Aggregate Bankruptcy Probabilities and Their Role in Explaining Banks' Loan Losses

Andreeva, Olga Working Paper Aggregate Bankruptcy Probabilities and Their Role in Explaining Banks' Loan Losses Working Paper, No. 2004/2 Provided in Cooperation with: Norges Bank, Oslo Suggested Citation: Andreeva, Olga (2004) : Aggregate Bankruptcy Probabilities and Their Role in Explaining Banks' Loan Losses, Working Paper, No. 2004/2, ISBN 82-7553-226-4, Norges Bank, Oslo, https://hdl.handle.net/11250/2498535 This Version is available at: https://hdl.handle.net/10419/209827 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/4.0/deed.no ANO 2004/2 Oslo February 26, 2004 Working Paper Research Department Aggregate bankruptcy probabilities and their role in explaining banks’ loan losses by Olga Andreeva ISSN 0801-2504 (printed), 1502-8143 (online) ISBN 82-7553-225-6 (printed), 82-7553-226-4 (online) Working papers from Norges Bank can be ordered by e-mail: [email protected] or from Norges Bank, Subscription service, P.O.Box. 1179 Sentrum N-0107Oslo, Norway. Tel. +47 22 31 63 83, Fax. +47 22 41 31 05 Working papers from 1999 onwards are available as pdf-files on the bank’s web site: www.norges-bank.no, under “Publications”. Norges Bank’s working papers present research projects and reports (not usually in their final form) and are intended inter alia to enable the author to benefit from the comments of colleagues and other interested parties. Views and conclusions expressed in working papers are the responsibility of the authors alone. Working papers fra Norges Bank kan bestilles over e-post: [email protected] eller ved henvendelse til: Norges Bank, Abonnementsservice Postboks 1179 Sentrum 0107 Oslo Telefon 22 31 63 83, Telefaks 22 41 31 05 Fra 1999 og senere er publikasjonene tilgjengelige som pdf-filer på www.norges-bank.no, under “Publikasjoner”. Working papers inneholder forskningsarbeider og utredninger som vanligvis ikke har fått sin endelige form. Hensikten er blant annet at forfatteren kan motta kommentarer fra kolleger og andre interesserte. Synspunkter og konklusjoner i arbeidene står for forfatternes regning. 1 Aggregate bankruptcy probabilities and their role in explaining banks’ loan losses Olga Andreeva 26 February 2004 Abstract Increased competition forces banks to narrow lending margins and at the same time relaxed lending standards worsen the pool of borrowers. To preserve sound banking system it is important task to monitor credit risk as one of the dominant factors leading to bank failures and financial vulnerability. Norwegian banks traditionally have a large share of loans to nonfinancial enterprises in their investment portfolios, and we focus on risk related to loans provided to limited liability enterprises. By combining statistics on loans to Norwegian industries and regions and bankruptcy probabilities for individual corporate borrowers, we construct a proxy reflecting risk profile of the banks’ loan portfolios. Aggregation within industries and counties provides a bank-level panel of risk indicators, which are used to estimate banks’ loan losses during the period 1988 – 2001. Constructed aggregate bankruptcy probabilities prove to be meaningful measures, which explain loan losses if we control for the macroeconomic and bank specific factors. JEL Code: G21, C81 Key words: Bank losses, bankruptcy probabilities, aggregation Acknowledgement: I would like to thank Bent Vale, Kjersti-Gro Lindquist, Glenn Hoggarth and participants of the seminars in the Research Department and Financial Stability Wing for valuable comments and discussions. 2 1. Introduction One of the most important roles of banks as financial intermediaries is allocation of credit, screening and monitoring of borrowers’ creditworthiness, and maintaining relationships with reliable customers, which they can do on a lower costs than individual agents. Bank loans are especially valuable for small firms that are not publicly traded and thus are constrained with financial resources due to the limited access to the financial markets.1 Well functioning financial markets and market discipline play an important role for preserving soundness of the banking system and keeping risks in adequate limits. However, market failures, free-rider problems of gaining benefits from collected information and other forms of distorted incentives of economic agents advocate for the presence of sound regulation.2 The New Basel Capital Accord also emphasises supervisory review process as an important part of controlling risks in banking. Credit risk and financial stability Financial system is exposed to four major types of risks related to the financial intermediaries: liquidity risk, market risk, credit risk, and operational risk. One of the central issues of the financial stability reports is to measure and monitor these risks, examine risks patterns and assess financial system vulnerability to them. Risk control policy is especially important in banks, the largest part of financial intermediaries, as bank failures induce large costs on the economy, society and government.3 It is widely recognised that credit risk is one of the dominant factors leading to bank failures and financial vulnerability. Lending is a main function of universal commercial banks and is even more inherent to savings banks, which allocate almost all attracted deposits to loans. Moreover, other types of risk reinforce credit risk to some extent, as for instance, due to the interest rate movements and changes in operational environment with counterparties bank may be exposed to higher credit risk. Banks may take excessive risks due to various factors from intentional risk taking and high risk tolerance in a competitive environment in situations of moral hazard and adverse selection.4 Even banks that apply good risk measurement techniques can underestimate potential risks due to low-frequency and high-severity event which may produce huge but almost unanticipated losses. As it is emphasised in Herring (1999), banks are often influenced by a special form of financial vulnerability, disaster myopia, when they undervalue default probabilities if failures do not arise for a long time. And even if a bank uses superior credit risk models that indicate higher risk pricing, it may lose in competition to other banks, which disregard this risk and therefore may choose herding behaviour. Increased competition from credit markets forces banks to narrow spreads and at the same time relaxed lending standards worsen the pool of borrowers.5 Strong competition with disaster myopia, short termism and herding may therefore increase financial vulnerability of banks. 1 Diamond (1991a ), Becketti and Morris (1992). 2 Financial Stability Review, Bank of England (2000-2002). 3 See in more details in Mailath and Mester (1994), Frydl (1999), Hoggarth, Reis and Saporta (2002). 4 See Mishkin (1991) on a discussion o asymmetric information and agency costs as causes of financial instability. 5 See a discussion in Salas and Saurina (2002a) and Matutes and Vives (2000) on risk taking behaviour of banks as a response to changes in competition and market power. 3 Market discipline is also diminished by insured liabilities of the banks since banks depositors are secured and thus have less incentive for control. Bank assets can be easily misallocated as they can borrow easier and therefore take higher risks in asset allocation. Since sound banking and financial health are essential factors for financial stability, it is important to monitor bank risk exposure to the corporate sector, changes in lending patterns and ensuing losses. Credit risk and loan losses Most of the borrowers on the credit market have limited liability on their obligations to the bank and therefore lenders are exposed to the risk of borrowers default. Problem loans are one of the major reasons of financial difficulties, especially for banks with a large scale of traditional lending activities. To insure themselves, at least partially, from the borrowers’ failure to repay, banks set aside loan loss provisions for expected losses on doubtful debts. Bank practices differ with respect to the rules used in definition of expected losses and estimations of loan loss provisions. Norwegian practice defines expected losses as losses inherent in the loan portfolio but not yet realised, and therefore loss provisions are based only on the current information. However, expected losses may also be defined as all possible future losses that can occur due to both current and future events, and thus indicate how much loss provisions a bank can make to account for possible future losses. Making such loan loss provisions, banks can write off losses against them and thus reduce the risk of weaker profitability and capital adequacy when losses are recognised. Systematic under-provisioning policy exposes bank credit portfolio to additional risk, as the bank may be unprepared to withstand shocks and maintain solvency. At the same time, variation of losses is uncertain, and therefore unexpected loss should also be considered a possible danger for bank financial situation that increases the probability of insolvency, especially if the bank does not maintain sufficient capital in relation to its assets. Uncertain magnitude of possible losses gives rise to the credit risk. While loss provisions may cover expected losses on loans, bank capital in excess of the required minimum helps to absorb unexpected losses so that a bank can maintain solvency. When banks decide on their lending policy they have a trade-off between short-term gain from risk-taking and long-term losses on loans and possible bankruptcy or takeover. Considered costs and losses also include expected loss, assessment of its possible variability and opportunity cost of allocating capital and liabilities. Expected loss can be calculated on the basis of borrowers’ creditworthiness and correlation of loss exposure of different loans in the portfolio. If allocation of credit is not profitable, a bank may increase interest rate on loans or collateral requirements to reduce expected loss if it cannot reduce costs. However, this policy is not always sustained due to the downward competition press on interest rates. Approaches to credit risk and motivation for the study Due to the common concern of regulators in many countries about the financial stability a lot of effort has been done in the direction of assessment of credit risk and construction of warning indicators based on these measures. Credit risk is associated with the possibility that the borrower will not fulfil its contractual obligations and depends on the general