SHOULD THE CAPITAL REQUIREMENT ON HOUSING LENDING BE REDUCED? EVIDENCE FROM AUSTRALIAN BANKS. Australian Prudential Regulation Authority

Transcription

1

2 SHOULD THE CAPITAL REQUIREMENT ON HOUSING LENDING BE REDUCED? EVIDENCE FROM AUSTRALIAN BANKS Neil Esho and Alvin Liaw Working Paper Australian Prudential Regulation Authority June 2002 Abstract In its recent submissions to the Basel Committee on Banking Supervision, the Australian Prudential Regulation Authority argued that the risk weight on lending secured by residential mortgage should be reduced from 50% to 20%. Using Australian banking data, we estimate the relative riskiness of assets according to their regulatory imposed risk weights. The results suggest that for an average Australian bank over the period 1991 to 2001, a 10 percentage point shift in assets, from the 50% to 100% risk weight bucket, resulted in a 30% increase in the ratio of impaired assets to total assets, and a 45% (relative to the mean) increase in the ratio of credit losses to total assets. There is however no significant difference in risk associated with varying the proportion of assets held between the 20% and 50% risk weight buckets. This suggests the current 50% risk weight on housing lending may be excessive. Both authors are from the Australian Prudential Regulation Authority, Level 26, 400 George Street, Sydney NSW Ph: ; Fax: neil.esho@apra.gov.au; alvin.liaw@apra.gov.au. The views and opinions expressed in this paper are those of the authors and do not necessarily reflect those of APRA. The authors would like to thank Bob Allen, Wayne Byres, Charles Littrell, Anthony Coleman, Marianne Gizycki, Glenn Homan, Bill Jones, Murray Jones, Terry Pittorino and Ian Sharpe for helpful comments. APRA

3 I. Introduction In its recent submission to the Basel Committee on Banking Supervision, the Australian Prudential Regulation Authority (APRA, 2001) recommended that the capital adequacy risk weight applied to housing lending be reduced from 50% to 20%. 1 In support of the lower risk weight, the APRA submission cites evidence of historically low rates of default in the Australian home lending market. The submission also reports that based on the January 2001 proposed amendments to the Basel Accord, risk weights for housing lending calculated using the Internal Ratings Based (IRB) approach ranged from 5% to 10% at surveyed Australian banks. Under the proposed new Capital Accord, financial institutions will have to choose between either continuing with the standard regulatory defined credit risk weighting function or, subject to regulatory approval, may choose to determine capital charges based on their own internal credit ratings. In this framework, the setting of appropriate risk weights under the standardised approach takes on increased importance, with implications for regulatory neutrality, the efficient management of capital in the financial services industry, and risk taking incentives. With internally calculated risk weights on housing lending likely to be significantly lower than the standard 50% risk weight, institutions that remain on the standardised approach will be required to hold substantially larger amounts of capital, and may therefore be at a competitive disadvantage (APRA, 2001). Moreover, where a substantial deviation exists between the standard and IRB risk weights, this may encourage increased risk taking, and lead some deposit-taking institutions to increase their portfolio of risky commercial loans, where the IRB and standardised risk weights are likely to be more closely aligned. 1 The lower risk weight would apply to lending fully secured by residential property and with a loan to valuation ratio (LVR) less than or equal to 80%. Mortgage insured loans with LVRs greater than 80% would be risk weighted according to the insurer s external rating (see APRA, 2001, p.9). APRA 1

4 In this paper we provide evidence on the relative riskiness of housing lending, by estimating the effect of changes in portfolio composition on loan quality. A unique feature of this study is that portfolio weights are defined by the proportion of total onbalance sheet assets held in each risk weight category. This enables us to estimate the effect of substituting assets between different categories, and therefore to determine whether the relative risk weights are appropriate. This paper therefore sheds lights on the following important questions: Should the 50% risk weight that has historically been applied to housing lending be maintained in the new Capital Accord? How do shifts in portfolio concentrations between the 0, 20, 50 and 100 percent risk weight categories affect impaired assets? Controlling for portfolio composition, how does loan growth, bank efficiency, capital adequacy, and economic growth affect loan quality? The literature examining the determinants of impaired assets at financial institutions is limited to relatively few studies. Clair (1992) and Salas and Saurina (1999) examine the macroeconomic and microeconomic determinants of impaired assets using bank level data in studies of Texas and Spanish banks respectively. Other studies either estimate models using data aggregated at the state or national level (Keeton (1999)), or focus exclusively on the macroeconomic determinants (Gizycki (2001)) or microeconomic determinants of loan quality (Kwan and Eisenbeis (1995)). In this paper we extend the existing literature in two ways. Firstly, we control for portfolio composition using regulatory defined risk weight categories, allowing us to examine the appropriateness of risk-based capital charges. Secondly, our study provides Australian evidence on the importance of both macroeconomic and microeconomic variables in determining loan quality. APRA 2

