converted to monthly form. 112 pm = 1-( I-CPrM) (A4) p... - monthly prepayment rate; cprm- constant prepayment rate- annualized Fixed Rate Mortgages (F RM) — (Refinancing equations are from the OTS NPVM) Representative Instrument: FNMA 30 yr. FRM Conventional, seasoned (over 60 months) and moderately seasoned (30 to 60 months) Data Source: Bloomberg System 95 Refinancing Equations: well seasoned (over 10 years old) refinancing = .l828-.0892*arctan[4.776*(1.083-coupon/mort)] moderately seasoned (less than 10 years old) refinancing = . 1992-. 1295 *arctan[3 .623 *( l .087-coupon/mort)] mort- mortgage rate = short rate at t+60 months Adjustafle Rate Mortgages (ARM)- Representative Instrument: 1 year reset - 1 year Treasury FNMA Data Source: Wall Street Journal Refinancing Equation: refinancing = .2006-.0950*arctan[2.401*(1.021*coupon/mort)] Calculation of fully indexed rate: fully indexed ratem = margin + indexj,..,,-2 j - adjustable rate index Projected index: index”,t = forecasted ratem + basis,- basisj = (mortgage rate (5 year) - short rate (1 month))*M,~ Mj - .5, .7, and 1 for 1, 3, and 5 year indexes respectively Calculation of new coupon: (A5) (A6) (A7) (A8) (A9) (A. 10) cm = max{ [min(fully indexed rate, cm-1+ period cap, life cap)], on,” - period floor, life floor} (All) 96 Assume the first resetting of the coupon occurs at half the reset period, ex: 1 yr. Treasury resets after the 6th month. Index and reset are equal, ex: 5 yr. index resets every 5 years. Balloon Payment Mortgages Representative Instrument: FNMA 30 year FRM Conventional; seasoned and moderately seasoned Data Source: Bloomberg System Refinancing Equation: same as F RM Assume balloon mortgages amortize according to a 30 year schedule. Remaining balance is paid at actual maturity (7.5, 15, or 25 yrs). Borrowings Cash Flow (CF) Equation: B + B CF. =—rL—2—fl+ (4.12) B, - balance at time t: r - reported cost of funds B, = BM except Baum-(y = 0 Retail and Wholesale Certificates of Deposit Next period’s balance: 3. =8.-i(1+r.) (A.13) rt- reported cost of funds 97 Non-interest cash flows: , B. + B,_ ‘ (F; Z—IILf—ll (A.14) n - noninterest costs Early withdrawal condition: CD — NOM ‘3'" r*m((l+( 12 j) -1 (A.15) r - one month interest rate; CD - current CD rate; NOM - contracted CD rate; and m - years remaining on the CD contract. Core Deposits Cash Flow Equation: CF. =— GO 2820:? "CO 37639 6.80 v.26; mzmoaoa no 23223 Ba 5:25 Bi 3.6; Q0 6.36.055 ”.00 $o§>n< mdfi v.26; DU 282055 ”50 mweaotom EEO .32.:33Q 33::on E335. coasooé FUD E2532: 22% + mcmoq bane; 5950-0 ”CD 3868800 365m + 33.3: cog—50-0 "50 250‘— 583280 @8ch + mowmmco—z bane; 8950-0 moo vacuum nmoam + .6336; com—50-0 moo mowmwtoZ EEO mowmwtovz $5 + 2:8 SN 95 emu 80:8 mowmwto—z $5 + 2:3 08 $5 23. 0323.3 mommwto—z 35 + 2:8 com 35 22 BE .énaww 8mm E3085 @0562 E2085 a3. 5::on :25 “2308502 352 - 4 2,5 113 Table 5- Price sensitivity table on December 31,1985 for fixed rate Federal National Mortgage Association (FNMA) mortgages. Current term structure Maturity 25 15 7.5 4 0.07 0.87798 0.90280 0.93929 0.96688 0.08 0.91587 0.92922 0.95537 0.97616 0.09 0.96157 0.96690 0.98123 0.99254 Term structure shock - up 200 bp Maturity 25 15 7.5 4 0.07 0.77600 0.82639 0.88661 0.93344 0.08 0.81661 0.85213 0.90241 0.94270 0.09 0.87225 0.89339 0.92992 0.95978 Term structure shock - down 200 bp Maturity 25 15 7.5 4 0.07 0.97572 0.98065 0.99218 1.00014 0.08 0.99687 0.99868 1.00446 1.00791 0.09 1.02569 1.02673 1.02565 1.02239 Coupon 0.1 1.00459 1.00608 1.00995 1.01163 0.11 1.03259 1.03345 1.03089 1.02603 Coupon 0.1 0.93015 0.94028 0.96234 0.98039 0.11 0.97152 0.97562 0.98725 0.99643 Coupon 0.1 1.05571 1.05806 1.05038 1.03984 0.11 1.07627 1.08087 1.06893 1.05326 Term structure shock - actual term structure change Coupon Maturity 25 15 7.5 4 0.07 0.97656 0.98081 0.99080 0.99777 0.08 0.99681 0.99833 1.00296 1.00558 0.09 1.02517 1.02611 1.02412 1.02013 0.1 1.05494 1.05728 1.04883 1.03762 0.11 1.07531 1.07994 1.06733 1.05103 0.12 1.04965 1.05131 1.04514 1.03616 0.12 0.99809 0.99957 1.00464 1.00788 0.12 1.08947 1.09641 1.08197 1.06292 0.12 1.08834 1.09533 1.08028 1.06066 0.13 1.05279 1.05532 1.04864 1.03888 0.13 0.14 1.06128 1.06512 1.05693 1.04508 0.14 1.00668 1.019364 1.00737 1.019984 1010151019975 1.01144 0.13 1.09099 1.09925 1.08488 1.06541 0.13 1.08968 1.09802 1.08311 1.06311 1.01828 0.14 1.09819 1.10828 1.09283 1.07150 0.14 1.09675 1.10695 1.09099 1.06917 Key: Prices are differentiated by term structure, coupon, and maturity. 114 Table 6 - Failure Record for Thrifts in months 1 through 12. Year Thrifts at the Receivership Conservatorshipl Total Failures beginning of (from months 0 (from months 0 (from months 0 the year to 12) to 12) to 12) 1986 3135 45 9 54 1987 3164 40 23 622 1988 3085 179 6 185 1989 2864 282 O 282 1990 2648 187 3 190 Total 14896 733 41 773 Key: The failure data is compiled from lists provided by the Office of Thrift Supervision and the Federal Deposit Insurance Corporation. Receivership implies a regulatory change of ownership while conservatorship implies a regulatory change of management. Total failures is the sum of both receivership and conservatorship. The number of thrifis is from the Thrift Financial Report of the previous December. 1 This includes management consignment programs and other conservator-like programs. 2 One firm enters conservatorship and receivership in the same year. 115 Table 7 - Failure Record for Thrifis in months 13 through 24. Year Thrifts at the Receivership Conservatorshipl Total Failures beginning of (from months (from months 13 (from months the year 13 to 24) to 24) 13 to 24) 1986 3135 40 23 622 1987 3164 197 6 203 1988 3085 291 o 291 1989 2864 191 2 193 1990 2648 134 8 142 Total 14896 853 39 391 Key: The failure data is compiled from lists provided by the Office of Thrifi Supervision and the Federal Deposit Insurance Corporation. Receivership implies a regulatory change of ownership while conservatorship implies a regulatory change of management. Total failures is the sum of both receivership and conservatorship. The number of thrifts is from the Thrift Financial Report of the previous December. 1 This includes management consignment programs and other conservator-like programs. 2 One firm enters conservatorship and receivership in the same year. Table 8 - Summary Statistics: Total Assets- (‘1.000) Capital- (net worth/til asts) Capital- (MVPE/market) Capital- (MVPE/ttl asts) Capital- exp(net worth/ttl asts) Asset Quality- (repos'd asts/111 asts) Asset Quality- (other real estate/ttl asts) Management Efficiency- (net inc/gr inc) Earnings Ability- (net inc/ttl asts) Earnings Ability- (int inc/int exp) Liquidity- (amt elig for reg liq/ttl asts) Liquidity- (brokered deposits/m asts) Liquidity- (cash and securities/it! asts) Uchg- (MVPE-MVPEUVMarket Asts ISF- (interest sensitive funds/til funds) SHORT- (s.t. Iiab‘s - s.t. asts)/til asts Interactive Variables Uchg*exp(capita|) ISF*exp(capital) SHORT*exp(capital) 116 Median Mean Standard Deviation $99,846 $400,151 $1,436,107 0.0520 0.0443 0.0807 0.0603 0.0406 0.1070 0.0596 0.0439 0.0930 1 .0534 1.0484 0.0735 0.0034 0.0159 0.0404 0.0000 0.0034 0.0163 0.0203 0.0228 0.0124 0.0012 -0.0015 0.0159 1 .1200 12.0307 434.8286 0.1005 0.1303 0.0972 0.0000 0.0138 0.0511 0.0371 0.0562 0.0625 0.0296 0.0301 0.0123 0.9278 0.9046 0.0851 0.1175 0.1282 0.2519 0.0309 0.0314 0.0122 0.9653 0.9433 0.1314 0.1218 0.1232 0.2014 Key: The accounting and maturity data are from, December 31, Thrifi Financial Reports (TFRs), 1985-1989. There are 14.896 observations with data for each variable. 