macroeconomic situation, lending standards, i.e. interest rate, collateral requirements and other loan covenants, and legal enforcement mechanism, including the capacity to recover part of the loan after the default. There exist many different approaches to measuring credit risk and assessing its influence on bank performance. Value at risk models (VaR), option-based and insurance 4 approach6 to risk measurement and also rating-based models try to quantify credit risks and exposures of the banks. The size of risk is measured as the amount of a potential loss that can be incurred by a bank with some probability. Some of the models are designed on quite a sophisticated level and they often require extensive data for different contingencies and even confidential information related to the banks’ internal accounts and customers’ financial position. Lack of this information or low quality information can widely decrease supervisory effects from these models. A natural approach to the credit risk measurement when credit claims are not tradable is to measure a probability of default to occur and amount of loss given that default. Loss in the event of default is the amount of money that the bank will not be able to recover less possible recoveries on collateral. Then expected loss is a probability of default over the next year multiplied by the loss given default. But accurate estimation of the default probabilities requires quite detailed information on borrowers. Norwegian banks are mainly engaged in traditional banking with loans constituting the largest part of their assets. Therefore, we concentrate on a narrow meaning of the credit risk, i.e. risk related to bank loans. The aim of the analysis is to construct a proxy for the credit risk measure to reflect risk profile of the banks’ loan portfolios. In order to do this we aggregate risk indicators for banks on the basis of bankruptcy probabilities for individual corporate borrowers7, and estimate how these indicators can explain banks’ loan losses during the period 1988 - 2001. Two types of annual data are combined for this study: detailed bank statistics on loans specified for each county and industry and statistics for individual non-financial enterprises with limited liability. To construct a risk measure for a bank, bankruptcy probabilities for enterprises are aggregated within county and/or industry groups and then weighted by the volume of loans granted to each of these groups by this bank. Commercial banks have higher share of corporate loans, while savings banks traditionally provide loans mostly to households. However, historically mortgages are safer than loans to corporations (within the present and the New Basel Capital Accord house mortgages are also considered less risky), therefore we do not lose much by focusing on industrial loans in our risk assessment. Constructing a risk measure for the banks’ loan portfolios which can explain bank loan losses is an important task in studying the banking system and preserving its soundness. 2. Description of the datasets Statistics on bank loans We consider annual aggregate volumes of domestic loans of the Norwegian savings and commercial banks and branches and subsidiaries of foreign banks in Norway to the nonfinancial institutions classified by industry and county.8 The number of Norwegian banks is gradually decreasing from around 150 savings banks and 20 commercial banks at the beginning of the sample period to 130 and 12 banks respectively in 1999/2000. At the same time, volume of loans adjusted for the Consumer price index (CPI) index is generally growing with exception of 1990-1991 and 1993-1994. The data in its most disaggregated form is represented by loans 6 See Saunders (1999) on VaR, KMV, insurance and other approaches to credit risk measurement. 7 See Bernhardsen (2001) and Eklund, et al (2001) for estimation of individual bankruptcy probabilities. 8 Information is taken from the banks financial reports (Report 60). Data on loans granted by other financial enterprises and mortgage companies, which constitute almost 40 per cent of all observations (around 20 per cent in volume of loans), are available only from 1996 and are not included in the data set. 5 to around thirty – sixty industries9 and nineteen counties10 because information on the individual borrowers of each bank is not available. According to this type of classification we combine data from the banks’ end of year balance sheets with annual statistics on individual enterprises along two dimensions: industry dimension and industry/county dimension. Later they are refered to as industry/year and industry/county/year groups.11 We use only the data on loans granted by banks to the sector of limited liability enterprises over the years 1988 - 2001. Data was controlled against negative observations for loans and positive observations for loan loss provisions. Observations with missing or zero industry and county codes were dropped. Statistics on enterprises (annual financial statements): SEBRA-database The SEBRA-database is a broad dataset on limited liabilities enterprises. We have excluded companies in the oil and gas industry, financial industry and public sector. It contains information from annual financial statements of the enterprises registered at the Norwegian register for business enterprises over the years 1988-2001. The data set contains 1,399,119 observations in total for 14 years. The number of enterprises submitting their financial records was constantly growing from 47,641 in 1988 to 137,201 in 2000 with a small decrease in 1994, but there is a large drop of more than 7 per cent in the last period of the data set, year 2001. At the same time, number of enterprises in different industries and counties varies from just a few to several thousands. This pattern is similar to the statistics on loans, which can be explained by a relatively low level of activities in some counties and industries. The dataset was checked for missing observations for those enterprises that provide accounting information not on a regular basis. The data was controlled against missing and zero industry and county codes, and also against observations with industry codes that do not correspond with aggregate codes in the bank statistics. The SEBRA model12 predicts bankruptcy probabilities for individual enterprises with book value of total assets exceeding 250,000-300,000 NOK on the basis of accounting statements. An observation is defined as a record with financial and other relevant information submitted by an enterprise (referring to its unique identification number) available in the database for a particular year. High average bankruptcy probabilities with large deviations, i.e. mean value larger than 0.036 and standard deviation larger than 0.065, which corresponds to the upper 25 per cent, are found in many industries especially during the Norwegian banking crisis years 1990-1993. High bankruptcy probabilities during the years beyond the crisis are found in the following industries: Fishing, Manufacture of office machinery and computers, Hotels and restaurants, Post and telecommunication, Recreation, cultural and sporting activities, Other service activities. These industries traditionally have high uncertainty in their activities, which is particularly true for the hotel, restaurants, recreation, service activities and fishing. However, Real estate activities, which are also considered risky, show quite stable and low values of bankruptcy probabilities throughout the sample period. 9 Standard classification includes 32 industries before 1991, 33 industries up to 1996, 58 industries in 1996-1997 and 59 industries up to 2001. 10 Observations for counties 21 – 23 were joined in county 21 (Svalbard) as counties 22 and 23 are not defined in the enterprise statistics, and observations for county 2 (Akershus) and county 3 (Oslo) were joined in county 3 (Oslo/Akershus) due to the geographical and economic interrelations of these counties. 11 Since we use data classified by industry, changes in the type of industry classification in the bank reports (i.e. the number and contents of specified industries) can explain the variation in the number of groups (e.g. introduction of a more detailed classification in 1996 gives a rise in the number of observations to more than 6,400 compared to around 4,200 in the previous years). 12 See Bernhardsen (2001) and Eklund, et al (2001) for a description of the model. 6 Linking of the datasets and aggregation of individual bankruptcy probabilites The SEBRA-database contains only industry codes consistent with SIC94 as they were previously converted from SIC83 for all enterprises, while the bank statistics use old aggregate classification of industries in Reports 60 up to 1996. Therefore, for the data before 1996 we assign old aggregate codes to enterprises using relationship patterns between old aggregate codes and SIC83, and between SIC83 and SIC94. For the data from 1996 to 2001, assignment of the aggregate industry codes, valid in the bank statistics after 1996, to enterprises in the SEBRA-database is made according to the relationship pattern between SIC94 and aggregate codes in the Report 60. In this respect, a formal correspondence pattern between two industry classifications is utilised, where possible; whereas some artificial relationship between them is suggested, where necessary.13 After establishing a correspondence between industry codes in the bank statistics and industry codes for the individual enterprises, we aggregate individual bankruptcy probabilities, obtained for each enterprise from the SEBRA-model. Referring to the two common dimensions for the banks’ reports and the SEBRA-database, we use industry/year groups, i.e. the aggregate across all counties, and industry/county/year groups. The first type of aggregation mixes observations across counties and can be in disagreement with the county specific type of activities of the medium-size savings banks. However, it provides a direct link between the two datasets. Moreover, it may be more accurate than the second one if banks in their annual reports assign counties on some other basis (e.g. location of the local branch which an enterprises uses for its loan application), than the formal registration criteria used in the SEBRA-database. The second type of aggregation allows utilisation of higher variation in risk indicators, i.e. over larger number of groups. Volumes of debt to the financial institutions or the levels of activities, represented, for example, by total assets or operating revenues are used as weights in aggregation. It is reasonable to focus on the enterprises with non-zero ‘debt in financial institutions’ since only these enterprises will inflict a loss for the bank in the event of bankruptcy. Probability of non-repayment of the loan may depend on the borrowers’ prospects and type of business as well as financial strength and liquidity characteristics. These factors are incorporated into the bankruptcy probabilities through financial ratios reflecting companies’ earnings, liquidity and solidity, as well as companies and industry characteristics (age, size, and deviations of the profitability, liquidity and solidity from industries averages).14 Therefore, aggregated bankruptcy probabilites serve as a good risk indicator and can be used to estimate loan losses for individual banks. However, we do not have a direct link to the borrowers of each bank and also financial information is subject to a quick change, which creates a scope for upward or downward biases in loan losses estimation based only on these risk measures. So banks’ risk profile is not completely reproduced and when we model bank loan losses we need to incorporate some proxies for distinguishing between banks’ lending policies. Therefore, we consider also macroeconomic data, interest rate, and some bank-specific information which is discussed below. 