5 The remainder of this paper is structured as follows. In the following section we specify a model of the determinants of loan quality. Section III summarises the data, the estimation and results are presented in section IV, and section V concludes. II. Model Specification In this section we model the determinants of loan quality as a function of portfolio composition, cost efficiency, loan growth, moral hazard costs and macroeconomic conditions. We use two measures of loan quality: impaired assets and credit losses. Initially we define the dependent variable as the log of the ratio of impaired assets to total assets denoted Impaired Assets, 2 where impaired assets include non-accrual items, restructured items (original contract terms have been modified to provide concessions of interest and/or principal), and assets acquired through security enforcement. 3 Impaired assets is commonly used to measure credit quality and has the advantage of providing a standardised approach to the recognition of actual and potential credit risk. However, a limitation of using impaired assets to measure credit quality, is that the amount of potential credit loss associated with a given impaired assets level can vary widely with changes in recovery rates. To illustrate, both nonaccrual loans and restructured loans reflect the rate of default, but do not account for the recovery rate on the loans, which can be highly variable and depend on the nature of security, type of borrower, and the stage of the credit cycle. As actual credit losses are a product of both the rate of default and the recovery rate, impaired assets may differ significantly from actual credit losses. To overcome the limitation of impaired assets 2 Clair (1992) and Salas and Saurina (1999) use a similar log transformation in their studies. 3 Non-accrual items include assets where there is doubt about the receipt of principal and interest, payments of principal and/or interest that are 90 or more days past due and where the fair value of security is insufficient to cover payments, overdrafts outside approved limits for more than 90 consecutive days where security is insufficient to cover the bank s exposure, and payments made on a cash basis due to a significant deterioration of the borrower. For more details on the definition of impaired assets see Guidance Note AGN 220.1, which is available from the APRA website at APRA 3

6 data, a second measure of credit quality is used which accounts for both default rates and recovery rates. The alternative dependent variable is denoted as Credit Losses, and is defined as the sum of net asset write-offs 4 and the change in specific provision, divided by total assets. 5 In addition to accounting for recovery rates, Credit Losses provides a measure of unexpected losses due to credit risk, which is the component of credit risk against which banks hold capital. Actual credit losses are however much more volatile than impaired assets and subject to the discretion of bankers decisions to write-off bad debts. Moreover, both Credit Losses and Impaired Assets measure the realisation of poor credit decisions and tend to ignore the build up of credit risk through the credit cycle. A. Portfolio Composition Underlying the international application of risk-based capital adequacy requirements for credit risk is an acceptance that portfolio composition is a fundamental determinant of bank risk. To control for the effects of portfolio composition on loan quality, Clair (1992) and Emmons (1996) include variables measuring the proportion of commercial and industrial loans to total loans and the proportion of real estate loans to total loans. Salas and Saurina (1999) also note the importance of portfolio composition as a determinant of loan quality, which they proxy by the ratio of loans without collateral to total loans. Several macroeconomic studies have highlighted the importance of portfolio composition on bank risk. Schwartz (2001), for example, provides evidence to show that prior to the 1930 s depression in the United States and Japanese recession of the 1990s, shifts in portfolio composition were important in fuelling excessive price rises. These periods were both followed by rapid asset price deflation, which left the 4 Net asset write-offs equals total write-offs, less recoveries made of previously written off amounts. 5 We are grateful to Bob Allen for suggesting the use of Credit Losses. APRA 4

7 respective financial systems in a severely weakened position. Carmichael and Esho (2001) present similar evidence showing that changes in the portfolio composition of Australian banks were an important factor in facilitating rapid asset price inflation in the late 1980s, which ultimately contributed to the build up of bad debts in the early 1990s. In this paper we define portfolio weights according to the proportion of onbalance sheet assets held in the 0, 20, 50, and 100 percent risk weight categories. 6 The risk weight (RW) variables are denoted RW0, RW20, RW50 and RW100 respectively. As the risk weight categories sum to one, we omit one of the risk weight categories from each regression, so that the coefficients on the remaining risk weight variables may be interpreted as measuring portfolio shifts relative to the omitted risk weight category. We expect that as an institution increases its reliance on assets in higher risk categories, that this will lead to greater risk and therefore an increase in Impaired Assets and Credit Losses. B. Loan Growth Keeton (1999) analyses the conditions under which rapid loan growth may lead to increases in loan losses. Where an increase in loan growth is due to a shift in the supply of loans, Keeton argues that this is likely to be caused by a reduction in credit standards, which ultimately leads to an increase in loan losses. Banks that attempt to increase market share typically do so by lowering both loan rates and credit standards. Even if a bank attempts to maintain credit standards, they face an adverse selection problem, as the pool of new customers attracted by the high growth bank may contain a higher proportion of riskier borrowers, since competition for the lowest quality borrowers will be weakest. The adverse selection problem may be compounded when banks attempt to increase growth rates by moving into new product or geographical 6 In August 1998 APRA removed the 10% risk weight category. To maintain consistency, prior to August 1998, RW0 includes assets that were previously assigned a 10% risk weight. Including off-balance sheet items has no effect on the results. APRA 5