117 Table 9 - Logit regression of year one failures onto base CAMEL and IRR variables. Constant- Capital- (net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/til asts) Earnings Ability- (net inc/til asts) Liquidity- (amt elig for reg qu/ttl asts) Uchg- (MVPE-MVPEUV Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 -2923 -3.357 3597 -2.826 4 .880 -2294 (48.875) (4.730) (-7.184) (-7317) (-3.408) (-7.183) -6.774 -24795 -7.088 -8.216 49.848 1.421 (42.137) (-7.406) (-5339) (7304) (41.848) (1.218) 4.007 1.215 2.706 4.830 9.773 2.798 (4.608) (0.288) (1.124) (3.211) (3.280) (1.437) 3.282 23.480 3.988 43.898 3.122 9.818 (1.122) (2.617) (0.733) (2040) (0.284) (1.783) -22.681 28.270 43.784 43.379 -30.255 -51.129 (-8.811) (2.908) (-3.186) (2943) (.2399) (8.589) -8.232 2995 -2.022 -2.344 -5.594 44.958 (-9.802) (4 .233) (4 .073) (4 .799) (-2.106) (8.930) 17.915 45.253 45.879 9.873 17.944 17.920 (5.000) (-0.840) (4 .023) (1.301) (1.880) (2.471) 4431 394 478 1020 81 1 1 148 0.273 0.303 0.243 0.281 0.878 0.172 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 118 Table 10 - Logit regression of year two failures onto base CAMEL and IRR variables. Constant- Capital- (net worth/m asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/til asts) Liquidity- (amt elig for reg liq/m asts) Uchg- (MVPE-MVPE..)/ Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 -2395 4 .302 -1.738 4 .247 -2.082 -2.594 (-14.874) (-2.110) (4984) (8.180) (8.351) (-7.782) 4731 47.123 40.933 -16.509 -30.749 4.757 (-7255) {-4.841) (8.805) (-7.792) (-8.448) (4.459) 10.551 4.845 7.387 24.343 20.438 3.218 (10.318) (1.220) (3.428) (8.325) (8.782) (1.531) 1 1.289 10.500 .4882 -7.348 24.990 14.284 (3.751) (1.050) (-0.671) (-0.880) (3.782) (2.338) 43.020 18.078 42.853 43.471 8.507 18.885 (-3.832) (1.889) (4.948) (4.844) (0.533) (1.803) -8.297 42.218 8.989 -8.308 4304 8.820 (-11.617) (8.891) (4.990) (8.299) (-2723) (8.881) 8.841 40.338 0.587 -0.102 18.154 3.819 (2.011) (-2593) (0.087) (-0.015) (2.539) (0.480) 5478 490 1 102 1205 1 102 1058 0.188 0.185 0.245 0.388 0.207 0.038 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 119 Table 11 - Logit regression of year one failures onto base CAMEL and Dchg variables. Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 Constant- -2.685 -3.604 -3.917 -2.967 -1.584 -2.048 (-17.874) (-5.334) (-8.595) (-9.396) (-3.401) (-7. 149) Capital- 6784 -24.874 -7.062 -8.199 -50.032 1.466 ("9‘ mm" 35“) (-12.170) (-7.449) (-5.290) (-7.312) (-11.740) (1.272) Asset 3.983 1.024 1.954 4.1 14 9.358 3.233 Quality- (4.547) (0.224) (0.818) (2.709) (3.132) (1.675) (repos'd asts/ttl asts) Management 2.627 24.939 4.733 -13.451 1.912 9.682 Efficiency- (0.907) (2.860) (0.866) (-2.040) (0.172) (1.732) (non-int exp/ttl asts) Earnings -22.355 26.081 -13.588 -13.484 -29.133 -50.803 Ability- (-8.690) (2.894) (-3.137) (2.903) (-2.314) (6556) (net inc/tn asts) Liquidity- -8.299 -3.052 -2.017 -2.289 -5.639 -14.811 fa” °“9 '°' "’9 (-9.896) (-1.246) (-1.063) (-1.755) (-2.129) (6879) WI" asts) Dchg- 13.199 -8.747 -2.959 16.312 13.694 10.810 affix?” (4.172) (-0.467) (-0.227) (2.733) (1.696) (1.680) AIC 4438 395 477 1015 61 1 1150 LRI 0.272 0.302 0.241 0.285 0.676 0.169 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) 120 Table 12 - Logit regression of year two failures onto base CAMEL and Dchg variables. Constant Capital- (net worth/m asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg liq/m asts) Dchg- (MVPEg-MVPEV Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 -2.618 -2.345 -1 .991 4 .888 -2.092 -2.751 (-18.321) (8.783) (-6.304) (4.934) (—6.651) (-9.169) —4.699 -16.763 -10.869 -16.284 80.708 -1.768 (-7.196) (4.842) (8.598) (-7712) (-8.399) (4 .478) 9.854 3.259 8.873 23.449 19.704 2.829 (9.586) (0.855) (3.105) (7.966) (8.498) (1.264) 12.335 17.220 8.885 -5.120 24.639 14.458 (4.098) (1.807) (-0.528) (8.823) (3.759) (2.371) 42.835 14.881 42.573 44.753 7.800 18.757 (-3.709) (1.602) (4 .904) (4 .727) (0.638) (1.825) -8.229 42.752 -6.882 -8.178 4.597 -5.730 (41.517) (4.004) (4.921) (8.228) (-2.869) (8.815) 15.410 -8.458 10.487 12.090 23.980 11.823 (5.460) (-0.518) (1.627) (2.240) (4.208) (1.883) 5455 498 1099 1200 1092 1055 0.172 0.174 0.248 0.371 0.215 0.040 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 121 Table 13 - Logit regression of year one failures onto base CAMEL and Achg variables. Constant- Capital- (net worth/m asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt eIig Ior reg qu/ttl asts) Achg- (MVPE-MVPE.)/ Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1989 -2.482 8.880 8.822 4 .878 4 .530 (20.327) (-6.422) (-9.781) (8.437) (8.710) 8.894 -24.776 -7.161 -7.890 1.744 (42.177) (-7.447) (-5.353) (8.990) (1.540) 5.107 0.486 0.868 5.989 3.284 (5.978) (0.107) (0.351) (3.978) (1.758) 1.488 28.011 3.734 4 4.258 8.105 (0.522) (2.993) (0.667) (-2143 (1.431) -22.466 25.790 43.753 42.804 80.953 (8.773) (2.880) (8.189) (-2.714) (8.538) 8.040 -3.007 4 .898 -2.291 45.015 (-9.641) (4.228) (4.005) (4.795) (8.980) 2.439 8.833 42.803 -23.082 8.612 (8.544) (8.038) (8.813) (-2.299) (2.992) 4418 395 477 1018 1144 0.276 0.302 0.242 0.284 0.173 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) N.C. - no convergence. 122 Table 14 - Logit regression of year two failures onto base CAMEL and Achg variables. Constant- Capital- (net worth/til asts) Asset Quality- (repos‘d asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg qu/ttl asts) Achg- (MVPE-MVPE.)/ Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 2.234 -2.469 -1 .821 8.840 -0.540 -2093 (—19.242) (4.378) (-5.538) (4 .813) (-1.624) (-7521) 8.011 48.877 41.071 4 5.728 81.401 4 .311 (-7.538) (-4.830) (8.847) (-7.378) (-8.533) (4 .091) 10.951 3.029 6.810 24.313 18.733 2.888 (10.777) (0.802) (3.083) (8.266) (8.029) (1.341) 9.905 17.745 -5.688 -7.699 23.900 11.501 (3.288) (1.868) (8.817) (8.955) (3.839) (1.823) 43.102 14.700 42.952 48.351 10.827 18.374 (8.851) (1.587) (4 .959) (-1.881) (0.881) (1.555) -8.163 42.788 8.888 -7.830 8.998 -6.302 (41.471) (4.005) (4.925) (8.055) (-2.662) (8.987) 1.409 4.728 -7.105 -25.348 4.032 10.986 (4.027) (0.314) (8.785) (-2.768) (8.785) (3.711) 5487 496 1 101 1 198 1092 1044 0.170 0.174 0.245 0.373 0.214 0.051 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 123 Table 15 - Logit regression of year one failures onto base CAMEL and ISF variables. Constant Capital- (net worth/til asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/til asts) Liquidity- (amt elig for reg qu/ltl asts) ISF- (interest sensitive funds/ttl funds) AIC LRI Data Period for Independent Variables 1985-89 1985 1986 1987 1989 8.473 8.470 8.424 8.180 8.897 (8.918) (2.277) (8.273) (4.722) (8.811) 8.71 1 24.793 -7.046 -8.042 1.398 (-12.016) (-7449) (8.221) (-7.178) (1.211) 4.811 0.518 1.573 4.804 4.144 (5.888) (0.116) (0.706) (3.282) (2.202) 1.773 25.945 4.874 4 5.121 9.427 (0.818) (3.070) (0.893) (2.224) (1.885) 22.319 25.645 43.142 43.353 81.184 (-8.721) (2.841) (8.045) (2.885) (-6.620) 8.827 -2.889 2.379 8.387 44.209 (4 0.249) (4 .185) (4.211) (2.490) (8.518) 1.328 8.458 1 .663 4.200 4 .298 (2.377) (8.288) (0.893) (2.935) (4.357) 4449 395 476 1012 1 150 0.271 0.302 0.242 0.287 0.169 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) N.C.- no convergeance 124 Table 16 - Logit regression of year two failures onto base CAMEL and ISF variables. Constant Capital- (net worthlttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg liq/til asts) ISF- (interest sensitive funds/til funds) AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 4 .219 -0.648 2.209 8.238 1.178 8.775 (8.280) (-0.579) (2.470) (8.285) (1.458) (-0.838) -4.817 48.890 40.908 -16.687 81.799 4 .908 (-7.360) (-4.874) (8.808) (-7.877) (8.592) (4 .585) 10.929 2.786 7.320 24.894 21.172 3.584 (10.800) (0.740) (3.484) (8.417) (6.977) (1.773) 10.108 17.969 4.588 -7.333 20.175 14.180 (3.405) (1.932) (-0.660) (8.899) (3.098) (2.319) 43.281 13.937 42.878 42.831 5.219 18.971 (8.928) (1.502) (4 .921) (4 .538) (0.425) (1.806) -7.995 4 1 .882 -7150 -7.920 8.282 -5.234 (41.111) (-3.647) (4.975) (4.997) (-2.106) (8.334) 4 .096 2.322 0.568 4 .178 2.904 4 .954 (-2.655) (4 .841) (0.571) (4 .295) (8.288) (4 .925) 5478 493 1101 1203 1098 1055 0.168 0.179 0.245 0.369 0.210 0.041 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (I- statistics are below parameter estimates.) 125 Table 17 - Logit regression of year one failures onto base CAMEL and SHORT variables. Constant- Capital- (net worth/ttl asts) Asset Quality- (repos’d astle asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg qu/ttl asts) SHORT- (s.t. liab's - s.t. asts)/ttl asts AIC LRI Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 -2.556 8.828 -3.793 2.721 4.518 2.219 (49.788) (8.475) (40.899) (40.132) (8.829) (8.253) 8.728 -26.090 8.230 -7.178 -48.173 3.200 (-9.699) (-7455) (8.297) (8.074) (41.370) (2.608) 4.355 1.440 2.584 4.520 8.948 2.828 (5.100) (0.324) (1.118) (3.031) (3.007) (1.481) 2.190 25.789 4.883 43.949 8.975 10.072 (0.769) (3.088) (0.918) (2.125) (8.091) (1.829) 22.724 27.924 43.119 -13.810 80.143 -49.850 (-8.916) (3.068) (8.020) (8.094) (2.378) (8.484) -8.194 8.182 -2.263 -2.288 8.197 44.247 (-9.839) (4.293) (4.195) (4.784) (4.972) (8.888) 1.198 4.222 4.381 1.358 2.083 1.984 (4.871) (4 .327) (4 .838) (2.666) (2.943) (4.294) 4431 393 475 1015 805 1134 0.273 0.305 0.245 0.285 0.679 0.181 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 126 Table 18 - Logit regression of year two failures onto base CAMEL and SHORT variables. Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 Constant -2.254 -2.532 -1.755 -1.548 -1.710 -2.476 (-18.644) (-5.992) (-6.260) (-5.208) (-6.239) (-9.343) Capital- -4.318 -17.004 -10.731 -15.313 -29.408 -1.820 ("etwmhm' 385) (8.370) (4.733) (8.128) (-7.126) (-8.091) (4 .447) Asset 10.587 3.040 7.306 23.973 20.114 3.583 Quality- (10.404) (0.812) (3.447) (8.196) (6.515) (1.753) (repos'd astle asts) Management 10.760 18.346 -4.596 -5.982 23.158 14.199 Efficiency- (3.631) (1.993) (-0.664) (-0.742) (3.489) (2.320) (non-int exp/ttl asts) Earnings -13.133 15.144 -13.039 -12.932 8.045 18.591 Ability- (8.908) (1.611) (4 .978) (4 .591) (0.662) (1.789) (net inc/ttl asts) Liquidity- -8.310 -12.791 -6.951 -8.073 -4.133 -5.842 (amt 6'19 W "39 (4 1.670) (4.013) (4.978) (-5.170) (2.857) (8.892) Irqlttl asts) SHORT- 0.455 -0.317 0.180 1.373 1.929 -0.131 (81- "ab's ' s-l- (2.088) (-0.383) (0.350) (2.912) (3.963) (-0.269) asts)/til asts AIC 5478 496 1 102 1 197 1092 1058 LRI 0.168 0.174 0.245 0.373 0.214 0.038 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) 127 Table 19 - Logit regression of year one failures onto base CAMEL and IRR variables except MVPE / market is used for capital. Constant- Capital- (MVPE/market) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int explttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg liq/n1 asts) Uchg- (MVPE-MVPEJ/ Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1986 1987 1988 1989 8.390 8.974 8.940 8.538 8.315 2.189 (-21.084) (-6.055) (8.042) (40.195) (-7.462) (-7254) 4.338 -5.800 8.083 -7.035 21.488 2.345 (44.253) (4.382) (4.587) (-7.971) (42.093) (4.889) 7.159 14.193 7.187 7.991 19.718 4.538 (9.758) (4.516) (3.815) (8.072) (7.547) (8.902) 10.713 25.529 9.150 4 .203 21.287 6.677 (3.997) (2.947) (1.737) (8.209) (2.880) (1.283) -26.991 -9.521 47.282 47.288 .71.932 88.978 (40.833) (4.394) (4.385) (8.998) (-7.459) (8.478) 8.245 4.199 4.435 4.575 8.388 45.578 (40.220) (4.790) (8.792) (4.225) (2.377) (-7.335) 20.568 -10.983 42.943 18.280 14.179 20.189 (8.012) (8.852) (8.844) (2.898) (1.888) (2.832) 4401 449 489 1002 714 1 126 0.278 0.203 0.221 0.294 0.620 0.187 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 128 Table 20 - Logit regression of year two failures onto base CAMEL and [RR variables except MVPE / market is used for capital. Constant Capital- (MVPE/market) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inclttl asts) Liquidity- (amt elig for reg qu/ttl asts) Uchg- (MVPE-MVPEJI Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 2.842 4 .881 2.084 2.043 2.200 -2.677 (47.817) (-3.209) (-6.091) (8.710) (8.025) (—8.476) 2.727 2.411 41.312 4 0.472 4 9.087 2.199 (—8.688) (4 .931) (8.189) (8.005) (4 0.785) (4.178) 11.829 10.408 9.511 29.128 24.143 1.907 (12.590) (3.245) (4.954) (10.348) (7.843) (0.978) 14.880 15.084 4 .516 6.285 25.851 15.769 (5.378) (1.823) (-0.197) (0.943) (3.777) (2.731) 47.802 43.395 47.817 21.921 8.049 20.722 (-5.630) (4 .823) (2.799) (2.838) (-0.718) (2.137) 8.548 43.828 8.818 -7.855 4.218 -7.110 (42.103) (4.394) (4.188) (4.931) (2.721) (4.597) 8.404 -36.323 5.388 8.017 4.789 4.490 (2.831) (2.388) (0.631) (0.918) (0.859) (0.546) 5485 51 1 1050 1198 1057 1044 0.170 0.147 0.280 0.373 0.240 0.051 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 129 Table 21 - Logit regression of year one failures onto base CAMEL and IRR variables except MVPE / total assets is used for capital. Constant- Capital- (MVPE/ttl asts) Asset Quality- (repos'd astsfttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg liq/til asts) Uchg- (MVPE-MVP Eu)/ Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 8.223 -3.939 8.835 8.097 -3.223 2.097 (49.795) (-5.999) (-7.776) (8.218) (7.193) (-6.972) 8.285 8.131 -5.568 8.422 23.897 -3.621 (4 5.858) (8.777) (4.230) (8.485) (-12.087) (8.104) 8.333 15.072 6.785 7.040 19.700 2.492 (8.428) (5.204) (3.354) (5.201) (7.441) (4 .409) 7.814 25.189 7.821 -6.894 21.102 5.093 (2.819) (2.896) (1.438) (4 .088) (2.591) (0.979) 25.378 -5.148 48.942 -15.516 -70.404 88.899 (40.258) (-0.724) (4.281) (8.583) (-7.172) (8.185) 8.254 8.552 4 .553 4 .606 8.204 4 8.183 (40.078) (-1.526) (-0.853) (4.233) (2.281) (-7.542) 19.599 42.001 4 3.428 13.223 12.897 20.221 (5.889) (8.711) (8.878) (2.081) (1.509) (2.847) 4345 438 490 981 701 1 123 0.288 0.223 0.219 0.309 0.627 0.189 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) ‘1 130 Table 22 - Logit regression of year two failures onto base CAMEL and [RR variables except MVPE / total assets is used for capital. Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 Constant -2.555 -1.848 -1 .987 -1.860 -2.135 -2.593 (4 7.012) (-3. 161 ) (8.808) (-5.280) (8.850) (8.155) Capital- -4.038 -3. 592 -1 2. 589 -12.999 -20.535 -3. 565 (MVPE/“l ”‘9 (-9.945) (-2.059) (-9.758) (-9.431) (-11.069) (4508) Asset 1 1.282 10.318 9.200 28.954 24.533 0.946 Quality- (11.925) (3.205) (4.784) (10.176) (7.873) (0.482) (repos'd asts/ttl asts) Management 14.163 15.185 ~3.332 5.231 26.214 14.064 Efficiency- (4.984) (1.632) (-0.432) (0.784) (3.865) (2.363) (non-int exp/ttl asts) Earnings -15.880 -12.066 -15.881 -19.166 -6.509 25.541 Ability- (-5.086) (-1.583) (2.552) (-2.313) (-0.582) (2.481) (net inc/ttl asts) Liquidity- -8.515 -13.569 -5.787 -7.653 -4.116 -7.577 ($419 for res W" (4 1.982) (4.297) (4.145) (4.741) (2.858) (-4.758) Uchg- 8.073 -36.554 5.570 4.482 3.800 4.747 ”1:2?st (2.518) (-2.406) (0.651 ) (0.691) (0.523) (0.579) AIC 5439 511 1042 1171 1050 1041 LRI 0.174 0.148 0.286 0.386 0.245 0.054 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 131 Table 23 - Logit regression of year one failures onto base CAMEL and IRR variables except exp (net worth / total assets) is used for capital. Constant Capital- exp(net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg lid/til asts) Uchg- (MVPE-MVPEUV Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 7.611 25.006 7.387 9.913 50.278 4 .053 (10.435) (6.688) (4.157) (6.612) (10.919) (8.783) 4 0.312 -28.215 4 0.819 -12.458 -52.006 4 .047 (45.850) (-7.826) (-6.671) (8.341) (41.879) (8.838) 2.145 8.140 0.766 2.837 9.285 0.575 (2.440) (8.030) (0.303) (1.871) (3.081) (0.294) 8.313 22.108 0.875 4 9.089 1.858 6.792 (8.108) (2.420) (0.188) (2.810) (0.165) (1.243) -18.646 27.158 40.909 8.701 27.558 45.815 (-7334) (2.998) (2.539) (4 .985) (2.182) (-6.052) -8.105 2.887 4 .583 2.323 8.380 4 4.721 (8.574) (4 .108) (-0.841) (4 .734) (2.008) (-6.806) 17.925 -16.985 48.185 10.131 18.909 18.473 (4.948) (8.928) (4 .028) (1.313) (1.747) (2.559) 4323 388 460 981 595 1 147 0.291 0.319 0.269 0.309 0.685 0.171 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the colurrm headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 132 Table 24 - Logit regression of year two failures onto base CAMEL and IRR variables except exp (net worth / total assets) is used for capital. Constant Capital- exp(net worthntl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg qu/ttl asts) Uchg- (MVPE-MVPEQ/ Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 5.385 18.295 12.929 19.528 29.997 1.391 (6.561) (4.854) (8.353) (7.888) (7.948) (0.885) -7.597 -19.480 44.451 20.493 -31.994 -3.827 (40.107) (8.233) (-7.698) (8.935) (-8.878) (2.891) 9.143 3.809 5.853 22.988 20.149 2.008 (8.831) (0.968) (2.687) (7.740) (8.822) (0.943) 9.389 9.487 -7. 166 4 0.298 24.870 11.589 (2.959) (0.938) (4 .085) (4 .192) (3.715) (1.838) 8.505 18.532 8.520 -7.946 8.584 24.835 (2.511) (1.797) (4 .420) (-0.971) (0.699) (2.245) -8.118 41.902 -6.784 8.108 4.050 8.545 (41.315) (8.798) (4.813) (8.122) (2.589) (8.495) 8.598 41.497 8.411 -0.804 17.926 4.182 (1.988) (2.857) (8.048) (8.120) (2.495) (0.503) 5421 485 1080 1 177 1089 1053 0.177 0.193 0.280 0.383 0.218 0.043 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 133 Table 25 - Logit regression of year one failures onto base CAMEL and [RR variables except other real estate / total assets is used for asset quality. Constant Capnah (net worth/ttl asts) Asset Quality- (other real estate/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig tor reg qu/ttl asts) Uchg- (MVPE-MVPEQ/ Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 2.881 8.344 8.853 2.774 4 .521 2.270 (-16.662) (4.711) (-7.198) (-7.083) (2.899) (-7097) 8.083 25.513 -7.996 8.850 83.733 0.545 (48.403) (8.485) (-6.613) (8.893) (42.797) (0.580) 0.747 2.855 2.327 6.298 21.373 4.752 (0.449) (8.510) (8.494) (2.038) (3.158) (0.924) 1.883 23.730 3.737 4 5.094 8.055 9.167 (0.852) (2.853) (0.884) (2.148) (8.281) (1.654) 24.872 28.782 44.059 44.523 83.717 -51.943 (8.807) (2.978) (8.194) (8.181) (2.551) (8.899) -8.417 8.172 2.174 2.587 4.945 44.703 (40.003) (4 .297) (4 .148) (4 .981) (4.919) (8.788) 21.271 4 3.514 8.195 14.049 17.899 19.511 (6.180) (8.758) (8.559) (1.926) (1.873) (2.787) 4452 394 477 1028 813 1 148 0.270 0.303 0.241 0.277 0.675 0.171 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 134 Table 26 - Logit regression of year two failures onto base CAMEL and IRR variables except other real estate / total assets is used for asset quality. Constant Capital- (net worth/ttl asts) Asset Quality- (other real state/m asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inclttl asts) Liquidity- (amt elig tor reg qu/ttl asts) Uchg- (MVPE-MVPEuy Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 2.221 4 .280 4.533 8.750 4 .891 2.587 (43.888) (2.039) (4.321) (-2.016) (4.372) -7.694 -7471 -18.198 42.882 23.029 88.017 2.749 (-12.308) (-5.575) (-7959) (40.588) (8.742) (2.881) 9.473 5.445 10.114 28.241 22.549 12.751 (5.484) (1.513) (3.023) (5.803) (4.187) (2.333) 9.082 10.410 40.483 8.015 25.087 13.551 (2.919) (1.039) (4 .385) (4 .035) (3.524) (2.179) 4 7.769 17.398 4 8.011 25.558 0.831 20.737 (8.089) (1.873) (2.889) (2.895) (0.047) (2.055) 8.343 42.217 -7.161 -8.692 8.805 8.530 (41.717) (-3.893) (-5.110) (8.482) (2.475) (8.508) 11.041 -39.736 4.753 4.128 19.583 5.469 (3.478) (2.575) (0.587) (0.855) (2.845) (0.885) 5557 489 1 105 1254 1 138 1056 0.158 0.188 0.243 0.342 0.183 0.040 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 135 Table 27 - Logit regression of year one failures onto base CAMEL and IRR variables except percentage change of median real estate is used for asset quality. Constant Capital- (net worth/ttl asts) Asset Quality- (percentage change in real estate value) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttl asts) Liquidity- (amt elig for reg qu/ttl asts) Uchg- (MVPE-MVPEQI Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 -2.627 8.222 8.331 2.308 4 .402 -2.198 (45.045) (4.487) (8.415) (-5.840) (-2.628) (-6.560) -7.783 -24.635 -7.588 8.380 82.878 0.571 (-16.048) (8.300) (8.481) (8.450) (42.542) (0.585) 4.931 8.587 4.182 8.387 2.492 8.981 (-6.414) (4 .128) (4 .858) (4.420) (4 .530) (-0.653) 3.327 23.725 3.459 4 2.483 1.813 9.274 (1.133) (2.827) (0.635) (4 .839) (0.141) (1.874) 22.719 25.541 4 3.311 4 2.890 -35.815 82.420 (8.898) (2.839) (2.987) (-2.847) (2.805) (-6.787) 8.188 2.892 4 .880 2.802 8.794 4 4.950 (-9.776) (4 .208) (4.011) (4 .992) (2.221) (8.893) 18.988 43.991 -14.738 10.575 17.184 19.798 (5.450) (8.784) (4 .006) (1.443) (1.822) (2.804) 4407 393 474 1008 818 1148 0.277 0.306 0.247 0.290 0.672 0.171 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) 136 Table 28 - Logit regression of year two failures onto base CAMEL and IRR variables except percentage change of median real estate is used for asset quality. Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 Constant -1.817 -1.054 -1.161 -0.001 -1.380 -2.480 (-11.156) (4 .683) (8.348) (-0.004) (8.525) (-7.018) Capital- -7.337 -1 7. 824 -12.834 -21.780 -34.930 -2.669 ("elm'll‘m' “'5’ (-12.332) (-5.391) (-8.227) (-1 0.255) (-9.475) (-2.633) Asset -7.867 -6.258 -7.815 -12.826 -6.504 -1.338 Quality- (4 0.827) (4 .996) (8.541) (-7.338) (4.881) (8.813) (percentage change in real estate value) Management 1 1 .609 10.846 -5.970 -3.252 30.935 13.912 Efficiency- (3.675) (1.064) (8.858) (-0.412) (4.358) (2.257) (non-int exp/ttl asts) Earnings -15.256 15.208 -12.782 -17.149 3.237 16.912 Ability- (4.502) (1.589) (4.981) (4 .918) (0.243) (1.857) (net inc/ttl asts) Liquidity- -8.415 -12.023 -6.822 -9.308 -4.633 -5.841 mfg '0' res "W (-11.814) (-3.878) (4.921) (8.803) (2.974) (8.881) Uchg- 8.773 -36.876 1 .436 0.054 16.390 6.033 (wpewpem (2.731) (2.420) (0.174) (0.008) (2.341) (0.747) Market Asts AIC 5451 486 1080 1216 1124 1059 LRI 0.172 0.192 0.260 0.363 0.191 0.037 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 137 Table 29 - Logit regression of year one failures onto base CAMEL and IRR variables except net income / total assets is used for earnings. Data Period for Independent Variables 1985-89 1985 1986 1987 1988 1989 Constant -2.867 -3.261 -3.538 -2.664 -1 .740 -2.293 (48.527) (4.578) (-7.078) (-7.134) (8.201) (-7.546) Capital- 8541 -19.901 -8.411 -9.480 -54. 160 -1.472 ("awn/masts») (45.787) (-7.326) (-6.447) (-9.029) (43.899) (4 .577) Asset 5.385 2.047 3.029 5.442 10.911 4.351 Quality- (6.337) (0.491) (1.291) (3.658) (3.787) (2.600) (repos'd asts/ttl asts) Management 8.226 12.680 6. 896 -1 2. 797 7.034 19.380 Efficiency- (2.810) (1.434) (1.247) (2.021) (0.885) (3.943) (non-int exp/ttl asts) Earnings 0.000 -0.001 -0.020 -0.001 -0.002 0.002 Ability- (0.648) (8.181) (8.835) (8.259) (-0.299) (2.842) (int incflnt exp) Liquidity- -8.657 -3.250 -2.240 -2.712 -5.622 -15.419 (amt 8'9 t" "=9 (4 0.239) (4 .339) (4.174) (2.053) (2.088) (-7.276) Irqlttl asts) Uchg- 17.138 -13.915 -12.691 8.230 16.926 18.856 mfim’w (4.828) (8.787) (8.830) (1.091) (1.782) (2.898) AIC 4522 403 486 1 030 61 7 1 198 LRI 0.258 0.287 0.227 0.274 0.673 0.134 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 138 Table 30 - Logit regression of year two failures onto base CAMEL and IRR variables except net income / total assets is used for earnings. Constant Capuak (net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (int inc/int exp) Liquidity- (amt elig for reg qu/ttl asts) Uchg- (MVPE-MVPEuy Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 2.398 4 .357 4.755 4 .223 -2.084 2.803 (4 4.805) (2.201) (8.070) (8.118) (8.387) (-7.896) -5.782 43.351 42.040 47.941 80.893 8.785 (8.424) (4.700) (-7577) (8.158) (-8.870) (-0.705) 11.147 4.985 8.145 25.147 19.198 2.599 (10.979) (1.339) (3.847) (8.885) (8.431) (1.247) 13.285 8.348 4 .478 8.792 24.872 13.285 (4.418) (0.835) (8.228) (-0.694) (3.742) (2.259) 0.000 8.015 8.014 0.000 0.002 8.002 (—0.155) (8.888) (4.228) (0.113) (2.050) (8.525) 8.308 42.208 8.805 8.243 4.413 8.737 (41.802) (8.858) (4.847) (8.254) (2.785) (-3.612) 8.797 89.150 1.290 -0.202 18.832 3.484 (2.081) (2.518) (0.151) (8.030) (2.605) (0.422) 5494 492 1 103 1208 1099 1081 0.166 0.182 0.244 0.387 0.210 0.035 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) _...1 139 Table 31 - Logit regression of year one failures onto base CAMEL and IRR variables except broker deposits / total assets is used for liquidity. Constant Capital- (net worth/ttl asts) Asset Quality- (repos'd astslttl asts) Management Efficiency- (non—int exp/ttl asts) Earnings Ability- (int inc/Int exp) Liquidity- (brokered deposits/Itl asts) Uchg- (MVPE-MVPE»! Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 8.932 4.028 8.835 8.148 2.280 8.737 (25.583) (8.859) (-8.785) (8.901) (4.283) (43.270) 8.818 25.517 -7.480 -8.080 -50.963 1.234 (41.932) (-7.652) (8.480) (-7.209) (-11.983) (1.171) 4.424 1.091 2.740 4.937 9.829 2.148 (5.040) (0.234) (1.124) (3.254) (3.284) (1.127) 3.924 27.042 3.437 44.072 1.808 10.273 (1.357) (2.969) (0.624) (2.073) (0.146) (1.887) 23.577 30.512 45.348 43.998 28.411 49.481 (8.138) (3.291) (8.530) (8.100) (2.240) (8.377) 2.075 4.059 8.392 0.827 0.808 2.886 (3.390) (1.897) (4 .058) (0.491 ) (0.423) (3.003) 21.487 -9.483 45.151 11.444 18.064 19.324 (5.999) (8.518) (-0.975) (1.518) (1.891) (2.877) 4552 393 476 1024 616 1213 0.254 0.305 0.243 0.279 0.873 0.123 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 140 Table 32 - Logit regression of year two failures onto base CAMEL and [RR variables except broker deposits / total assets is used for liquidity. Constant Capital- (net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inclltl asts) Liquidity- (brokend depositle asts) Uchg- (MVPE-MVPEUN Market Asts AIC LRI Data Period for Independent Variables 1985-89 1985 1988 1987 1988 1989 8.444 2.587 2.880 2.301 2.391 8.302 (23.814) (4.292) (8.404) (-6.247) (-6.260) (-10.764) 8.072 4 9.784 41.552 4 7.348 -33.164 4 .798 (-7.766) (8.810) (8.877) (8.027) (-9.301) (4.554) 10.979 6.111 8.993 25.288 19.893 2.769 (10.895) (1.696) (4.213) (8.547) (8.595) (1.309) 11.880 12.888 8.528 8.989 22.710 14.325 (4.079) (1.224) (8.515) (8.717) (3.357) (2.309) -12.483 20.574 42.321 8.328 7.045 23.480 (-3.665) (2.106) (4 .829) (4 .150) (0.591) (2.178) 2.928 2.633 2.315 4.430 2.547 3.124 (5.320) (1.302) (1.473) (4.293) (2.128) (2.852) 10.936 -39.722 1.028 4.620 19.883 5.828 (3.317) (2.508) (0.119) (0.706) (2.793) (0.707) 5843 51 1 1 132 1227 1 107 1070 0.143 0.148 0.224 0.358 0.204 0.027 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t- statistics are below parameter estimates.) 141 Table 33 - Logit regression of year one failures onto base CAMEL and [RR variables except cash and securities / total assets is used for liquidity. Constant- Capital- (net worth/ttl asts) Asset Quality- (repos’d asts/Itl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net incfttl asts) Liquidity- (cash and securitieslttl asts) Uchg- (MVPE-MVPE.» Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1986 1987 1988 1989 -3.884 8.721 8.884 8.158 2.231 8.583 (24.815) (8.592) (8.738) (-8.805) (4.214) (42.825) 8.898 25.494 -7.122 8.012 -50.884 0.720 (42.117) (-7.603) (8.283) (-7.158) (41.998) (0.700) 4.325 1.839 2.819 4.830 10.090 1.899 (4.913) (0.414) (1.158) (3.202) (3.348) (0.979) 3.905 24.154 4.189 4 3.822 1.220 9.484 (1.332) (2.679) (0.773) (2.014) (0.110) (1.707) 24.182 28.921 44.158 44.232 29.727 81.730 (8.329) (2.948) (8.287) (-3.131) (2.403) (-6.677) 2.517 0.100 8.902 2.735 0.643 2.818 (8.009) (0.032) (8.305) (4 .539) (0.313) (4.578) 25.412 45.119 44.288 15.988 18.837 21.819 (8.505) (8.824) (8.888) (1.946) (1.588) (2.792) 4552 398 477 1021 616 1218 0.254 0.300 0.241 0.280 0.673 0.119 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the pooled sample period as shown by the column headings. Failure is from the immediately following twelve months (year one). (t- statistics are below parameter estimates.) 142 Table 34 - Logit regression of year two failures onto base CAMEL and IRR variables except cash and securities / total assets is used for liquidity. Constant- Capital- (net mnh/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inclttl asts) Liquidity- (cash and securitieslttl asts) Uchg- (MVPE-MVPE»! Market Asts AIC LRI Data Period for Independent Variables 198589 1985 1988 1987 1988 1989 8.359 2.441 2.807 2.113 2.357 -3.187 (22.975) (4.128) (-7.983) (8.722) (-6.184) (40.392) 8.233 4 9.984 41.294 4 7.425 83.058 2.147 (8.071) (-5.851) (8.748) (8.152) (8.375) (4 .887) 1 1 .094 6.456 8.537 28.185 20.079 2.897 (10.821) (1.820) (3.938) (8.825) (6.678) (1.351) 12.215 11.345 8.144 8.509 23.535 14.288 (4.148) (1.091) (8.450) (4 .005) (3.558) (2.322) 42.998 18.249 -12.610 41.505 7.130 19.818 (-3.829) (1.927) (4 .866) (4.392) (0.595) (1.855) 4 .585 0.827 8.498 -0.115 4 .012 8.789 (2.248) (0.337) (4 .984) (8.087) (8.858) (-0.443) 12.440 43.489 7.548 3.279 20.873 5.391 (3.589) (2.791) (0.786) (0.463) (2.747) (0.834) 5881 513 1130 1243 1110 1075 0.140 0.146 0.225 0.348 0.201 0.022 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end data, either individual years or the entire sample period as shown by the column headings. Failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) N.C.- no convergeance 143 Table 35 - Logit regression of year one failures onto optimal CAMEL and IRR variables. Constant- 8.316 (1 3. 059) Capital-alternate -10.769 exp(net worth/ttl asts) (-1 9, O1 6) Asset Quality- -4.574 (percentage change in real estate value) (-5 923) Management Efficiency- 0.298 (non-int exp/ttl asts) (O. 1 02) Earnings Ability- -17.850 (net inc/ttl asts) (-7_097) Liquidity- -7.938 (amt elig for reg qu/ttl asts) (-9436) Uchg- 1 7.594 (MVPE-MVPEQIMarket Asts (4. 956) AIC 4290 LRI 0.297 Key: Optimality is defined as the variable which provides the highest LRI when substituted in with the base CAMEL variables and the base IRR. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) 144 Table 36 - Logit regression of year two failures onto optimal CAMEL and IRR variables. Constant- 8.360 (1 1 .188) Capital- -10.019 exp(net worth/ttl asts) (_1 4.71 7) Asset Quality- -7.696 (percentage change in real estate value) (-1 0_ 554) Management Efficiency- 9.723 (non-int exp/ttl asts) (2.961 ) Earnings Ability- -9.936 (net inclttl asts) (-2955) Liquidity- -8.187 (amt elig for reg qulttl asts) (-1 1 .457) Uchg- 8.102 (MVPE-MVPEQIMarket Asts (2493) AIC 5372 LRI 0.184 Key: Optimality is defined as the variable which provides the highest LRI when substituted in with the base CAMEL variables and the base IRR. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) Table 37 - Logit regression of year one failures onto base and optimal CAMEL and all three IRR variables. Base CAMEL Variables Optimal CAMEL Variables Constant -3. 577 Constant 7.544 (-7.166) (8.815) Capital- 6049 Capital- -10.563 (net worthfttl asts) (_9. 907) exp(net worth/ttl asts) (_1 6. 1 73) Asset Quality- 3.916 Asset Quality- -4.549 (repos'd asts/ttl asts) (4. 51 3) (percentage change in real estate (_5. 880) value) Management 2.951 Management 0.209 Efficiency- (1 .018) Efficiency- (0.071 ) (non-int exp/ttl asts) (mint exp/ttl asts) Earnings Ability- -22.766 Earnings Ability- -17.861 (net inclttl asts) (_8. 866) (net inc/ttl asts) (7.094) Liquidity- -8.438 Liquidity- -8. 144 (amt elig for reg qu/ttl asts) (_9. 81 5) (amt elig for reg qu/ttl asts) (_9. 447) Uchg- 1 0.463 Uchg- 15.637 (MVPE-MVPEQ/Market Asts (2 . 359) (MVPE-MVPEQ/Market Asts (3 541 ) ISF- 0.865 ISF- 0.691 (interest sensitive funds/m funds) (1 .542) (interest sensitive fundslttl funds) (1 . 229) SHORT- 0.769 SHORT— 0.137 (s.t. Iiab's - s.t. asts)/ttl asts (2 586) (st. Iiab's - s.t. asts)/III asts (0. 449) AIC 4426 AIC 4293 LRI 0.275 LRI 0.297 LRT 9.056 LRT 1 .735 (0.011) (0.420) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) Table 38 - Logit regression of year two failures onto base and optimal CAMEL and all three IRR variables. Base CAMEL Variables Optimal CAMEL Variables Constant -1 .308 Constant 9.760 (-3.489) (10.888) Capital- -4. 561 Capital- -10.337 (net worth/ttl asts) (-6. 544) exp(net worth/ttl asts) (-1 3744) Asset Quality- 10.530 Asset Quality- -7.709 (repos'd asts/ttl asts) (1 0. 31 5) (percentage change in real (_10572) estate value) Management 1 1.051 Management 9.556 Efflciency- (3.701 ) Efficiency- (2.929) (non-int exp/ttl asts) (non-int exp/ttl asts) Earnings Ability- -13.282 Earnings Ability- -10.057 (net inclttl asts) (-3952) (net "Wt" asts) (2.995) Liquidity- -7.833 Liquidity- —7.736 (amt elig for reg qu/ttl asts) (_1 O. 863) (amt elig for reg liq/ttl asts) (_1071 1 ) Uchg- 6.151 Uchg- 11.473 (MVPE-MVPEQ/Market Asts (1 . 51 0) (MVPE-MVP Eu)/Market Asts (2 846) ISF- ,. 4.295 ISF- .. -1.336 (1:13:39 sensrtrve funds/ttl (_3. 089) (3:338 sensitive funds/ttl (_3.174) SHORT- 0.280 SHORT- -0.170 (s.t. Iiab's - s.t. asts)/ttl asts (1 .051 ) (s.t. Iiab's - s.t. asts)/ttl asts (.0625) AIC 5472 AIC 5366 LRI 0.170 LRI 0.186 LRT 10.377 LRT 9.83 (0.006) (0.007) Key: Failure is represented by a one and health by a zero in the logit regression. Independent variables are from year-end, 1985-1989, and failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 147 Table 39 - Logit regression of year one failures onto base CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Constant- Capital- (net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Management Efficiency- (non-int exp/ttl asts) Earnings Ability- (net inc/ttI asts) Liquidity- (amt elig for reg liq/til asts) IRR IRR*exp(capita|) AIC LRI LRT Interest Rate Risk (IRR)- Uchg- (MVPE-MVPEJ/Market Asts ISF— (interest sensitive funds/ttl funds) SHORT- (s.t. Iiab's - s.t. asts)/ttl asts 2.933 (48.700) -1.122 (4.041) 1.821 (1.847) 4.015 (1.357) 22.838 (8.775) 8.217 (8.778) 240.400 (6.179) 220.000 (8.743) 4399 0.279 33.685 (0.000) 8.520 (-7.858) 39.848 (14.865) 2.540 (3.058) -1 .355 (8.478) 45.328 (8.289) -7711 (8.043) 63.161 (17.581) -61.456 (-17.495) 4128 0.324 323.13 (0.000) 2.433 (48.240) 20.452 (49.495) 5.298 (8.798) 2.554 (0.923) 44.177 (8.972) -7.724 (8.902) 82.300 (48.827) 33.877 (17.522) 4091 0.330 341.518 (0.000) Key: Results are shown for each IRR variable tested. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-I989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 148 Table 40 - Logit regression of year two failures onto base CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Interest Rate Risk (IRR)- Uchg- ISF- SHORT- (MVPE-MVPEu)/Market Asts (interest sensitive funds/ttl (s.t. Iiab's - s.t. asts)/ttl asts funds) Constant 2.512 8.178 4.733 (45.195) (-7957) (-13.611) Capital- -1.388 48.021 -22.160 ("9‘ WW" “5‘51 (-1.020) (1 1.738) (-1 6.428) Asset Quality- 9.756 9.515 11.789 (repos'd asts“ asts) (9.067) (9.243) (11.514) Management 1 1 .563 9.188 13.244 Efficiency- (3.939) (2.987) (4.482) (non~int exp/ttl asts) Earnings Ability- -12.810 -8.518 -5.399 (”I "m" as“) (-3.795) (-2.551) (-1.656) Liquidity- 8.249 -7.403 -7.614 (34.119101218118515) (41.584) (40.391) (40.347) IRR 140.000 67.678 -42.813 (2.813) (12.977) (45.324) |RR*exp(capitaI) -128.400 -66.226 42.141 (2.883) (43.289) (15.534) AIC 5473 5260 5145 LRI 0.169 0.201 0.219 LRT 7.133 217.48 334.625 (0.008) (0.000) (0.000) Key: Results are shown for each IRR variable tested. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 149 Table 41 - Logit regression of year one failures onto optimal CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Interest Rate Risk (IRR)- Uchg- ISF- SHORT- (MVPE-MVPEQ/Market Asts (interest sensitive funds/ttl (st. Iiab's - s.t. asts)/ttl asts funds) Constant- 9.483 32.956 22.421 (6.751 ) (4.786) (20.509) Capital- -1 1.931 -35.696 -24.404 WWt WW" ”'5’ (-8.710) (-5.215) (-23.145) Asset Quality- -4.640 -4.867 -4.073 (99mm Change 1" '95" (-5.970) (-6.280) (-5.275) estate value) Management 0.1 74 -0.249 0.041 Efficiency- (0.059) (-0.084) (0.014) (non-int explttl asts) Earnings Ability- -17.796 -16.857 -11.488 ("9‘ Ml ”'51 (-7.085) (-6.761 ) (-4.976) Liquidity- -7.914 -8.504 -7.298 (amt el‘9 '0' '99 "9"" 3515) (-9.420) (-9.874) (-8.457) IRR -15.659 -25.380 -27.091 (-0.439) (-3.475) (-1 6.957) IRR*exp(capital) 33.301 26.451 28.682 (0.938) (3.641) (17.586) AIC 4292 4297 3958 LRI 0.297 0.296 0.352 LRT 0.872 (0.350) Key: Results are shown for each IRR variable tested. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 150 Table 42 - Logit regression of year two failures onto optimal CAMEL variables, each IRR variable, and an interactive product between capital and IRR. Interest Rate Risk (IRR)- Uchg- ISF- SHORT- (MVPE-MVPEQ/Market Asts (interest sensitive funds/ttl (s.t. Iiab's - s.t. asts)/ttl asts funds) Constant 1 0.485 47.666 24.895 (6.034) (5.867) (17.888) Capital- -12.078 -47.027 -25.904 exp(net WW “‘51 (-7.250) (-5.958) (-19.417) Asset QualityI- -7.745 -7.727 -7.272 (Zig'mthange '" "33' (-1 0.590) (-1 0.604) (-9.934) Management 9.595 10.214 1 1 .407 Efficiency- (2.933) (3.174) (3.635) (non-int exp/ttl asts) Earnings Ability- -9.997 -9.153 -5.571 ("9‘ inc/I“ 3515) (-2.984) (-2.815) (-1.716) Liquidity- -8.180 -7.869 -7.521 (amt el‘9 '°' "-‘9 "W351“ (-1 1.457) (-10.897) (-10.271) IRR ~55.270 -41.657 -33.936 (-1.195) (-4.833) (-14.157) IRR*exp(capital) 61 .590 39.398 33.603 (1.373) (4.707) (14.324) AIC 5371 5348 5081 LRI 0.184 0.188 0.229 LRT 1 .857 (0.173) Key: Results are shown for each IRR variable tested. Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the following thirteen to twenty-four months (year two). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) Table 43 - Logit regression of year one failures onto base CAMEL variables, each IRR variable, and a series of interactive terms which include sequential dummies for capital levels. Base CAMEL Variables Optimal CAMEL Variables Constant- -2.285 Constant -0.792 (42.967) (4.291) Capital- -1 .004 Capital-alternate -1 .363 (net worth/ttl asts) (_2. 1 57) exp(net worthfttl asts) (_2. 456) Asset Quality- 0366 Asset Quality- -2.389 (repos'd asts/ttl asts) (_0. 483) (percentage change in real 03.199) estate value) Management 3. 879 Management 4.177 Efficiency- (1 .501 ) Efficiency- (1.602) (non-int exp/ttl asts) (non-int exp/til asts) Earnings Ability- -11.225 Earnings Ability- 9990 (net inc/ttl asts) (-5.626) (net inc/ttl asts) (-5.036) Liquidity- -4.897 Liquidity- -4.713 (amt eIig for reg qulttl asts) (-6 05) (amt elig for reg liq/til asts) (_5. 842) Uchg*C1- 53.942 Uchg*C1- 50.61 1 .1411“: (12719 1859:4892: (12.246, Uchg*CZ- 18.1 Uchg*C2- 16.389 (iiiovgen 0.05) (_10464) (C4=‘I when NW/ttl asts>.05) (_1 0436) AIC 3451 AIC 3439 LRI 0.435 LRI 0.438 LRT 984.6 LRT 857.4 (0.000) (0.000) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are in parenthesis.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) Table 44 - Logit regression of year one failures onto base or optimal CAMEL and IRR variables and a series of interactive terms which include sequential dummies for capital levels. Base CAMEL Variables Optimal CAMEL Variables Constant -2.285 Constant -0.792 (-12.967) (-1.291) Capital- -1 .004 Capital-alternate -1 .363 (net worth/ttl asts) (_2. 1 57) exp(net worth/ttl asts) (_2- 455) Asset Quality- 0366 Asset Quality- -2.389 (repos'd asts/ttl asts) (41483) (percentage change in real (_3 199) estate value) ' Management 3.879 Management 4.177 Efficiency- (1 .501 ) Efficiency- (1.602) (non-int exp/ttl asts) (non-int exp/ttl asts) Earnings Ability- -11.225 Earnings Ability- -9.990 (net inc/til asts) “5626) (net inc/ttl asts) (-5036) Liquidity- -4.897 Liquidity- -4.713 (amt elig for reg qu/ttl asts) (-6 050) (amt elig for reg qu/ttl asts) (-5842) Uchg- -1 12.800 Uchg- -112.100 MVPEoMVPEu I _ (MVPE-MVPE“ rket Asts _ 11.4.8.1. ’ (10464) W“ (10.436) Uchg*C1- 166.700 Uchg*C1- 182.700 (C1=1 when NW/ttl asts<0) (17.023) (C1=1 when NW/ttl asts<0) (16413) Uchg*CZ- 130.900 Uchg*C2- 128.500 = C = 8331614122882, 03-403) 88431328202, 113-‘32) Uchg*C3- 66.249 Uchg*C3- 65.065 = CB= fg§1,000,000,000) AIC LRI LRT 8.272 (12.897) -10.761 (-18.929) -4.660 (6.022) 0.875 (0.293) -17.750 (-7.074) -7.656 (8.932) 15.993 (1.815) 17.668 (1.461) 23.408 (4.847) 4291 0.297 3.2 (0.202) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) Table 46 - Logit regression of year one failures onto base or optimal CAMEL and IRR variables and a series of interactive terms which include sequential dummies for asset levels. Base CAMEL Variables Optimal CAMEL Variables Constant -2.959 Constant 8.272 (-16.783) (12.897) Capital- 6757 Capital- alternate -10.761 (net worth/til asts) (_1 2.099) exp(net worth/ttI asts) (_1 8929) Asset Quality- 4.033 Asset Quality- -4.660 (repos'd asts/ttl asts) (4630) (percentage change in real (5022) estate value) Management 3.792 Management 0.875 Efficiency- (1 .282) Efficiency- (0.293) (non-int exp/ttl asts) (non-int exp/ttl asts) Earnings Ability- -22.606 Earnings Ability- -17.750 (net inc/ttl asts) (-8_ 795) ("9‘ ‘ncm' “‘51 (~7.074) Liquidity- -7.990 Liquidity- -7.656 (amt elig for reg liq/ttl asts) (_9. 