13 See a detailed description of these procedures in the appendix “Combining bank statistics on loans with statistics on non-financial enterprises”. 14 See Bernhardsen (2001) and Eklund, et al (2001). 13 ratio can create additional stimulus for risk-taking behaviour, while decreasing equity market accompanied by increased uncertainty and lower expectations lead to lower capital buffers and also increased risks due to the worsening of corporate accounts. Then, other things equal, banks may be less willing to take risks. At the same time, if bank managers value bank solvency and soundness quite low and prefer to keep low equity-assets ratio, they may also prefer higher and more volatile profits and may take higher risks. Thus, the size of the equity-asset ratio allows us to incorporate the influence of bank buffer to withstand risk and shows bank willingness to take risk. A more general measure is a capitalasset ratio but it is less informative as its rise may also happen due to the increase in loss reserves.22 Large banks usually have changes in capital-asset ratio due to the increase in loss reserves or decrease in assets, while we are more interested to track changes in equity. Capital buffer safeguards against unexpected risks of the banks loan portfolio. These risks can be connected to the macroeconomic downturns, payment problems or bankruptcies of individual enterprises, increased lending to the corporate sector, concentration in particular industries, lower risk pricing and expansion to new customers. The latter factors are associated with intensified competition in banking. A bank with low capital, i.e. just above the minimum capital adequacy requirements, have high probability of being perceived as risky in the market, and therefore will have to borrow on worse terms and may experience liquidity problems. Growth in loan portfolio Rate of growth in loan portfolio reflects a rate of bank expansion in lending. High loan growth contributes to the reduction in capital adequacy, and therefore banks need solid profits to maintain funding and cannot sustain high loan growth for a long time. So lending is limited by the capital adequacy requirement when banks would like to raise new equity capital through new issues. Fast increase in lending may also cause higher loan losses through lower credit standards and larger increase in bad loans than in loans to creditworthy customers.23 Moreover, lowering of credit standard compensated by lending margin may be followed by higher degree of moral hazard and adverse selection. Non-performing loans Non-performing loans are loans that have not been written off but are at least 90 days overdue, non-accruing or other problem loans with renegotiated terms. A change in the credit risk has an impact on the size of non-performing loans and non-performing loans net of loss provisions (net non-performing loans). The size of non-performing loans reflects already defaulted (overdue) loans and can be different for banks with dissimilar lending specializations, and therefore reveals different information than aggregate bankruptcy probabilities. There is a time span between changes in credit risk and recorded problems with loans, as an enterprise with liquidity problems may not default on the loan if its shareholders agree to inject new capital. Banks may also undertake some loan restructuring, i.e. payment extensions, favourable change in terms of loan agreements, etc. Then loans are not considered non-performing. Aggregate bankruptcy 22 Loss reserves are not used in the bank balances after 1995 and are excluded from equity in the data before 1995. 23 However, Keeton (1999) argues that a relation between loan growth and losses does not occur only due to supply side which can be associated with softening of lending terms, i.e. lower interest rate, lower collateral requirements, lenient assessment of borrowers, etc. Changes in the demand and productivity can also cause an increase in lending when it comes along with the tightening of the credit standards. 14 probabilities and size of non-performing loans or a ratio of non-performing loans are only weakly correlated with coefficient of correlation around -0.02/0.12, while rates of loan losses are strongly related to the current level of non-performing loans. The development pattern of the non-performing loans differs also from the change in the number of bankruptcies. Establishments and bankruptcies of small enterprises are quite common especially in some sectors of the economy. But some of these businesses are considered risky from the start and do not get ordinary loans. Lending margin Banks’ financial results depend to a large extent on structure of lending and associated risks, and therefore loan pricing and credit risk measurement is an important component of banks’ financial strategy. Lending margin reflects credit risk, goals for long-term profitability, administration costs and costs of funding. The size of risk included in the lending margin can also depend on the valuation of collateral because during the upward trend in the housing prices banks have lower risk of loan portfolio default. European financial reviews reflect a common tendency to a better risk management and greater importance of adequate risk pricing of loans. Increased lending margin and holdings of larger equity capital can lead to the same conclusion in Norway. Lending margin was reduced after the years of crisis only in 1994 and then after a short time was increased again in 1998. In this study lending margin is calculated as a difference between bank’s interest rate on loans and interest rate paid on the three month treasure bills, a proxy for the money market rate (i.e. marginal funding costs). Due to the varying banks’ policies with respect to costs and risk pricing, lending margin is a more useful variable for explaining loan losses than macroeconomic changes in the interest rate, which have less direct effect on the ability of enterprises to serve their debt. A decrease in banks’ lending margin carries a possibility that banks’ pricing policy is too mild and does not adequately reflect risks associated with corporate lending. In the conditions of intensified competition some banks review their pricing policy to win market shares. They can do this by reducing cost, by pricing risk lower and by decreasing their profits on loans. If this happens as a consequence of lower cost and better risk management then the bank can compete on the loan market maintaining its financial wealth, otherwise the risk of loan portfolio will markedly increase while earnings will deteriorate. Therefore, a decline in lending margins may increase banks’ vulnerability to future losses on loans as risk may be priced inadequately. Risk management/management quality It is quite difficult to find an adequate proxy for the quality of banks credit policy. Management quality may be proxied by various profitability characteristics (i.e. the size of earnings before losses related to assets, return on equity) or cost effectiveness (i.e. total operating expenses related to average total assets). Profitability reflects bank’s ability to generate revenue to cover incurred costs, pay dividends and retain profit. Banks may have increased profitability due to the increase in the rate of return or due to the change in the composition of assets and liabilities. But changes in return to assets, which are net of loss provisions, usually reflect changes in the size of the latter and thus may be misleading for our model. 15 Risk aversion Lower risk aversion may cause banks to value profit possibilities more than possible costs of risk taking. Then unstable profits will cause much higher losses to the banks in case of macroeconomic shock, and this will be also aggravated by the influenced of these adverse shocks on financial situation of the banks’ risky borrowers. Large variability in earnings and higher than average losses can serve as an indicator of risk-taking, i.e. banks with higher losses tend to have superior profits in the previous years and possibly charge higher interest on their loans to compensate for risk. As an indirect evidence of high-risk taking we can consider a loan to asset ratio, especially to risky industries or industries where higher interest rates are charged. Risk diversification A well-diversified bank may have lower risk as investments are spread over various industries and regions. If a bank provides loans to the industry or region with high bankruptcy probability it increases the bankruptcy probability of its total loan portfolio, while loans to the industries and regions with low bankruptcy probabilities have a mitigation effect. Specialising in a particular group of loans will carry higher risks and therefore an increase in the expected loss because of the higher probability of bankruptcy in this group. Moreover, large investment in a particular group reduces diversification in loan portfolio. A bank with low degree of diversification may still have comparable risk due to the higher expertise in particular industries. However, small low diversified banks may also have to accept higher risk due to the stronger competition. A proper diversification of credit risk may lead to a much lower risk associated with loans. Savings banks have lower risk due to the higher share of mortgages, and thus they can and may be willing