8 markets. (Shaffer (1998)). It is also possible that rapidly growing banks may not devote sufficient additional resources to loan administration and monitoring, which can lead to higher future loan losses (Clair (1992)). Where loan growth occurs due to an increase in demand, Keeton argues that overall credit standards are likely to increase and loan losses should be expected to fall. Increased demand for bank loans may occur due to a shift away from internally generated cash flows and/or external capital markets, or because of a productivity shock which increases the creditworthiness of individual borrowers. Increased demand for bank loans is likely to raise the overall credit quality of loan applicants, allowing banks to raise credit standards and loan rates, and therefore reduce future loan losses. 7 Using quarterly data from the Senior Loan Officer (SLO) Survey conducted by the Federal Reserve, Keeton finds that the relationship between credit standards and loan growth in the 1990s is consistent with both supply shifts and productivity growth, providing partial support for the view that faster loan growth leads to higher loan losses. In the 1970s and 1980s however, there is evidence that loan growth and credit standards were positively related, suggesting that some of the loan growth in this period was driven by demand shifts, and would therefore not be expected to lead to an increase in future loan losses. Keeton also estimates the relationship between the level of consumer and industrial loans and associated loan losses directly using Federal Reserve call report data over the period 1982 to This data provides stronger support for a direct 7 In the case of a positive productivity shock the effect on credit standards (in the sense of accepting a greater proportion of below average credit quality borrowers) is uncertain, however loan losses are likely to fall as a result of the increase in ability of all borrowers to repay their debt. However, Clair (1992) notes that if loan demand is driven by investment in speculative asset bubbles, then the long run relationship between loan growth and loan losses may be negative. APRA 6

9 relationship between loan growth and loan losses, with loan losses occurring roughly two years after a positive shock to loan growth. 8 Clair (1992) examines the importance of loan growth in determining nonperforming loans and charge-off rates at Texas banks in the 1980s. A unique feature of the study is the disaggregation of loan growth according to whether the growth is internally generated, due to mergers, or the result of an acquisition of a failed bank with assistance of the Federal Deposit Insurance Corporation (FDIC). Clair finds that internally generated loan growth (current and one year lagged) leads to an improvement in both the non-performing loan ratio and loan charge-off rate. However, after a lag of three years there is statistically significant evidence that increasing internal loan growth leads to a significant positive increase in the rate of loan charge-offs. Interestingly, loan growth that is the result of mergers has an immediate positive impact on the loan charge-off rate, which reflects poor lending decisions made by the target bank prior to the merger. However, FDIC assisted acquisitions have an immediate negative effect on both the problem loans ratio and the loan charge-off rate, which is consistent with the FDIC absorbing the risk of failed institutions. A limitation of Clair s study is that the author does not control for unobserved bank effects in the estimation. Salas and Saurina (1999) also find that loan growth has a significant direct effect on problem loans in their sample of Spanish savings banks (commercial non-profit organisations focused on retail banking), however no significant relationship is observed in the sample of commercial banks. The existing literature provides theoretical and empirical support for the hypothesis that rapid loan growth leads to an initial improvement in credit quality. As it can take several years before the effects of rapid growth are reflected in deteriorating 8 It is unclear why Keeton (1999) uses the level of loans, rather than the loan growth rate when using the call report data. APRA 7

10 loan quality, the immediate impact of rapid asset growth is to lower Impaired Assets and Credit Losses. In the longer term however, the effects of rapid growth are reflected in higher levels of credit risk, leading to an increase in Impaired Assets and Credit Losses. In this study Loan Growth is measured by the percentage change in total assets, and is expected to have an immediate negative impact on Impaired Assets and Credit Losses. The lagged effect of Loan Growth, however is expected to be positively related to Impaired Assets and Credit Losses. The existing literature suggests that a lag of two to four years is sufficient to observe the positive effect of loan growth on impaired assets. Clair (1992) uses three annual lags of loan growth, while Salas and Saurina (1999) use three annual lags (lags two to four years), and Keeton (1999) uses 12 quarterly lags within a vector auto-regression estimation approach. In this study we initially include 8 quarterly lags of loan growth, and examine the sensitivity of the results to alternative lag structures. Maintaining a relatively short lag length allows us to maximise the sample size, estimate the model over a complete credit cycle, minimise the effects of multicollinearity associated with introducing lagged variables, and generally simplify the model. C. Cost Efficiency Within the efficiency literature, Berger and DeYoung (1997) examine several hypotheses that link non-performing loans (NPLs) with cost efficiency. They suggest that under the bad management hypothesis, managers that are inefficient in controlling operating costs are also inefficient in loan origination and monitoring. Poor cost control is reflected almost immediately in inefficiency, but the effect on non-performing loans is lagged as the consequences of poor loan origination and monitoring are eventually realised. The reverse relationship is suggested by the skimping hypothesis, which argues that banks may cut costs on loan origination and monitoring to increase short term cost efficiency, however in the longer term this leads to higher loan losses and increased APRA 8

