Does Pooling of Financial Statements and Default Data across Specialized Banks Improve Internal Credit Rating Systems?

Frerichs, Hergen Article Does Pooling of Financial Statements and Default Data across Specialized Banks Improve Internal Credit Rating Systems? Kredit und Kapital Provided in Cooperation with: Duncker & Humblot, Berlin Suggested Citation: Frerichs, Hergen (2005) : Does Pooling of Financial Statements and Default Data across Specialized Banks Improve Internal Credit Rating Systems?, Kredit und Kapital, ISSN 1865-5734, Duncker & Humblot, Berlin, Vol. 38, Iss. 3, pp. 401-433, https://doi.org/10.3790/ccm.38.3.401 This Version is available at: https://hdl.handle.net/10419/293530 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/4.0/ Kredit und Kapital, 38. Jahrgang, Heft 3 Seiten 401-433 Does Pooling of Financial Statements and Default Data across Specialized Banks Improve Internal Credit Rating Systems? By Hergen Frerichs, Frankfurt/M.* I. Introduction Under the new Basel capital accord (Basel II)1, banks will have the opportunity to estimate borrowers' default probabilities for regulatory capital calculation. Bank regulators will have to decide on the eligibility of these internal credit rating systems for the accord's internal ratings based approach (IRB approach). The European Union plans to oblige all banks in its jurisdiction to abide by Basel II. For many European banks, it will also be worthwhile to implement the IRB approach. Under the standard approach most European corporate customers would obtain a risk weight of 100% as they are not externally rated. For many of these customers, risk weights might be reduced if banks derive default probabilities internally. Yet, particularly for small banks, it is costly to implement the IRB approach. In addition, small banks are not likely to be able to calibrate an internal credit rating system of high quality due to their small databases. For these banks it might make sense to share their databases with other banks to save costs and to improve system quality. In Germany, both savings banks and cooperative banks have started respective projects. Coordination in these bank groups is particularly easy as banks generally do not compete with each other within a group. Each bank of the group is largely restricted to doing business with borrowers located in a prespecified region. Coordination becomes more difficult if banks compete with each other because the content of a bank's credit database is at least partly proprietary and constitutes part of a bank's competitive advantage. Yet, the improvement in system quality * This paper is a result of a research cooperation with Deutsche Bundesbank. i Basel Committee (2003). Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 402 Hergen Frerichs due to data pooling might outweigh the costs of coordination. Especially, when thinking about small banks, there ought to exist sufficiently many small banks in an economy that are not direct competitors. Moreover, the problem of releasing proprietary data to competitors can be handled by creating a neutral third party that collects data such that confidentiality is assured. In a very general sense, pooling always dominates non-pooling if nonpooling is understood to be a subset of pooling. If the result of a pooling project is that it is optimal for each bank to stick to its own internal credit rating systems, and if this is regarded as a pooling solution, then pooling will always be better than non-pooling.2 In this paper, pooling is defined in a narrower sense. A pooled credit rating system is defined to be a system that is derived from a pooled data set disregarding certain regional and sectoral attributes of this data set. This definition more closely captures the spirit of current pooling projects. For the segment of medium-sized companies, the group of German savings banks and that of cooperative banks are likely to design just one credit scoring function for the entire group. While there might be a further segmentation into producing and trading companies, it is quite unlikely that there will be any further regional or sectoral segmentations. Banks will be interested to keep the overall number of rating systems (e.g. for retail customers, for residential and commercial real estate, for small businesses etc.) small due to large administration, backtesting and updating expenses. Internal credit rating systems will be submitted to the bank regulator who has to decide on their admission to the IRB approach. As these systems will be based on a vast amount of data, they can be expected to be of high quality. Yet, when analyzing these systems, regulators do not only have to evaluate the overall quality of a system, but also its quality with respect to each participating bank of the group. It might well be possible that the savings banks' system is appropriate for the sector as a whole, but it might be of inadequate quality for savings banks in east Germany, for example, if regional differences are not considered by the system. The same reasoning can be applied to a large commercial bank such as Deutsche Bank. If Deutsche Bank calibrates a system for all German corporate borrowers, it might not be adequate for every regional or sectoral subportfolio. Therefore, the research question posed in the title is important for both banks and regulators. As it is very difficult to obtain access to credit 2 Thanks to an anonymous referee for raising this point. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 403 rating data for a large number of banks, a simulation approach is applied based on financial statements and default data from Deutsche Bundesbank.3 In each simulation step, a set of financial statements and default data is randomly drawn from the Deutsche Bundesbank database, which forms the credit portfolio of one particular bank. This bank may either use its own data, or it may pool its data with other banks to calibrate its internal credit rating system.4 Based on the bank's own portfolio, rating system quality is evaluated for both options using either the area-under- curve (AUC) or the Brier score.5 Over many simulation steps, it is investigated which of the two options is superior. Simulations are done for different bank sizes and different forms of specialization (regional, sectoral, sectoral-regional). The study is based on Deutsche Bundesbank's annual accounts database, which is the most comprehensive collection of annual accounts of German non-financial companies. While the major strength of the database certainly is its size, some weaknesses are: 1. The database is somewhat biased towards large public limited west German manufacturing companies due to its rediscount business origin.6 2. Companies primarily submit annual accounts based on tax law to Deutsche Bundesbank, which are characterized by compilation periods of up to one year decreasing the database's value for default prediction. 