to decrease their credit standards and make loans with a higher default probability among corporate borrowers. Then they can profit from the possibility of charging higher interest rate to a variety of borrowers of lower credit class but at the same time incur costs of higher probability of bankruptcy of their borrowers. Large share of mortgages decreases the variability of possible loan losses and therefore makes banks more willing to engage in such policy as they benefit more than they lose. At the same time, risk-weighted debt shows similar patterns with cyclical movements for most of the primary industries and counties in Norway.24 Therefore, loan losses in banks are also expected to have some cyclical pattern with limited diversification opportunities across industry groups. Moreover, data availability constraints us by the assumption that banks have the same bankruptcy probabilities on loans inside a particular industry or region. Thus we should be aware that while possible diversification opportunities across major industries are limited, they are not taken into account at all within the industries and counties. Market power The degree of market power of a bank has implication for bank loan losses through its influence on the size of lending and deposit margins and also incentives to monitor borrowers.25 At the same time, the degree of market power lowers incentives to take excessive risk through increased charter value of the bank and thus the size of losses in the case of failure due to the excessive risk-taking.26 24 Calculation of the risk-weighed debt was done in Eklund et al (2001) 25 Caminal and Matutes (2002). 26 See a discussion in Perotti and Suarez (2002). 16 Competition in banking Bank competition has a positive effect on the efficiency but it may also lead to an excessive risk-taking. A bank can expand its credit portfolio by underbidding its competitors or by accepting borrowers with lower creditworthiness. In the situation of intensified competition, in order to have compatible earnings banks may either try to compete by cost reduction or begin to expand aggressively and attract new clients that may highly increase their risk exposure. The latter contributes to the strategy of entering new industries and regions where banks do not have information advantage. There was a sharp increase in the number of bank branches as a result of increased competition and larger freedom in new branch establishment. However, rapid expansion in new industries and geographical regions put banks’ lending portfolios under higher risk than average in these industries and regions. Expanding banks possess limited information about customers from new market segments where they have little experience in specific conditions and particular characteristics of the borrowers. So they either should increase their screening and monitoring costs or tolerate higher risk and compensate it with larger lending margin. The latter was more apparent in expanding and optimistic economic conditions. However, this provided wider scope for unexpected risk, which together softened capital regulations27 created higher fragility in the banking. The other side of the expansion into new sectors was a myopic and herding behaviour of bank managers. Steigum (1992) suggests that deficient accounting made it possible for them to show high profits at the first stages independent of the loan quality due to the large initial charges on loans apart from the interest rate. Herd behaviour is consistent with a strategy to show high profits and expand when other financial institutions are doing so, otherwise bank managers are punished for unsuccessful policy in the short-term. This they can trade off with long-term benefit of non-herd behaviour. But under some conditions, herding is a prevailing rational strategy for all agents and can be another cause of following financial fragility. Lower risk pricing contributes to a decrease in lending margin. The size of lending and deposit margin, and spreads between banks can serve as indicators of the strength of competition. Narrowing difference between interest margins in different market segments indicates stronger competition both for new and existing customers. Difference between large and small/medium size banks Large banks have proven to have sound loan portfolio partly due to the better risk management strategies, higher possibilities for diversification and advantage in monitoring (cost reduction). Default costs are relatively higher for banks with small borrowers, as they have to administrate more bankruptcies with small repayment amounts. Moreover, the probability of borrowers’ default may increase even more if higher interest rate will lead to moral hazard problems and cause firms to take larger risks. At the same time, small banks are more likely to deal with small businesses, are more flexible and have better possibilities in resolving conflicts of interest. According to Boyd and Runkle (1993) small banks, which operate in restricted markets, receive higher economic rents. However, risk increases due to the expansion to new industries, regions and customer from new market segments of which they have little information and experience. Therefore, variables reflecting changes in the industry/region 27 Following Steigum (1992), capital requirements for Norwegian banks were reduced to 6.5 per cent in 1985 and then even further, when regulation allowed equity capital to be replaced by subordinated loan capital. 17 composition in banks’ loan portfolio may reflect not only willingness to expand to new market segments because of risk-taking or stronger competition, but also the difference between large and small banks. 5. Background information and estimation methods Separately aggregated data for loans to households and non-financial enterprises is used to explain corresponding loan losses. The essential component in the regression equation for nonfinancial enterprises is therefore risk-weighted debt ii iN p L ∈ ∑ , where L is amount of short and long-term debt of enterprises and p is bankruptcy probability for each enterprise from the set N of non-financial enterprises. Theoretically L should be a loss given default, as generally the bank loses not the whole amount of the loan after the borrower’s bankruptcy. From empirical data we can conclude that only around 30 – 50 per cent of the loan can be restored in the case of bankruptcy, but more detailed data on all loans is not available. A simple regression of total loan losses on the risk-weighted debt and housing index as a collateral proxy produces statistically and economically significant results with a good explanatory power. It is reasonable to assume that banks are heterogeneous from their external characteristics, as size, scope of operations, earnings, to internal characteristics such as client and investment policy, risk management, i.e. indicators on risk taken, tolerance to risk and following amount of buffer capital, competitive behaviour and costs. At the same time, it is even more interesting to look at the heterogeneity over time due to the known bank crisis in Norway in the beginning of 90-s, as we would like to have good explanatory variables, which can reflect variation in loan loss before, during and after the crisis. The aim of the analysis is thus to estimate how risk profile imposed on the bank by chosen loan portfolios can explain loan losses during the period 1988 – 2001 and especially during the banking crisis of the early 90-s. In this study we use panel estimation methods, which have higher estimation ability of the heterogeneous data by utilising two sources of variation in the data. While cross-sectional data helps to explain some relations relying only on the heterogeneity between individuals at the given moment in time and time series capture variations over time, longitudinal data addresses both inter-individual and between-individual variation. Even quite short time series but moderate size cross-sectional data provide good possibilities for explaining variations in the data. We have unbalanced characteristics of the dataset because of the bank mergers, closure of banks and their subsidiaries, and new bank establishments. We observe around 150 - 170 banks in the sample, and the largest fraction, near 70 per cent of the banks, is observed during all the years. However, around 16 per cent are observed only in the first or first two years and then were merged and stopped to submit financial information. In general, quite a small fraction of banks appeared or dropped out from the sample after the crisis, but in general we can observe mostly sample attrition because of the mergers. It is possible to argue about existence of the selection problem in this context, as banks that are taken over are mostly inefficient ones and possibly suffered losses in the previous periods. But it is only one side of the problem as this cannot be the only reason of mergers (i.e. banks can merge due to the possible cost savings and economies of scale after the merger) and also some banks are established during the sample period. So I will assume that appearance and dropping of banks from the sample is exogenous and is not dependent on the bank losses. 