11 inefficiency. Using Granger causality tests, Berger and DeYoung find support for the bad management hypothesis, which dominates the skimping hypothesis in their full sample results. Kwan and Eisenbeis (1995) examine the relationship between risk and inefficiency in a sample of 254 U.S. bank holding companies from 1986 to The authors use several measures of risk including: the standard deviation of daily stock returns, the standard deviation of residuals from the market model, the market value of equity to book value of assets, the ratio of charge-offs to loans outstanding and the book value of equity to asset ratio. Within each of four size quartiles, the authors find increased risk is positively correlated with inefficiency. Finally, in the Australian context, Esho (2001) finds a significant negative relationship between cost efficiency and the ratio of loans in arrears to total loans in a sample of Australian credit unions. To account for the effects of inefficiency on credit quality, we include the cost to income ratio denoted Inefficiency, where cost is measured as total non-interest operating expenses, and income is the sum of net interest income and non-interest income. Due to the offsetting predictions of the skimping and bad management hypotheses the relationship between Inefficiency and Impaired Assets (and Credit Losses) is uncertain. As the skimping and bad management hypotheses suggest the effect of Inefficiency on credit quality will occur with a lag, we allow for a two year lagged effect. D. Moral Hazard The moral hazard hypothesis suggests that shareholders at weakly capitalised banks have an incentive to take advantage of incorrectly priced deposit insurance (implicit or explicit) by increasing risk taking. In this framework, banks operating at low capital levels will take on increased risk, which will be reflected in an increase in impaired assets and credit losses. Using Granger causality tests, Berger and DeYoung (1997) find that an increase in the capital ratio leads to a reduction in NPLs at thinly APRA 9

12 capitalised U.S. banks, providing some support for the hypothesis that shareholders at weak banks exploit bondholders and incorrectly priced deposit insurance, by taking on riskier loans, which eventually leads to an increase in non-performing loans. 9 Clair (1992) and Salas and Saurina (1999) also use the capital ratio to proxy moral hazard costs and find evidence consistent with the moral hazard hypothesis that lowly capitalised banks take on increased risk. We include the ratio of regulatory capital to risk adjusted assets, denoted Capital Ratio, with eight quarterly lags and expect it to be negatively related to Impaired Assets and Credit Losses. E. Macroeconomic Conditions The relationship between macroeconomic conditions and performance of the financial system has been extensively examined in the academic literature. (See for example Bernanke, Gertler and Gilchrist (1996) and Borio, Furfine and Lowe (2000)). It is important to control for the effect of the business cycle in this study, as loan growth, changes in portfolio composition and changes in impaired assets may appear to be related, when in fact the variables may simply be correlated with the business cycle. In a recent study, Gizycki (2001) examines the determinants of impaired assets at Australian banks, focusing on the role of macroeconomic variables. Using a fixed effects estimation approach, Gizycki finds support for the effect of various macroeconomic variables on impaired assets growth. Emmons (1996) examines the relative importance of local economic conditions and bank risk taking in predicting the probability of bank failure in the U.S. over the period 1970 to Using state level data, Emmons proxies local economic conditions by the failure rate of business in each 9 Although Australia is one of few countries without a system of explicit deposit insurance, governments in the past have provided guarantees to depositors in troubled financial institutions (see Fitz-Gibbon and Gizycki, 2001). This creates a potential moral hazard problem. APRA 10

13 state, and finds that both local economic conditions and bank risk taking are important in predicting bank failure rates across states. 10 To control for the influence of the business cycle on the ability of borrowers to repay their debt, we include Real GDP Growth (with two quarterly lags) in the model and expect it to be negatively related to Impaired Assets and Credit Losses. F. Controls To control for the effects of mergers on impaired assets, we include a dummy variable, Merger, which in a given quarter, takes the value of one if the institution was involved in a merger and zero otherwise. In September 1994, the regulatory definition of impaired assets was altered. 11 To account for a break in the Impaired Assets series we include a dummy variable denoted Sept94, which takes on the value of one from September 1994, and zero prior to September Finally, a Time Trend variable is included to account for trends in the dependent variable, and quarterly dummy variables, denoted Quarter 2, Quarter 3 and Quarter 4 to control for seasonal variation in the data. III. Data Quarterly data submitted by Australian banks over the period December 1991 to June 2001 were obtained from statistical returns submitted to the Australian Prudential Regulation Authority (and its predecessor as bank supervisor, the Reserve Bank of Australia). The sample includes all Australian banks for which data is available for at least 15 quarters, and excludes foreign banking subsidiaries and foreign branches, which for the most part undertake specialised investment banking functions. 12 The sample 10 Clair (1992) and Salas and Saurina (1999) also find favourable macroeconomic conditions are directly related to loan quality. 11 The objective of the changes was to establish a set of minimum standards and a common approach for identifying, measuring and reporting impaired assets in order to improve the quality and consistency of data received from banks. 12 Appendix 1 contains a list of the banks in the sample. APRA 11