3. Default is defined as the formal initiation of insolvency proceedings, which is a narrower definition than the Basel II definition7 causing relatively low default rates in the database. 3 The general approach is taken from Frerichs and Wahrenburg (2003). See there for other related literature. Frerichs and Wahrenburg (2003) treat the general question of evaluating internal credit rating systems, but do not at all address the effects of specialization. 4 Internal credit rating systems are defined to consist of eight rating classes of equal size. Borrowers are classified into rating classes according to their logistic regression credit score. Pooled systems are calibrated based on the complete Deutsche Bundesbank database or on large regional or sectoral subsamples. 5 The AUC is used in Sobehart et al. (2000), Blochwitz et al. (2000), Engelmann et al. (2003). The Brier score is used in Diebold and Rudebusch (1989), Winkler (1994), Lopez (1999, 2001). 6 Deutsche Bundesbank (1998). 7 Basel Committee (2004), § 452. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 404 Hergen Frerichs We use data of medium-sized and large customers according to German corporate law8 from 1994-1999, forming a 1994-1998 training sample and a 1999 validation sample. The training sample consists of 98,910 observations of 29,607 companies with an average default rate of 0.58%. The validation sample consists of 18,671 observations (= companies) with an average default rate of 0.74%. When discussing the question whether pooling is worthwhile or not, portfolio size and particularly a larger number of defaults is the most important argument in favour of pooling.9 If banks are not specialized in any way, pooling would generally be expected to improve system quality. Yet, as banks specialize in certain fields of business, the information contained in the overall economy might become less valuable. Therefore, specialization effects might lead to a situation in which pooling is not beneficial any more. The major result of the study is that out-of-time pooling is beneficial for small and medium-sized credit portfolios (375 and 750 borrowers out-of-time) in most cases. This result holds for regionally and sectorally specialized banks, and also for banks that specialize both in a sector and a region. There are quite a few examples, where an increasing degree of specialization does not lead to a smaller benefit from pooling, but to a larger one. This means that specialization sometimes does not cause a specialization advantage, but a disadvantage. There seems to be information in other sectors, in which a bank does not specialize, that helps predicting credit risks in the sector, the bank is specialized in. Overall, differences in system quality are not large, which means, that in most cases we are not able to state rigorously that pooling significantly outperforms non-pooling. Yet, as expected, we observe that the benefits of pooling increase with decreasing credit portfolio size. There are two points that might partly explain the out-of-time dominance of pooling for small and medium-sized portfolios. The first point is, that the pooled credit scoring functions are derived either on the complete Deutsche Bundesbank dataset or on some large subportfolio. This reflects the effort of German savings and cooperative banks that actually pool their data on this scale. If only two or three banks decide to pool 8 § 267 HGB (Handelsgesetzbuch/German Corporate Law, defining the size criterion for public limited companies. Companies have to satisfy at least two of the following criteria: 1. total assets larger than 3,438 million Euros, 2. revenues larger than 6,875 million Euros, 3. a yearly average of more than 50 employees. 9 Frerichs and Wahrenburg (2003). Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 405 their data, then the benefits from pooling are likely to be smaller. The second point concerns data restrictions. Due to the low default rate in the Deutsche Bundesbank database, randomly drawn portfolios often contain the same defaulters. On the one hand, this reflects reality in so far that if a borrower defaults, then in most cases more than one bank specialized in the borrower's sector is affected. On the other hand, although the Deutsche Bundesbank database is quite large, it does not nearly cover all medium-sized and large German companies. Therefore, the small number of defaults in the database reduces the randomness of samples, which might reduce the average quality of systems that are based on banks' own data. There is one major criticism against the approach of this paper.10 From a purely statistical point of view, the importance of explanatory variables indicating sectoral or regional specialization can be very easily investigated by including them (e.g. as dummy variables) in the pooled credit scoring function and testing for their statistical significance. As a consequence, the complex simulation setup in this study would be superfluous, and given the correct scoring function is chosen, larger portfolios always dominate smaller ones. It is true that the overall significance of regional or sectoral specialization can be tested in this way. Yet, it must be pointed out that the significance of specialization is only tested for the optimal credit scoring function on the pool level. For each particular bank, not only the specialization variable changes, but the bank's optimal credit scoring function might differ considerably from the pooled function in the number and type of explanatory variables. As this study allows each specialized bank to choose its own optimal credit scoring function, it allows for more degrees of freedom than the dummy variable approach on the pool level. With its simulation approach, this paper also takes into consideration that, as each particular bank is autonomous in choosing its rating approach, regulators are only able to dismiss a particular credit scoring function as inferior if they are able to prove this inferiority on the bank's own data. As it is technically inevitable to restrict the range of credit rating models, and as banks may specialize in many more dimensions apart from regions and sectors, the question posed in the title cannot be answered generally for all credit rating systems and all sorts of specializations. Nevertheless, the study gives first answers with respect to two obviously important dimensions for banks in Germany. 