18 Variable | Mean Std. Dev. Min Max | Observations ---------------------------+--------------------------------------------+---------------- Bankruptcy prob. overall | 2.014316 .9619911 .1475461 8.172152 | N = 1956 (per cent) between | .6454108 .5982204 4.414628 | n = 186 within | .793504 -.4746066 6.862702 | T-bar = 10.5161 | | Loan loss overall | 27957.18 213727.3 -802694.1 5117471 | N = 1956 (mil NOK) between | 113285.9 -2653.735 872844.9 | n = 186 within | 185283.9 -1468845 4451320 | T-bar = 10.5161 | | Ratio loss-assets overall | .0057483 .0167285 -.4049709 .3741996 | N = 1956 between | .0160808 -.0245717 .131978 | n = 186 within | .0153142 -.374651 .4045196 | T-bar = 10.5161 Overall and within deviation is calculated for N bank-years of data. Between deviation is calculated over n banks. The average number of years a bank is observed is 10.5. For example, average bankruptcy probability is 2 per cent with standard deviation 0.962 per cent and it varies between min = 0.148 per cent and max= 8 per cent over the considered 13 years. Average risk indicators for each bank for 13 years have lower standard deviation of 0.645 per cent and lie in a smaller range between 0.598 and 4.4 per cent. Within number show deviation from each bank’s average over time which also explains negative sign for the minimum, but we also need to deduct global means28 and so they vary between -0.475 – 2.014 to 6.863 – 2.014. We also see that a deviation observed within banks over time is higher for risk indicators and loan losses but lower for the loan loss ratio than variation across banks. But we observe high variation in the data both between banks and over the years. Econometric model The analysis of the constructed longitudinal dataset is aimed to investigate whether calculated aggregate risk indicators for banks are significant and can explain, at least to some extent, bank loan losses. The following model is considered: LAit = αi + ABPit-1 β + Mtξ + Sit ρ + εit , i ∈1:N, t ∈ 1:T (1) where N is number of banks, T is the number of periods equal to 12 and disturbances εit are identically and normally distributed with zero mean and constant variance σ2. Variable LA is calculated as a ratio of loan losses to the total bank assets. This variable is more of interest than simply bank loan losses, as the variability of the loan losses can be huge not only due to the risk in lending but also due to the diversification effect related to the bank size and size of the loan portfolio, along with other factors. Variable ABPit-1 is a one period lagged aggregate bankruptcy probability, a risk indicator for a bank’s portfolio of corporate loans. We use lagged values as mostly values realised in the previous period may influence losses of the current period.29 Variable Mt stands for some of the macroeconomic variables (e.g. GDP, unemployment, and housing price index) and variable Sit stands for a vector of time- and bank- 28 We can transform the model as follows: LAit - •i AL = (ABPit-1 - •i PBA )β + (Mt - M)ξ + (Sit - •i S)ρ +( εit - •i ε ), where averages over years are calculated as: T t t ∑ = • i i AL AL . Estimated model have also global means added to each intraindividual difference. LAit - •i AL +AL = α+(ABPit-1 -•i PBA +PBA )β + (Mt - M)ξ +(Sit - •i S+S)ρ + ( εit - •i ε + η )+ ε 29 Presence of the lagged regressors makes it necessary to take into account bank mergers, which were especially widespread during the beginning of 90-s. 19 specific characteristics. Loan losses can take negative values because banks make reversals of previously made loss provisions if they overestimated their size and some of the breached contracts were repaid next period or they value given default happened to be higher than expected. Therefore, we do not use logarithmic form of the equation, which would be useful for log-normal distribution of positive values of loan losses. Due to data construction of variables are predetermined in the model and are assumed to be exogenous and uncorrelated with the disturbance term. 6. Estimation and hypothesis testing Two different banks may invest in the same industry and region but have different investment results due to the diverse credit policies and different client base. A major shortcoming of the constructed aggregate bankruptcy probabilities is that we have to assume the same average credit risk for the banks that have loans to the same industries and region. However, loan losses dependent not only on the size of loans to riskier industries but also on the size of loans provided to more financial fragile enterprises. Therefore we have to use proxies that can help to distinguish banks with respect to their lending policies, i.e. quality of risk management, inclination to take risks and expansion into the new regions and industries, see discussion in section 4. We conducted an estimation of the random effect model, for which individual specific effects ηi are correspondently assumed to be constant or randomly distributed. As αi can be decomposed into a constant and individually variable part, we can rewrite the model as: LAit = α + ABPit β + Mtξ + Sit ρ + ηi + εit , i ∈1:N, t ∈ 1:T, (2) where ηi + εit is a composite error term composed of the genuine disturbance and individual effect part, which is supposed to be randomly distributed and ηi to be drawn from the same probability distribution with IID (0, σ2 α ) and εit is IID (0, σ2 ) as before. We also make a strong assumption of independency of ηi, εit and explanatory variables. So we have non classical gross disturbance due to heteroskedasticity and autocorrelation through the variance of the individual random effect, and thus estimate the model by GLS.30 30 Generalised least squares provides a weighted estimate of β using both within and between variation and assigning a smaller share to the ‘between’ one. This share is smaller when we have larger part of the gross disturbance variance due to the individual random effect. In our case we have almost 1/5 of gross disturbance variance due to the random individual effect. 20 Table 2: Random effect GLS regression Dependent variable Ratio loss/assets Ratio loss/assets Ratio loss provisions/assets Aggregate bankruptcy probability (ABP) 0.002 *** (0.0006) 0.009 *** (0.002) 0.004 *** (0.0006) Difference of the ratio of nonperforming loans to assets 0.392 *** (0.019) 0.306 *** (0.014) Ratio of non-performing loans to assets 0.170 *** (0.007) Unemployment 0.001 *** (0.0004) -0.00005 (0.0001) Ratio equity to assets 0.004 *** (0.001) -0.003 *** (0.0004) Share of risky loans 0.002 (0.003) 0.005 ** (0.003) Interest rate t-1 0.0005 *** (0.0002) Number of regions 0.0008 ** (0.0003) 0.0005 *** (0.0002) 0.00001 *** (0.000) Dummy sb (if saving bank then 1) 0.011 *** (0.003) 0.0005 (0.0013) Dummy sb*ABP -0.007 *** (0.002) -0.003 *** (0.0007) Constant -0.008 *** (0.002) -0.017 *** (0.004) -0.001 (0.001) Breusch and Pagan LM test for RE31: Test: Var(u) = 0 chi2(1)= 18.33 Prob>chi2=0.000 Hausman test: Ho difference in coefficients not systematic32 chi2(6)= 3.88 Prob>chi2=0.794 R2: within 0.279 0.250 0.329 between 0.671 0.436 0.651 overall 0.293 0.289 0.428 Macroeconomic variables are strongly correlated with coefficient of correlation -0.88 for unemployment and GDP, and 0.9 housing index and GDP. As macroeconomic variable we choose unemployment due to the above mentioned valuable properties for explaining loan losses. To reflect bank-specific variables we consider a ratio of non-performing loans to assets, a share of risky loans33, interest rate on loans, a ratio of equity to assets and number of regions in bank loan portfolio. The model provides economically and statistically significant results with the expected coefficient signs but not very high explanatory power. Due to the asymptotic properties of the 31 Breusch-Pagan (1980) Lagrange multiplier test supports the idea of the random effect model. Statistics distributed as χ2 with one degree of freedom, under the null hypothesis of no random effects (i.e. zero variance of the individual specific part of the gross disturbance), is equal to 18.33 and hypothesis can be rejected. 32 Assuming our correctly specified model and uncorrelation of ηi and RHS variables, we check that two models do not give statistically different results. Hausman’s (1980) specification test checks the null hypothesis that difference in coefficients is not systematic and it cannot be rejected with p-value equal to 0.79. Difference between coefficients is statistically insignificant as null hypothesis cannot be rejected (probability of error is 79 per cent), and we can use random effects estimator for our model. 33 Defined as a share of loans to non-financial firms with bankruptcy probabilities higher than three per cent to total loans in the bank’s loan portfolio. 21 random effect estimator, Wald statistics confirm presence of significant regression on the 95 per cent significance level. Both unemployment and non-performing loans are found to have a positive effect on bank loan losses. In addition ratio of equity to assets and share of risky loans also have positive influence reflecting adverse incentives arising from higher capital buffer. The number of regions has a statistically significant positive coefficient, suggesting that larger expansion increases risks and creates adverse effect for the loan losses. In addition to evaluating statistical significance of the sign of the coefficients it is useful to check the plausibility of the size of the obtained effects. Non-performing loans have highest effect on loan losses as an increase on 0.01 in the ratio of non-performing loans to assets with the average value of 0.023 leads to a 0.033 percentage points increase in the ratio of loan losses from 0.0057 to 0.0087 on average. Aggregate bankruptcy probabilities have less apparent but still quite large and statistically significant result. An increase on a 0.1 percent from the average value of 2 per cent leads to an increase from 0.0057 to 0.0066 in the ratio of loan losses to assets, which is around 15 per cent increase compared to the average value of this ratio. A dynamic specification of the model was also estimated as losses in one period can be driven by the previous periods losses, which, for instance, can capture prevalence of banks’ inefficient policy in assessment of the borrowers’ credit risks or/and reflect influence of the banks’ financial conditions on losses through the past performance.34 Meaningful and statistically significant results for the aggregate bankruptcy probabilities are robust to the dynamic specification of the model. 34 Here we have to deal with the endogeniety problem, as explanatory variables are correlated with the disturbance term (i.e. violation of the weak exogeniety). 