14 comprises 16 banks with a total of 503 quarterly observations and is based on the global consolidated operations of each bank. 13 Summary statistics for the dependent and independent variables are reported in Table 1. Over the sample period, the mean bank held roughly 13% of assets in the 0% risk weight category, 8% in the 20% risk weight category, 40% in the 50% risk weight category, and 39% of assets in the 100% risk weight category. The mean bank has an impaired assets ratio of 1.8%, a credit losses ratio of 0.065%, has a risk-adjusted capital ratio of 11.32%, grew at an average quarterly rate of 2.8% and has a cost-to-income ratio equal to 65%. Finally, the average quarterly growth in real GDP over the sample period was 1%. A correlation matrix of first differenced variables is presented in Table Consistent with the expectations regarding portfolio composition, Impaired Assets is significantly negatively correlated with changes in RW0 and positively correlated with changes in RW100. There is also a significant negative correlation between current Loan Growth and Impaired Assets, supporting the view that rapid loan growth leads to an initial improvement in loan quality. The seven-quarter lagged effect of Loan Growth on Impaired Assets is positive, but statistically insignificant. The correlation between changes in the Capital Ratio and Impaired Assets is insignificantly positive for the current period capital ratio, but significantly negatively for the seven-quarter lagged Capital Ratio, providing some support for the hypothesis that banks take on increased risk when capital levels are low. Correlations between changes in Credit Losses and the explanatory variables follow a similar pattern to that observed with Impaired Assets, though generally less 13 The sample of 503 observations is determined after first differencing all variables, including lagged variables, and excludes the Inefficiency variable, for which only annual data is available. One observation is lost using credit losses, reducing the sample to 502 quarterly observations. 14 Based on a sample of 503 observations a correlation of (0.088) is statistically significant at the 1% (5%) level. APRA 12

15 significant. Credit Losses is as expected significantly positively correlated with Impaired Assets, and significantly negatively correlated with changes in the holding of assets in RW20. There is also a positive correlation between Credit Losses and RW100, however the correlation is not statistically significant at the 5% level. The sample period covers one of the most severe downturns in the credit cycle confronted by Australian banks. The cycle in loan quality and its relationship to the level of economic growth is illustrated in Figure 1. The figure shows a strong positive relationship between the level of gross domestic product (inverse log scale) and Impaired Assets and Credit Losses. Impaired assets peaked at around 6% of total assets in early Average Credit Losses peaked at 0.37% and are highly volatile up until As noted previously, the sharp spikes in Credit Losses reflects the fact that writeoffs ignore the build up of credit risk through the credit cycle and are more subjective in their timing. 15 Shifts in the portfolio composition of Australian banks are illustrated in Figures 2 and 3. In Figure 2 the average portfolio weights are constructed using equally weighted data, while Figure 3 gives greater weight to the large banks in the sample by calculating the value weighted (or aggregate) portfolio weights. 16 Over the sample period, both charts indicate that banks increased their holdings of assets in the 50% risk weight category, while reducing the proportion of assets held in the 100% risk weight category. The shift towards housing lending was most evident in the period 1991 to 15 The increase in bad debts was preceded by major deregulation of the banking system, increased competition in the financial services industry, and a significant increase in corporate sector leverage. See Carmichael and Esho (2001) for further discussion of the causes of the increase in bad debts in the early 1990s. 16 The Australian banking market is dominated by 4 national banks which, as at June 2001, accounted for 83% of total banking industry assets. APRA 13

16 The proportion of assets held in the 0% and 20% risk weight categories remained stable through the period. The most notable difference between Figures 2 and 3 is the proportion of assets held in the 100% risk weight category is substantially higher in Figure 3, which gives greater emphasis to the 4 major Australian banks. There is also greater variability in the portfolio weights in Figure 2, which implies that the portfolio mix of regional banks is more volatile than that of the four major banks. IV Estimation and Results Summarising the hypotheses detailed in Section III, the model may be written as follows: LoanQuality it = α RW0 h= 8 h= 2 h= 8 h= 0 + α RW 20 + α RW100 α Quarter2 + α Quarter3 + α Quarter4 + µ + ε 6 1 h h it λ LoanGrowth γ Capital Ratio t 2 7 i, t h i, t h it 1 4 t 3 + δ Inefficiency 8 it it + i, t h= 2 h= 0 + δ + α Merger + α TimeTrend + 5 β Real GDP Growth t 2 h Inefficiency i it t i, t 4 + t h + (1) where two alternative measures of Loan Quality are utilised, Impaired Assets and Credit Losses, µ i denotes unobserved firm specific effects, ε it is the error term, and all other variables are as defined in Section III. As the data required to calculate Inefficiency is only available annually and is limited to 402 observations, we estimated the model with and without the Inefficiency variable. 18 Note that in the above specification the proportion of assets held in the 50% risk weight category has been omitted. The 17 During this period, banks were also increasing their use of mortgage securitisation, suggesting that the shift towards housing lending may have been even greater than suggested by Figure 2. The increase in housing lending was also partly due to changes in regulation. In December 1993, the 50% risk weight on loans secured by mortgage over residential property was extended to include property not owned by the borrower. In September 1994, the 50% risk weight applying to loans secured by residential mortgage was restricted to loans with a maximum LVR of 80%. However, this rule was amended in August 1998 allowing housing loans with greater than an 80% LVR to qualify for the 50% risk-weight, provided the loan is covered with an adequate level of mortgage insurance through an acceptable mortgage insurer. 18 As the inefficiency measure only changes annually, a minimum lag of four quarters is required. To avoid spurious correlation with the denominator of the dependent variable the current and one period lagged values of Loan Growth have been omitted. APRA 14