340) (amt elig for reg IIthtI asts) (3.932) Uchg- 22.898 Uchg- 23.408 -M PE“ MVPE-MVPE“ IM A mg“); v (4.687) < > W sts (4.847) Uchg*A1— -6.183 Uchg*A1- -7.413 = _ A = _ itflst‘srthef‘CIQOOODOO) ( 1 '521 ) 1tl 1152531“ ($8,000,000) ( 1 815) Uchg*AZ- -4.622 Uchg*A2- ~5.738 = . ' < _ A = ' . _ 225’5218‘8'70898‘3‘3” "' (“80> 2.352181323092311 °°°‘“' (”61> AIC 4432 AIC 4291 LRI 0.274 LRI 0.297 LRT 2.3 LRT 3.2 (0.317) (0.202) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 155 Table 47 - Logit regression of year one failures onto base CAMEL and [RR variables and a series of interactive terms which include sequential dummies for IRR levels. Capital- Capital- Capital- (net worth/ttl asts) (mnpv/market) exp(net worth/ttl asts) Constant -2.927 -3.444 7.527 (-16.853) (-21.139) (10.083) I1 *Capital- 8330 -0.630 -10.222 (“=1 when ”CW-02> (-6.057) (~3.790) (-15.202) I2*Capital- -7.102 -4.122 -1 0.325 1'2“ when -°2‘=U°h9<-°4) (-11.1 16) (~10.770) (-15.607) l3*CapitaI- -5.237 -2.556 -10.392 "3“ when -°4‘=U°h9l (-6.341) (-4.107) (-15.417) Asset Quality— 4.661 7.745 2.126 “em“ 35‘5“" as“) (5.214) (10.388) (2.414) Management Efficiency- 3.120 1 1.852 0372 (net inc/9' Inc) (1 .080) (4.583) (0128) Earnings Ability- -22.150 -28.288 -18.663 ("9' inch" 35") (-8.622) (-1 1 .378) (-7.338) Liquidity- -8.1 14 -8.047 8105 Ian“ 9'19 '°' "*9 11qu asts) (-9.665) (-10.008) (-9.575) Uchg- 17.488 20.143 21.050 (MVPE-MVPE..)/Market Asts (4.887) (5.784) (3_ 304) AIC 4428 4460 4326 LRI 0.275 0.269 0.291 LRT 6.3 1 .7 0.4 @043) (0.427) (0.819) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) Akaike Information Criteria (AIC) considers the number of parameters in the regression and allows cross specification comparison. The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 156 Table 48 - Logit regression of year one failures onto base CAMEL and IRR variables and a series of interactive terms which include sequential dummies for IRR levels. Capital- Capital- Capital- (net worth/ttl asts) (mnpv/market) exp(net worth/ttl asts) Constant- -2.927 -3.354 7.527 (-16.853) (-20.411) (10.083) Capital- -5.237 -4.621 -10.392 (-6.343) (-7.982) (-15.417) l1*Capital- -3.093 0.431 0.171 ("=1 when ”“9109 (-2.027) (0.941 ) (0.643) I2*Capital- -1.865 0.191 0.067 02“ When -°2‘=UCh9‘-°4l (-2. 168) (0.306) (0.420) Asset Quality- 4.661 7.092 2.126 ("3906“ asts/l" asts) (5.214) (9.545) (2.414) Management Efficiency— 3.120 10.524 0372 (net inC’SJ' “ml (1 .080) (3.923) (0128) Earnings Ability- -22.150 -26.896 -18.663 (net WI" 3515) (-8.622) (-10.804) (-7.338) Liquidity- -8.1 14 -8.246 8105 Inn" 9|‘9 '°' "-‘9 "W as“) (-9.665) (-10.217) (-9.575) Uchg- 17.488 19.719 21.050 (MVPE-MVPEJIMarket Asts (4_ 887) ( 5. 587) (3. 304) AIC 4428 4403 4326 LRI 0.275 0.279 0.291 LRT 6.3 1 .7 0.4 (0.043) (0.427) (0.819) Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). (t-statistics are below parameter estimates.) Akaike Information Criteria (AIC) considers the number of parameters in the regression and allows cross specification comparison. The Likelihood Ratio Test (LRT) tests the hypothesis that parameters for the dummy variables are different. (Chi-square statistics are in parenthesis.) 157 Table 49 - Logit regression of year one failures onto all CAMEL and IRR variables, a variable for hedging activities, and a size variable. Constant- Capital- (net worth/ttl asts) Capital- (MVPE/market) Capital- exp(net worth/ttl asts) Asset Quality- (repos'd asts/ttl asts) Asset Quality- (other real estate/ttl asts) Asset Quality- (percentage change in real estate value) Management Efficiency- (net inc/gr inc) Earnings Ability- (net inc/ttl asts) Earnings Ability- (int incfrnt exp) Liquidity- (amt elig for reg qulttl asts) Liquidity- (brokered deposits/til asts) Liquidity- (cash and securities/ttl asts) Hedging Securities- (Sum of hedging securities/Total Assets) Size- Log(Total Assets) Uchg- (MVPE-MVPEQIMarket Asts SHORT- (s.t. Iiab's - s.t. asts)/ttl asts ISF- (interest sensitive funds/ttl funds) AIC LRI -5.161 (-6.235) -4.737 (-7.569) -2.999 (-8.268) 2.556 (2.885) 0.261 (0.152) -4.246 (-5.3251 5.105 (1.672) -20.387 (-8.080) 0.000 (0.753) -8.096 (-9.170) 0.688 (1.059) 0.294 (0.341) 0.719 (1.760) 0.079 (2.169) 15.596 (3.285) 0.023 (0.073) 1.630 (2.639) 4314 0.296 3.876 (3.322) -2.539 (-6.793) -8.710 (-11.364) 0.878 (0.982) -0.183 (-0.105) -4.455 (-5,519) 1 .543 (0.506) -16.800 (-6.723) 0.000 (1.500) -7.955 (-8.923) 0.762 (1.170) 0.227 (0.261) 0.980 (2.451) 0.080 (2.177) 18.752 (3.858) -0.423 (-1 .343) 1.476 (2.373) 4233 0.309 -2.920 (-17.008) -6.875 (-12.299) 4.001 (4.626) 2.632 (0.923) -22.905 (8906) -8.025 (-9.578) 1 .389 (3.756) 17.028 (4.778) 4421 0.275 Key: Failure is represented by a one and health by a zero. Independent variables are from year-end, 1985-1989, and failure is from the immediately following twelve months (year one). Akaike Information Criteria (AIC) considers the number of parameters in the regression and allows cross specification comparison. (t-statistics are below parameter estimates.) Table 50 - Results of the two index model with various portfolios of S&Ls as the independent variable. 158 Model 1 Model 2 MKT INTLT MKTr INTT TILFIRM 0.665 0.001 0.680 -4.322 (5.99) (0.00) (6.68) (2.41) HCRTNS 0.628 0.000 0.640 -3.521 (5.68) (0.00) (6.20) (4.94) HIGHCAP 0.615 0.026 0.628 -2.828 (6.54) (0.14) (7.13) (-1.82) MEDCAP 0.630 0.029 0.648 -4.116 (4.76) (0.11) (5.23) (-1.89) LOWCAP 0.785 -.0858 0.798 -7.303 (4.61) (-0.26) (5.17) (-2.68) HIGHAST 0.858 0.017 0.878 -5.123 (4.73) (0.05) (5.15) (-1.71) LOWAST 0.624 -.016 0.636 -4.266 (6.11) (8.08) (6.84) (2.81) Quarterly Data: 1986:1-1989:4 lbbotsen/Asso 1 .147 1 .267 1 .431 6.990 (2.84) (4.93) (4.98) (1.62) CRSP -.644 .145 .030 -7.748 (0.94) (0.33) (0.07) (-1.21) Key: The data are monthly returns from CRSP for the independent variable and from Ibbotsen & Associates for the indices. The market index GVIKT) is the S&P 500, the long term interest rate index (INTLT) is the total return on long term government bonds, and the short term index (INTI) is the total return on twelve month Treasury bills. The test period is 1985:] through 1989:12. All returns are excess returns and the interest rate indices are ‘whitened’. (t-statistics are in parenthesis below.) 159 Table 51 - Results of the single stage model. Uchg SHORT ISF Uchg SHORT ISF Model 1 MKT INTLT*IRR INTLT 0.568 -5.158 0.150 (21.47) (-1.01) (0.94) 0.566 -0.855 0.729 (21.46) (-1.55) (1.53) 0.566 -0.186 0.012 (21.47) (-0.56) (0.21) Model 2 MKT INT1*IRR INT1 0.572 44.053 -3.921 (22.27) (1.25) (-3.22) 0.575 -2.171 -0.724 (22.49) (—0.75) (-0.29) 0.576 -1 .933 -2.390 (22.52) (-0.60) (-4.42) Key: The data are monthly returns from CRSP for the independent variable and from Ibbotsen & Associates for the indices. The market index (MKT) is the S&P 500, the long term interest rate index (INTLT) is the total return on long term government bonds, and the short term index (INT1) is the total return on twelve month Treasury bills. The test period is 1985:] through 1989:12. All returns are excess returns and the interest rate indices are ‘whitened’. (t-statistics are in parenthesis below.)