22 Table 3. Dynamic model GMM estimation (robust to heteroskedasticity) Dependent variable Ratio loan losses/assets Ratio loan losses/assets Ratio provisions/assets Ratio loan losses to assets t-1 -0.524 *** (0.144) -0.467 *** (0.142) Ratio loan losses provisions to assets t-1 0.704 *** (0.087) Aggregate bankruptcy probability (ABP) 0.002 ** (0.0008) 0.002 * (0.0009) 0.0002 (0.0006) Aggregate bankruptcy probability (ABPt-1) 0.003 *** (0.0008) 0.004 *** (0.001) 0.0006 (0.0006) Aggregate bankruptcy probability (ABPt-2) -0.001 * (0.0006) Ratio non-performing loans to assets 0.303 * (0.178) 0.292 (0.187) 0.031 (0.013) Equity/assets 0.022 (0.083) 0.027 (0.079) 0.033 *** (0.008) Equity/ assets t-1 0.475 ** (0.239) 0.482 * (0.247) 0.075 *** (0.027) Unemployment 0.001 ** (0.0006) 0.0008 *** (0.0003) Real loan interest rate (RIL) 0.003 *** (0.0009) 0.003 *** (0.0008) 0.0003 *** (0.000) RIL t-1 0.003 *** (0.0008) 0.002 *** (0.0006) 0.0002 * (0.000) RIL t-2 0.0009 *** (0.0003) Share of risky loans 0.006 (0.004) Share of risky loans t-1 0.005 * (0.003) Number of regions 0.0005 * (0.0002) 0.0003 (0.0003) -0.0001 (.0001) Number of regions t-1 0.0004 (0.0003) 0.0005 * (0.0003) 0.00007 (0.000) Constant -0.0007 *** (0.0003) -0.0003 (0.0002) 0.0002 * (0.000) A-B test: for zero 1-order aurocovariance in residuals z = -2.28 Pr > z = 0.0228 z = -2.30 Pr > z = 0.0215 z = -2.03 Pr > z = 0.0423 A-B test: for zero 2-order aurocovariance in residuals z = 0.41 Pr > z = 0.6821 z = 0.72 Pr > z = 0.4721 z = 0.21 Pr > z = 0.8302 7. Conclusions We found that aggregate bankruptcy probability as a proxy for risks in lending can explain bank loan losses. This means that banks with higher bankruptcy probabilities of their loan portfolios tend to have higher loan losses if we control for the general macroeconomic conditions and bank specific factors. However for a given phase of the economic development, banks with more efficient credit risk management may be able to control risks more efficiently and reduce possible loan losses. Agenda for the future research contains a possibility of testing of the following hypothesis: 29 Manufacture of other engineering products (code 469), Business services in 1988 (code 840), Recreational, cultural and sporting activities (code 920) in 1996 and 2000 and Air transport (code 620) over 1996-2000 in county 21, enterprises in Manufacture of wearing apparel, dressing and dyeing of fur (code 180) in 1996, Manufacture of pulp, paper and paper products (code 210) over years 1996-1999, Manufacture of other electrical machinery and devices (code 310) in 1998 and Manufacture of medical, precision and optical instruments, watches and clocks (code 330) over 1996-2000 in the county 20, enterprises in Conversion of wood (code 320) over 1988-1994 in counties 19 and 20 and in Recycling (code 370) over 1996-2000 in county 19. For the corresponding years, there are no enterprises registered in the corresponding industries and counties, and submitting financial statements according to the SEBRA-database. At the same time, a closer accord between the two statistics is obtained for the industry/year groups. For all groups constructed using the bank statistics, relevant aggregated bankruptcy probabilities can be calculated directly form the SEBRA-database after establishing a correspondence between industry codes in the two statistics. A minor exception is industry Public administration (code 750), to which loans are not registered in the bank statistics in 1996 and 1997, but some enterprises in this industry reported ‘debt in the financial institutions’ during the same years. However, a correct connection between groups made for enterprises with debt in financial institutions and groups of loans obtained from the banking statistics cannot be checked without full information on all financial institutions (Note: we consider only bank loans). Therefore, industry/year groups can be directly used for linking the bank and enterprise statistics, while industry/county/year groups require additional transformations. Information from financial statements also involves some problems. For example, in 17,835 cases out of all submitted financial statements for the period 1988 - 2000, assets do not match with the liabilities. For 296 observations assets are negative and are equal to the liabilities. Negative assets appear mainly due to the posts: cash, debtors, or investment in financial assets; however it is not the only source of problems. Negative liabilities show up mainly because of negative equity on the liability side of the balance sheet. Moreover, the same problem arises due to negative debt (4,452 observations), while it should be posted as assets, for example, as cash or deposits. Negative debt arises mainly due to bank overdrafts (short-term loans in financial institutions), but long-term negative liabilities are also present in the data. Negative operating revenue appears 625 times in the sample. However, it is not possible to find a concrete source of mistakes and to correct the figures without having more specific information about enterprises than accounting data contained in the SEBRA-database. In some cases, values of aggregated posts are not equal to the sum of the detailed posts. For example, the value of total assets cannot be obtained from the sum of fixed and current assets, and values of total liabilities cannot be obtained from the sum of equity and liabilities. However, using the calculated instead of the given numbers for total assets and total liabilities leads to a discrepancy between assets and liabilities in a large number of cases. Another property of the data is that enterprises with recorded negative or zero assets have extremely low predicted probabilities of bankruptcy, almost all belonging to the range with p < 0.01, where p is the probability of bankruptcy. Those with negative operating revenue have more reasonable values of risk measures but most of them still belong to the range with lowest probabilities, i.e. p < 0.01. Therefore, bankruptcy probabilities estimated on these records are not trustworthy. Estimation of bankruptcy probabilities on the sample of enterprises with book value of total assets larger than 300,000 NOK still includes some records with negative revenues. The model was re-estimated setting them equal to zero; however, it did not 30 significantly affect the results as these records constitute a small share of total observations. However, records with negative asset values should be excluded as predicted bankruptcy probabilities are not reliable and will introduce errors into the aggregated values. Average bankruptcy probabilities in different industries show diverse patterns over the years 1988-2000. Most of them are decreasing, while some have slightly rising tendency over the last years, e.g. Forestry (code 021), Manufacture of pulp and paper (code 210), Manufacture of coke, refined petroleum products and nuclear fuel (code 230), Transport via pipelines (code 603) and Manufacture of tobacco products (code 160). Some had this tendency only till the middle of the sample period, with a following decrease after 1994 for Mining of coal and lignite (code 100) and after 1996 for Recycling (code 370). However, data on these industries, except 021 and 210 contains quite small number of observations, and results may be influenced by extreme values. Implausible values of the average bankruptcy probabilities are also found in the following industries: Mining of coal and lignite, Collection, purification and distribution of water, Private households with employed persons, Post and telecommunication. Mining of coal and lignite 0 0,01 0,02 0,03 1988 1990 1992 1994 1996 1998 2000 Collection, purification and distribution of water 0 0,01 0,02 0,03 0,04 0,05 1988 1990 1992 1994 1996 1998 2000 Private households with employed persons 0 0.02 0.04 0.06 0.08 1990 1992 1994 1996 1998 2000 Post and telecommunication 0,02 0,04 0,06 0,08 1988 1990 1992 1994 1996 1998 2000 High average bankruptcy probabilities with large deviations, i.e. mean value larger than 0.036 and standard deviation larger than 0.065, which corresponds to the upper 25 per cent, are found in many industries especially during 1990-1993. Values in the upper 20-10 per cent correspond are found in the following industries: Agriculture and forestry (code 111) in 1990-1992, Raising of fish (code 051) in 1988-1993, Fishery (code 052) in 1988-1993, Publishing, printing and reproduction of recorded media (code 220) in 1988-1993, Manufacture of office machinery and computers (code 300) in 1996-1999, Hotels and restaurants (code 550) in the whole sample 1988-2000, Construction (code 450) in 1989-1993, also industries 461, 462, 469 and 490 in 1990-1992, industry Wholesale and commercial agency (code 610), Trade retailing (code 621) in 1990-1991, Post and telecommunication (code 640) in 1990-1991 and 1996-1998, Recreation, cultural and sporting activities (code 920) in 1996-1999, Other service activities (code 930) in 1996-2000. These industries traditionally have high uncertainty in their activities; this is particularly true for hotel, restaurants, recreation and service activities. However, industry Real estate activities (code 700), which is also considered risky, show quite stable and low values of bankruptcy probabilities throughout the sample period with mean values decreasing from 0.013-0.016 to 0.007. 31 Data transformations: ¾ Drop observations with missing and zero industry code or zero county number (for years 1988-2001, 8124 and 262 respectively observations were deleted). ¾ Drop observations with industry codes that do not correspond with aggregate codes in the bank statistics: - Codes 65000 - 68000 (financial operations and insurance), (for years 1988-2001, 49793 observations were deleted); - Code 99000 (international organisations), (for years 1988-2001, 1450 observations were deleted); ¾ Drop observations with industry codes 75000 – 76000 before year 1996, because they do not correspond to the aggregate industry codes from the old classification (aggregate industry 190 according to the new classification). ¾ Drop observations with industry codes that do not correspond with SIC94 and cannot logically be added to one of the existing groups: - Code 38399 (contains records on one enterprise with average assets around 1.200.000 NOK, average revenue and debt around 800.000 NOK for the years 1992-1997, the highest value of debt was 2.146.000 NOK in 1995), (for years 1988-2001, 6 observations were deleted). - Code 83299 (contains records on one enterprise with average assets around 170.000 NOK, average revenue around 650.000 NOK for the years 1994-1997, the highest value of debt was 13.000 NOK in 1996, and on one enterprise with average assets around 1.200.000 NOK, average revenue around 2.000.000 NOK and average debt around 280.000 NOK over the years 1990-1996, the highest value of debt was 988.000 NOK in 1994), (for years 1988-2001, 11 observations were deleted). - Code 88888 (contains records on one enterprise with 1.845.000 NOK assets, zero revenue and 1.674.000 NOK debt in 2000), (for years 1988-2001, 2 observations were deleted). ¾ Classify observations with industry codes that do not correspond with SIC94 due to higher precision level and assign them codes from upper designation correspondent to SIC94. For the years 1988-2000: - 53 observations with code 01222 are added to the code 01220. Added records constitute 62 % of total observations (01222 and 01220 together), and 67 % in terms of extended loans; - 990 observations with code 01411 are added to the code 01410. Added records constitute 73% of total observations and 82% in terms of extended loans; - 9 observations with code 11111 are added to the code 11100. Added records constitute 0.9% of total observations and almost 0% in terms of extended loans; - 266 observations with code 20511 are added to the code 20510. Added records constitute 36% of total observations and 57% in terms of extended