17 coefficients on the remaining risk weight variables therefore measure the effect of shifting assets out of the 50% risk weight category and into the 0%, 20% and 100 % risk weight categories respectively. A. Impaired Assets Using Impaired Assets as the dependent variable, regression results with and without the Inefficiency variable are reported as Models 1 and 2 respectively in Table 3. Summary statistics reported at the foot of Table 3 indicate there is no evidence of serial correlation in the first differenced error terms, which implies that in levels the errors are strongly positively correlated. Under the fixed and random effects assumption that the residuals are serially uncorrelated, correlation of the first differenced residuals is implied to be 0.5 (see Wooldridge (2002)). As the serial correlation tests strongly reject the fixed and random effects assumptions, the models are estimated in first difference form, which results in the removal of the unobserved firm effects from the estimation. To account for heteroskedasticity, White adjusted standard errors are computed. Given that the data has been first differenced, the models have good explanatory power, explaining roughly 17.3% and 15.0% of the growth in impaired assets respectively. Across the two models there is consistent support for the portfolio composition, loan growth, moral hazard and macroeconomic hypotheses, but no support for the hypothesis that managers that are poor in controlling costs are also poor in managing loan quality. The coefficients on the risk weight variables reveal some interesting information about the relative riskiness of asset classes, as defined by regulators. Consistent with the low risk of assets in the zero risk weight bucket, the coefficient on RW0 is negative, though statistically insignificant. The effect of substituting assets between the 50% and 20% risk weight buckets, has no statistically significant impact on the impaired assets APRA 15

18 ratio. This suggests that the differential capital treatment of assets currently in the 20% and 50% risk weight buckets may be inappropriate. Moreover, given the positive coefficient on RW20 and the insignificance of RW0, the results support the lowering of the capital charge on assets currently in the RW50, rather than an increase of the capital charge on assets in RW20. In both models, the coefficient on RW100 is significant at the 1% level, indicating that shifting assets out of RW50 and into RW100 is associated with a significant increase in impaired assets. Specifically, a 10 percentage point shift in assets, from RW50 to RW100, leads to an estimated 30% (25%) increase in the ratio of impaired assets to total assets in Model 1 (2). 19 The most notable difference in the results between Models 1 and 2 is observed in the significance of Real GDP Growth. While consistently taking a negative coefficient, Real GDP is statistically significant in Model 2 with up to a six months lag, but only significant at the 10% level in Model 1 at a 6 months lag. This may be explained however by the shorter sample period in Model 1, which because of missing inefficiency data begins in 1993, and therefore does not include a complete credit cycle. Consistent with the expectation that rapid loan growth has a lagged adverse effect on loan quality, loan growth lagged seven quarters has a significant positive coefficient in both models. This finding is similar to existing studies which suggest that it takes roughly two years before the effects of rapid loan growth are reflected in a deterioration in loan quality. Evidence supporting the moral hazard hypothesis is mixed, with the current period Capital Ratio positive and significant at the 1% level, but significantly negative with a seven quarter lag. The latter results suggests that banks with low capital levels have an incentive to increase risk taking, which ultimately leads to an increase in 19 The percentage change in Impaired Assets is given by e (α3 RW100) 1, where α3 is the estimated coefficient on RW100 and RW100 is the assumed change in portfolio weight. APRA 16

19 Impaired Assets, consistent with the moral hazard hypothesis. The positive coefficient on the current period capital ratio is however inconsistent with the moral hazard hypothesis, but may be explained by the actions of regulators, which require higher risk banks to hold higher levels of capital. That is, capital ratios are increased as increases in impaired assets are realised, implying a positive relationship. Finally, there is no evidence to suggest that banks that are cost inefficient are also less efficient in credit evaluation and monitoring, with the coefficient on Inefficiency(t-4) insignificant. There is evidence of a significant adjustment in Impaired Assets associated with the regulatory changes implemented in September The Trend variable is also negative and highly significant, consistent with the general decline in the impaired asset ratio observed over the sample period. A.1 Impaired Assets - Sensitivity Analysis The choice of lag length used in the previous analysis, and existing literature, may be criticised as being ad hoc. To examine the robustness of our results, we tried numerous alternative lag structures. Firstly, we re-estimated the models omitting the insignificant lags of the Capital Ratio and Loan Growth variables. Secondly, we extended the lag length of these variables to twelve quarterly lags. In both alternatives the results (not reported) were largely unchanged. 20 To remove the noise associated with using quarterly data and significantly simplify the lag structure, the models were re-estimated using annual data. The quarterly dummy variables are omitted from the annual data regressions, and the Sept94 variable is replaced by a dummy variable, denoted Year95, which takes the value of zero prior to 1995 and one otherwise. The results from the annual data regressions are reported in Table 4, and despite the considerable reduction in sample size, are consistent with the 20 The Akaike Information Criteria and Schwarz Information Criteria were both minimised at zero lags. However, given the strong economic reasons for expecting lagged effects, we decided against applying a purely statistical approach to determining lag length. APRA 17