loans; - 3677 observations with code 28751 are added to the code 28750. Added records constitute 58% of total observations and 69% in terms of extended loans; - 1 observation with code 29012 are added to the code 28750. Added records constitute almost 0% in terms of observations and extended loans; - 173 observations with code 31630 are added to the code 31620. Added records constitute 84% of total observations and 95% in terms of extended loans; - 454 and 40 observations with codes 45001 and 45002 respectively are added to the code 45000. Added records constitute 51% of total observations and 6% in terms of extended loans; - 7902 observations with code 45111 are added to the code 45110. Added records constitute 81% of total observations and 86% in terms of extended loans; - 89 and 2173 observations with codes 45251 and 45252 are added to the code 45250. Added records constitute 47% of total observations and 23% in terms of extended loans; - 54 observations with code 51331 are added to the code 51330. Added records constitute 24% of total observations and 1.3% in terms of extended loans; 32 - 400 observations with code 51411 are added to the code 51410. Added records constitute 24% of total observations and 6% in terms of extended loans; - 1256 and 1 observation with codes 51435 and 51439 respectively are added to the code 51430. Added records constitute 55% of total observations and 43% in terms of extended loans; - 307 observations with code 51443 are added to the code 51440. Added records constitute 95% of total observations and 99% in terms of extended loans; - 33 observations with code 51451 are added to the code 51450. Added records constitute 2.5% of total observations and 0.2% in terms of extended loans; - 193 observations with code 51461 are added to the code 51460. Added records constitute 7% of total observations and 0.6% in terms of extended loans; - 83 observations with code 51521 are added to the code 51520. Added records constitute 5% of total observations and 6% in terms of extended loans; - 73 and 2 observations with codes 51552 and 51551 respectively are added to the code 51550. Added records constitute 4% of total observations and 1% in terms of extended loans; - 70 and 391 observations with codes 51643 and 51641 respectively are added to the code 51640. Added records constitute 11% of total observations and 26% in terms of extended loans; - 377 total observations with code 51701-51703 are added to the code 51700. Added records constitute 8.5% of total observations and 20% in terms of extended loans; - 1127 total observations with code 51810, 51820, 51830, 51840, 51850, 51860, 51860, 51870, 51880 are added to the code 51000. Added records constitute 13% of total observations and 27% in terms of extended loans; - 154 observations with code 52101 are added to the code 52100. Added records constitute 21.5% of total observations and 37% in terms of extended loans; - 13812 total observations with code 52111-52113 are added to the code 52110. Added records constitute 47% of total observations and 62% in terms of extended loans; - 28 observations with code 52221 are added to the code 512220. Added records constitute 2.5% of total observations and 1% in terms of extended loans; - 328 observations with code 52601 are added to the code 52600. Added records constitute 97% of total observations and 99.5% in terms of extended loans; - 154 observations with code 52741 are added to the code 52740. Added records constitute 61% of total observations and 80% in terms of extended loans; - 337 observations with code 55401 are added to the code 55400. Added records constitute 24% of total observations and 29% in terms of extended loans; - 8165 observations with code 61001 are added to the code 61000. Added records constitute 89% of total observations and 91% in terms of extended loans; - 2 observations with code 63202 are added to the code 63200. Added records constitute 2% of total observations and almost 0% in terms of extended loans; - 107 observations with code 64201 are added to the code 64200. Added records constitute 7.5% of total observations and almost 0% in terms of extended loans; - 2 observations with code 63202 are added to the code 63200. Added records constitute 2% of total observations and almost 0% in terms of extended loans; - 363 and 19 observations with codes 71402 and 71401 respectively are added to the code 71400. Added records constitute 18% of total observations and 10% in terms of extended loans; - 3 observations with code 71911 are added to the code 71000. Added records constitute 2.6% of total observations and almost 0% in terms of extended loans; - 245 observations with code 72301 are added to the code 72300. Added records constitute 7.5% of total observations and 1% in terms of extended loans; - 13395 total observations with code 74401-74409 are added to the code 74400. Added records constitute 71% of total observations and 66% in terms of extended loans; - 300 observations with code 74601 are added to the code 74600. Added records constitute 20% of total observations and 59% in terms of extended loans; - 61 observations with code 74811 are added to the code 74810. Added records constitute 2% of total observations and 10% in terms of extended loans; - 5147 and 350 observations with codes 74832 and 74831 respectively are added to the code 74830. Added records constitute 96% of total observations and 99% in terms of extended loans; 33 - 369 observations with code 74841 are added to the code 74840. Added records constitute 1.6% of total observations and 8% in terms of extended loans; - 61 observations with code 74811 are added to the code 74810. Added records constitute 2% of total observations and 10% in terms of extended loans; - 54 and 37 observations with codes 91331 and 91332 respectively are added to the code 91330. Added records constitute 3% of total observations and 3.5% in terms of extended loans; - 513 observations with code 93012 are added to the code 93010. Added records constitute 22% of total observations and 4% in terms of extended loans; - 5516 total observations with codes 93021-93024 are added to the code 93020. Added records constitute 97% of total observations and 96.5% in terms of extended loans; - 1445 and 759 observations with codes 93041 and 93042 respectively are added to the code 93040. Added records constitute 88% of total observations and 95% in terms of extended loans; In some cases added observations constitute a large part of the newly obtained groups both in terms of the number of observations and amount of loans. However, the SEBRA-database contains many enterprises, with industry codes not included in SIC94, that cannot be omitted. Aggregation of the bankruptcy probabilities by industry and county Using industry and industry/county dimension, common for the banks’ Reports and the SEBRA-database, we aggregate bankruptcy probabilities for each enterprise from the SEBRA- model by industry or by industry/county. Volumes of debt in the financial institutions or the levels of activities, represented, for example, by total assets or operating revenues, are proposed as weights in aggregation. It is reasonable to concentrate on the enterprises with nonzero post ‘debt in financial institutions’ since only these enterprises will inflict a loss for the bank in the event of bankruptcy. However, aggregation using only debt as weights may cause some biases because the SEBRA-database contains post ‘debt in financial institutions’, which has a wider meaning than debt in banks. Therefore, some of the selected enterprises will still be irrelevant for the calculation of the bank’s risk on the loan portfolio. Moreover, banks have also loans to some industry/year groups that are not reflected in the SEBRA-database as groups of enterprises with debt in financial institutions’. For example, the SEBRA-database does not contain enterprises with debt in financial institutions in industry Transport via pipelines (code 603) in 1997 and in industry Private households with employed persons (code 950) in 1996, 1998 and 2000. Enterprises with debt, which belong to a particular industry/county/year group in the SEBRA-database, do not always correspond to those enterprises that banks have actually given loans in this group, also due to the discrepancies in the county classification. However, due to the discussed shortcomings of the data it is also problematic to choose one of the possible activity measures, i.e. total assets or operating revenue. Both of them include negative observations, also for enterprises with bank loans. Therefore, we use the level of debt in the financial institutions as weights in aggregation.38 38 A composite measure taking into account both characteristics can also be relevant because of the high correlation between them: corr(total assets, debt) = 0.9, corr(revenue, debt) = 0.7, corr(total assets, revenue) = 0.8. For example, we can assign half weight to the debt size and half weight to the measure of the activity size, or even use both activity measures as enterprises with negative total assets may have operating revenue and vice versa. 34 Linking of the bank statistics and the SEBRA-database Assigning new aggregate industry codes Assignment of the aggregate industry codes, valid in the bank statistics from 1996, to enterprises in the SEBRA-database is made for the years 1996 - 2001 according to the relationship pattern between SIC94 and aggregate codes in the Report 60. Observations from the SEBRA-database with industry codes that do not have a direct correspondence with aggregate codes are subdivided as follows: ¾ 143 observations from industry Fishing, operation of fish hatcheries and fish farms (code 5000) are randomly equally divided between the aggregate industries Fishing and operation of fish hatcheries (code 051) and Fish farms (code 052). These randomly added observations constitute 1.48 and 0.63 per cent respectively of the total number of observations in these industries. ¾ All 2071 observations from industry Land transport, transport via pipelines (code 60000) are included in the aggregate industry Land transport (code 601), and constitute 7.55 per cent of the total number of observations in this industry. (Note: Industry Transport via pipelines (code 603) has only few enterprises with a total of 33 observations for the whole sample period.) ¾ 1046 observations from industry Water transport (code 61000) are randomly divided between the aggregate industries Foreign water transport (code 611) and Inland water transport (code 612). These randomly added observations constitute around 2 and 19 per cent respectively of the observations in these industries. Aggregate industries with codes 051 and 052 are joined before 1991, because during this period industry 053 contains information on both 051 and 052. It is also suggested to exclude industry Private households with employed persons (code 950), which does not contain sufficient information. It has around 2-5 observations each year, of which in general one enterprise has debt. The total sum of debt for this industry from 1988 to 2000 is 619,000 NOK, total revenue is 106,139,000 NOK. Assigning old aggregate industry codes The SEBRA-database contains only industry codes consistent with SIC94 becasue they were previously converted from SIC83 for all enterprises, while bank statistics uses old aggregate classification of industries in Reports 60 up to 1996. Therefore, for the data before 1996, we should assign old aggregate codes to enterprises using relationship patterns between old aggregate codes and SIC83, and between SIC83 and SIC94. This correspondence is not one-to- one, i.e. a particular industry code in SIC94 can correspond to several different codes in SIC83 due to the different types of the classification applied, and hence to several aggregate codes. Since we do not possess detail information about sphere of activities of the individual enterprises, distribution of some observations can be only done randomly. Therefore, some groups of observations with assigned old aggregate codes contain an arbitrary part. How these random observations may influence our data set is discussed below. The following procedure is applied when defining the old and new aggregate codes: 1. Each observation with industry code that has a one-to-one correspondence with SIC83 is attributed to a correspondent aggregate industry; 2. Observations with industry code that corresponds to more than one industry in SIC83 are randomly divided between relevant aggregate industries; 35 3. Observations that do not have detailed industry codes consistent with the relationship pattern between SIC94 and aggregate codes are randomly divided between corresponding aggregate industries. Consequently, they will increase the number of randomly distributed observations. After the first examination, it turned out that some of the industries contain a very large random part, i.e. close to 100%. To avoid this, industries Transport and storage (code 911) and Foreign water transport (code 711) were united under the one industry 911 due to the similarities in their main activities. Aggregate industries Extraction of oil and gas (code 211), Financial operations relevant to extraction of oil and gas (code 231) and Drilling for oil on contact base (code 721) were joined in one industry 211 because they all have activities relating to oil and gas. Then following old aggregate tree-digit industry codes were assigned to the observations with consistent with SIC94 five-digit numbers, using the correspondence between SIC94 and SIC83, and between SIC83 and aggregate codes: - >=01000&<01400, 01420, >=02000&<02020 correspond to 111; - A half of observations with >=01400&<01420, 01500, 02020 was randomly selected for 111; - >=05010&<05020 correspond to 051; - A half of observations with 01500, 05000 was randomly selected for 051; - >=05020&<10000 correspond to 052; - A half of observations with 5000 was randomly selected for 052; - >=11000&<13000 correspond to 211; - A half of observations with 74203 was randomly selected for 211; - 60300 corresponds to 221; - >=10000&<11000, >=13000&<15000 correspond to 311; - >=21000&<21230, >=21240&<22000 correspond to 321; - A half of observations with 20200 and 21230 was randomly selected for 321; - >=23000&<24139, >=24140&<25130, >=25200&<25240 correspond to 331; - A half of observations with 24139, 26820, 29600, 35116, 35120, 36400, one third of observations with >=36100&<36200, 19300, 36500 and one fourth of 36630 were randomly selected for 331; - Also 80 per cent of the observations in 25130 and 25240 was randomly selected for 331 (Note: 20% of observations in the industries 25130 and 25240 was allocated to 420 and the rest 80% to 331 because of the small extend of the textile industry); - >=27000&<27210, >=27300&<27340, >=27350&<28000, 28400, 28500, 28520, 29140 correspond to 341; - A half of observations with 24139, 28000, 29229, 34300, 27210, 27220 was randomly selected for 341; - A one third of observations with 28700, 28750, 28751, 29200 and 29220 was randomly selected for 341; - >=15000&<17000 correspond to 411; - A half of observations with 55500 and 55520 was randomly selected for 411; - >=17000&<19300 correspond to 420; - A half of observations with 20520, one third of observations with 19300, one fourth of observations with 36630 and one fifth with were randomly selected for 420; - >=20000&<20200 and >=20300&<20520 correspond to 430; - A half of observations with 20200, 20520, 36000, 36150, 32300 was randomly selected for 430; - A one third of observations with 19300, 31500, >=36100&<36150 was randomly selected for 430; - >=22000&<22140 and >=22150&<22300 correspond to 440; - >=26000&<26820 correspond to 451; - A half of observations with 26820 and a one third of observations with 31500 were randomly selected for 451; 36 - >=35114& <35116 correspond to 461; - A half of observations with 45212 was randomly selected for 461; - 29111, >=35100&<35114 and 35117 correspond to 462; - A half of observations with 29110, 29221, 35000, 29200, 35116, 35120 and a one third with 29220 was randomly selected for 462; - 27340, >=28100&<28400, 28510, >=28600&<28700, >=28710&<28750, >=29000&<29110, >=28119&<29140, 28210, >=28230&<29600, >=28710&<31500, >=31600&<32300, 33000, >=32200&<34300, >=35201&fn<36000 correspond to 469; - A half of observations with 27210, 27220, 45300, 45310, 45330, 51570, 72500, 28000, 29110, 29221, 29229, 29600, 32300, 33100, 34300, 35000 , a one third of observations with 28700, 28750, 28751, 29200, 29220, 31500, >=36100&<36150, 36500 and one fourth with 36630 were randomly selected for 469; - >=36200&<36400 and >=36600&<36630 correspond to 490; - A half of observations with 21230, 36000, 36400, one third with 28700, 28750, 28751, 36500 na one fourth with 36630 were randomly selected for 490; - >=45000&<45120, >=45200&<45212, >=45220&<45300, 45320, >=45340&<50100 correspond to 450; - A half of observations with >=1400&<1420, 2020, 45212, 45300, 45310, 45330 was randomly selected for 450; - >=40000&<45000 correspond to 540; - >=37000&<40000, 50101, 50301, 50401, 52464, 52485, >=51000&<51570, >=51600&<52000 correspond to 610; - A half of observations with 52460, 52461, 52469, 50100, 50300, 51570 and a one third with 50400 were randomly selected for 610; - 50102, 50302, 50402, 50500, >=52000&<52460, 52462, 52463, >=52470&<52485, >=52486&<52700 correspond to 621; - A half of observations with 50100, 50300, 50200, 52460, 52461, 52469, a one third with 50400, >=71400&<72000 and a one fourth with 71000 were randomly selected for 621; - >=55000&<55500, 55510 correspond to 550; - A half of observations with 55500 and 55520 was randomly selected for 550; - >=70000&<71000 correspond to 700; - 22330, 45120, 71220, >=71300&<71400, >=72000&<72500, 72600, >=74000&<74203, 74209, >=74400&<74700, 74800, >=74820&<75000 correspond to 840; - A half of observations with 22300, 92400, 71200, 71230, 72500, 74203, a one third with 74300 and a one fourth with 71000 were randomly selected for 840; - >=60000&<60300, >=61000&<64100, 71100, 71210 correspond to 911; - A half of observations with 64100, 64120, 71200, 71230, a one third with 74300 and a one fourth with 71000 were randomly selected for 911; - 64110, 64200 and 64201 correspond to 640; - A half of observations with 64100 and 64120 was randomly selected for 640; - 22140, 22310, 22320, 50403, >=52700&<55000, >=73000&<74000, 74700, 74810, 74811, >=80000&<92400, >=92500&<99000 correspond to 931; - A half of observations with 22300, 92400, 33100, 50200, >=71400&<72000, a one third with 50400, 74300 and a one fourth with 71000 were randomly selected for 640; We analyse also the effect of the randomly added observations for the distribution of the bankruptcy probabilities (p) for the aggregate industries. The following graphs show the means and standard deviations for p’s before and after the randomly selected observations were added, as well as for these random observations themselves. Variables m_random, m_main and m_whole are respectively mean values of the bankruptcy probabilities in the arbitrary added part, and mean values before and after this part is added over years. Variables r-/r+, m-/m+ and w-/w+ show respectively standard deviations in bankruptcy probabilities in the arbitrary added part, before and after this part is added. 37 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 1988 1989 1990 1991 1992 1993 1994 1995 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 w + w - m_whole m_main r - m - m_random m + r + Wood coversion -0.050 -0.030 -0.010 0.010 0.030 0.050 0.070 0.090 0.110 0.130 1988 1989 1990 1991 1992 1993 1994 1995 -0.050 -0.030 -0.010 0.010 0.030 0.050 0.070 0.090 0.110 0.130 w + w - m_whole m_main r - m - m_random m + r + Manufacture of chemical products -0.060 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 0.140 1988 1989 1990 1991 1992 1993 1994 1995 -0.060 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 0.140 w + w - m_whole m_main r - m - m_random m + r + Manufacture of metals -0.070 -0.020 0.030 0.080 0.130 0.180 1988 1989 1990 1991 1992 1993 1994 1995 -0.070 -0.020 0.030 0.080 0.130 0.180 w + w - m_whole m_main r - m - m_random m + r + Manufacture of oil rigs 38 -0.050 0.000 0.050 0.100 0.150 19 8 8 19 8 9 19 9 0 19 9 1 19 9 2 19 9 3 19 9 4 19 9 5 -0.050 0.000 0.050 0.100 0.150 w + w - m_w hole m_main r - m - m_random m + r + Other industry production -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 1988 1989 1990 1991 1992 1993 1994 1995 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 w + w - m_whole m_main r - m - m_random m + r + Other engineering production -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 1988 1989 1990 1991 1992 1993 1994 1995 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 w + w - m_w hole m_main r - m - m_random m + r + Construction of crafts and boats -0.060 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 1988 1989 1990 1991 1992 1993 1994 1995 -0.060 -0.040 -0.020 0.000 0.020 0.040 0.060 0.080 0.100 0.120 w + w - m_w hole m_main r - m - m_random m + r + Manufacture of food, beverages and tobacco Olga Andreeva: Aggregate bankruptcy probabilities and their role in explaining banks’ loan losses Working Paper 2004/2 KEYWORDS: Bank losses Bankruptcy probabilities Aggregation - 17627