20 quarterly data results presented in Table 3. There are however two differences between the quarterly and annual data results that should be noted. Firstly, in the annual data results there is now significant evidence supporting the inefficiency hypothesis, with the coefficient on Inefficiency (t-4) significantly positive at the 5% level. Secondly, the explanatory power of models 1 and 2 increases significantly to 37.4% and 31.4% respectively when using annual data. This is not surprising as the annual data removes some of the noise associated with using quarterly impaired assets data. B. Credit Losses Due to the high quarterly volatility in credit losses observed in the early 1990s (see Figure 1), regressions using Credit Losses as the dependent variable were estimated using annual data with the results reported in Table Summary statistics reported at the foot of Tables 5 support the first differenced estimation approach, with the test for serial correlation in the first differenced residuals significantly different from the fixed and random effects assumption of Models 1 and 2 explain roughly 18.6% and 14.8% of Credit Losses respectively, which given the dependent variable proxies unexpected losses is substantial. In both models, Credit Losses are declining over time and significantly negatively related to Real GDP Growth and the lagged Capital Ratio. There is however no evidence to support the inefficiency or rapid loan growth hypotheses, while the effects of mergers on credit losses are also insignificant over the sample period. These results also show that portfolio composition and in particular the proportion of assets held in the 100% risk weight category is a major determinant of Credit Losses. While there is no significant variation in Credit Losses associated with substituting assets between the 20% and 50% risk weight buckets, the coefficient on RW0 is unexpectedly positive and significant at the 5% level. This anomalous finding 21 The quarterly loss data is averaged to form annual observations. APRA 18

21 may be driven by the high volatility in credit losses observed in the early part of the sample period, and/or the timing subjectivity associated with writing off bad debts. In order to examine the sensitivity of the results to the high volatility in Credit Losses in the early 1990 s, the models were re-estimated using a post 1993 sample period. The results reported in Table 6 show that the coefficient on RW0 is no longer significant, while the coefficient on RW100 remains positive and statistically significant. A 10 percentage point substitution of assets from RW50 to RW100, leads to an % increase in the ratio of credit losses to total assets in Model 2, which represents a 45% increase when measured relative to the mean credit loss to total assets ratio. The sub-sample results also show that varying the proportion of assets held among the 0%, 20% or 50% risk weight category has no significant effect on Credit Losses. Consistent with the regressions using impaired assets, the current period capital ratio is directly related to credit losses, while the lagged capital ratio is significantly negatively related to credit losses. Interestingly, the merger dummy variable is positive and statistically significant, suggesting that banks engaged in mergers incurred high credit losses in the year of the merger. Consistent with the results in Table 5, there is no support for the loan growth or inefficiency hypotheses, while the insignificance of Real GDP Growth may be explained by the shorter sample period. V. Conclusion In response to the primary question posed at the outset of this paper, we conclude that there is evidence to support the lowering of the risk weight on lending secured by residential mortgage in the Australian market. This finding supports the industry view that housing lending in the Australian market has been a historically low risk investment. The low loss rate on housing lending in Australia may be due to the high level of home ownership, favourable bankruptcy laws that permit lenders to recover residual balances from borrowers where sale of the security is insufficient, the non-tax APRA 19

22 deductibility of interest on home loans which encourages borrowers to reduce debt and, the significance of mortgage insurance in mitigating losses for lenders (Standard and Poor s (1999)). 22 The results also show that banks with low capital levels increase risk taking, and that as expected, aggregate economic output is inversely related to both impaired assets and credit losses. Support for the hypothesis that rapid loan growth has a lagged adverse effect on loan quality, and the hypothesis that cost efficiency is positively related to loan quality is only observed when credit quality is proxied by Impaired Assets. It is important to interpret the findings of this study carefully. Firstly, the results in this paper do not suggest that overall capital requirements are too high, but that the relative risk weight on lending secured by residential mortgage may be excessive. Hence, the lowering of the risk weight on housing lending should be accompanied by other regulatory changes, if the overall level of capital is not to be reduced. Secondly, the results apply for an average bank in an average year. To the extent that capital provides a cushion against unexpected events, then a higher capital charge on housing lending may still be appropriate, if unexpected losses from housing lending are likely to be higher than unexpected losses associated with holding assets in the 20% risk weight category. While the above limitation is most relevant to the model using impaired assets data, it also applies to models using actual credit losses, which do not account for potential credit losses. Thirdly, these results should not be interpreted as suggesting that any institution can simply increase its portfolio of housing loans at little or no risk. In addition to requiring the necessary lending expertise, institutions require sufficient portfolio diversification, and must have in place adequate procedures and processes. Moreover, institutions that attempt to rapidly expand in any product market still face the 22 In Australia, lending institutions with exposure to high loan-to-valuation ratios (ie with LVR>80%) are generally mortgage insured. This effectively transfers the credit risk of these exposures to mortgage insurers (APRA, 2001). APRA 20

23 problem of adverse selection. Finally, history has shown that property markets are on occasion prone to sharp price corrections, which is generally not observed for assets in the 20% risk weight category. We therefore conclude that while there are clear grounds for applying a risk weight below 50% for housing lending in Australia, there are also valid reasons for not reducing the risk weight all the way to 20%. APRA 21

24 REFERENCES Australian Prudential Regulation Authority (2001), Submission to the Basel Committee on Banking Supervision: The New Basel Capital Accord. Berger, A.N. and R. DeYoung (1997), Problem Loans and Cost Efficiency in Commercial Banks, Journal of Banking and Finance, 21(6), Bernanke, B., Gertler, M. and S.Gilchrist (1996), The Financial Accelerator and the Flight to Quality, The Review of Economics and Statistics, 78(1), Borio, C., Furfine, C. and P.Lowe (2000), Procyclicality of the Financial System and Financial Stability: Issues and Policy Options, Bank for International Settlements. Carmichael, J. and N. Esho (2001), Asset Price Bubbles and Prudential Regulation, Australian Prudential Regulation Authority, Working Paper Clair, R. (1992), Loan Growth and Loan Quality: Some Preliminary Evidence from Texas Banks, Federal Reserve Bank of Dallas Economic Review, Third Quarter, 1992, Emmons, W. R. (1996), Increased Risk-Taking Versus Local Economic Conditions as Causes of Bank Failures, Bank Structure and Competition, Esho, N. (2001), The Determinants of Cost Efficiency in Cooperative Financial Institutions: Australian Evidence, Journal of Banking and Finance, 25, Fitz-Gibbon, B. and M. Gizycki (2001), A History of Last-resort Lending and Other Support for Troubled Financial Institutions in Australia, Reserve Bank of Australia, Research Discussion Paper. Gizycki, M. (2001), The Effect of Macroeconomic Conditions on Banks Risk and Profitability, Reserve Bank of Australia, Research Discussion Paper. Keeton, W. (1999), Does Faster Loan Growth Lead to Higher Loan Losses?, Federal Reserve Bank of Kansas City Economic Review, 84(2), Kwan S. H. and R.A. Eisenbeis (1995), An Analysis of Inefficiencies in Banking: A Stochastic Cost Frontier Approach Federal Reserve Bank of San Francisco, Working Paper Salas, V. and J. Saurina (1999), Credit Risk in Two Institutional Regimes: Spanish Commercial and Savings Banks, Mimeo. Shaffer, S. (1998), The Winner s Curse in Banking, Journal of Financial Intermediation 7, Schwartz, A.J. (2001), Asset Price Inflation and Monetary Policy, Paper presented at the Annual Meeting of the American Financial Association, New Orleans, 7 January. APRA 22

25 Standard and Poor s (1999), Structured Finance: Australia & New Zealand. Wooldridge, J. M. (2002), Econometric Analysis of Cross Section and Panel Data, The MIT Press, Cambridge, Massachusetts. APRA 23

26 Table 1 Summary Statistics: Full Sample (N = 503) Variable Mean Standard Error Minimum Maximum Median Impaired Assets (log) (Impaired Assets/Total Assets) Credit Losses* (%) RW RW RW RW Loan Growth Real GDP Growth Capital Ratio Inefficiency * *Credit Losses data is limited to 502 observations, while data required to measure Inefficiency is only available annually and is limited to 402 observations. APRA

27 Table 2 Correlation Matrix Correlations are calculated using quarterly observations after first differencing all variables. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (1) Impaired Assets (2) Credit Losses (3) RW (4) RW (5) RW (6) RW (7) Loan Growth (8) Loan Growth (t-7) (9) Real GDP Growth (10) Capital Ratio (11) Capital Ratio (t-7) (12) Inefficiency (13) Inefficiency (t-4) APRA - 25

28 Table 3 The Determinants of Impaired Assets at Australian Banks: The dependent variable, Impaired Assets, is defined as the log (impaired assets / total assets). Ordinary Least Squares regression estimates are calculated after first differencing all variables. T- statistics are reported in parentheses and are calculated using White adjusted standard errors. The sample includes all Australian banks for which at least 15 quarterly observations are available over the period December 1991 to June In each case the 50% risk weight bucket is omitted from the regressions. ***,**,* indicate that the regression coefficients are statistically significant at the 1%, 5% and 10% levels respectively. Variable Model 1 Model 2 RW0 RW20 RW100 Real GDP Growth Real GDP Growth (t-1) Real GDP Growth (t-2) Loan Growth (t-2) Loan Growth (t-3) Loan Growth (t-4) Loan Growth (t-5) Loan Growth (t-6) Loan Growth (t-7) Loan Growth (t-8) Capital Ratio Capital Ratio (t-1) Capital Ratio (t-2) Capital Ratio (t-3) Capital Ratio (t-4) Capital Ratio (t-5) Capital Ratio (t-6) Capital Ratio (t-7) Capital Ratio (t-8) Inefficiency (t-4) Merger Sept94 Trend Quarter2 Quarter3 Quarter4 Observations Adjusted R Squared Serial Correlation (-1.49) (1.27) (3.25) *** (-1.25) (-1.41) (-1.72) * (0.96) (0.13) (0.18) (0.46) (0.29) (2.51) ** (1.42) (3.38) *** (-0.23) (0.46) (0.40) (0.11) (-0.01) (-1.22) (-1.76) * (-0.67) (0.99) (-0.25) (-3.77) *** (-4.93) *** (-1.57) (-0.48) (-3.57) *** (-1.46) (0.87) (3.16) *** (-1.76) * (-1.95) * (-2.03) ** (0.96) (-0.00) (0.27) (0.24) (-0.16) (1.87) * (0.63) (3.50) *** (-0.34) (0.04) (0.03) (0.63) (0.10) (-1.29) (-1.87) * (-1.09) (-0.17) (-3.79) *** (-5.35) *** (-2.09) ** (-1.12) (-3.65) *** APRA - 26