10 Thanks to an anonymous referee who made this criticism. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 406 Hergen Frerichs The remainder of the paper is organized as follows. In section II, the simulation set-up is described in more detail. Section III reports results, and section IV concludes. II. Simulation Set-up For ease of exposition, we use the term "bank" to stand for a bank's credit portfolio of medium-sized and large companies. We neglect all other businesses a bank might be engaged in. One assumption needed for our approach is that banks calibrate a specific internal credit rating system for the segment of medium-sized and large corporate borrowers, which seems to be a reasonable assumption due to peculiarities of this market segment. Bank size is defined by the number of balance sheets in a bank's 1994- 1998 training sample. We use bank sizes of 1,875, 3,750, 7,500, and 15,000 observations. As there are on average 3.34 balance sheets per company, the number of training sample companies ranges from 561 to 4,490. We do not consider smaller portfolios because it becomes difficult to derive a sensible credit scoring function for such small portfolios due to the small number of defaults. We do not consider larger portfolios because due to the limited overall size of the database, there would not be enough randomness in the composition of larger portfolios. The 1999 out-of-time sample of each bank consists of all companies that are part of the training sample and that stay customers in 1999. In addition, we simulate new business by randomly drawing new customers such that the bank's portfolio size and portfolio default rate stays constant. In doing this, we construct the situation that the out-of-time sample reflects an average year. As a result there are between 375 and 3,000 companies in the out-of-time samples. For each company, banks have access to the complete annual accounts history as it is available in the Deutsche Bundesbank database. For each bank size, we randomly draw five hundred credit portfolio compositions representing different banks. Each bank has the choice whether it calibrates its rating system only on its own historical financial statements and default data, or uses an externally given credit scoring function, which is based on a larger data set. This larger data set may either be one that is based on data from the whole economy (the complete Bundesbank data set) or one that is based Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 407 on data from a large region or sector of the economy like south or northwest Germany or manufacturing or trade. The system for which banks use only their own data is termed the "stepwise" system, and the one which uses an externally given credit scoring function is termed the "pooled" system. Both the stepwise and the pooled system are based on logistic stepwise selection procedures; only the samples on which system calibration is based differ. Stepwise selection procedures choose those financial variables from a set of forty-nine financial variables that are significant given predefined enter and stay criteria. The set of financial variables consists of forty-eight variables, which have been found to be good default indicators in the German credit risk literature, and one variable from Altman (1968) (Table 1). Forty-one variables are taken from Niehaus (1987), three ratios from Hiils (1995), and four ratios from Deutsche Bundesbank (1999). After selecting financial variables, a logistic regression is carried out to derive credit scores and default probabilities. We employ four different kinds of pooled systems that are based either on the complete Deutsche Bundesbank database or large regional or sectoral subportfolios (south, north-west, manufacturing, and trade). We avoid overfitting by applying a conservative procedure proposed by Shumway (2001) designed to account for the lack of independence between firm-year observations. In this procedure, the significance level for the stepwise selection procedure is corrected by multiplying the value of the partial F test statistic, which is necessary to obtain a confidence level of 90%, by the average number of firm-years per company in each training sample. These new values of the partial F test are used as thresholds to decide on the significance of a variable. This procedure is conservative as it is assumed that there is only one observation per company although on average there are three financial statements. As a result, the four pooled systems are based on four to seven financial ratios. For the stepwise system, we use two different sets of entry and stay criteria for the stepwise selection procedures. On the one hand, we employ the Shumway (2001) procedure used for the pooled system. On the other hand, we neglect the lack of independence between firm-year observations, and simply use a significance level of 5% for the stepwise selection procedures. Our analysis shows that, as expected, the number of variables that are selected by the stepwise selection procedures significantly increases if we implement the second alternative. Using the first Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 408 Hergen Frerichs alternative, there are about 1-3 financial variables in the credit scoring systems of the smallest banks (1,875 in-sample observations), about 2-3 in those of medium-sized banks (3,750 in-sample observations, and about 3-5 in those of the large banks (7,500 in-sample observations). Using the second choice, these numbers change to 3-5, 4-7, and 6-9, respectively. There clearly is a trade-off between system quality and overfitting. The fact that the Shumway procedure is conservative might lead to inferior system quality because important financial variables might be dropped from the credit scoring functions. To neglect the lack of independence might on the other hand lead to overfitting in-sample such that system quality appears to be higher than it actually is. Our simulation findings are that implementing the significance level of 5% considerably increases the quality of the stepwise system relative to the pooled system. This quality increase is not only observed in-sample, where it might be caused by overfitting, but also out-of-time although to a smaller extent. In the following, we report results for the significance level of 5 %. If we used the Shumway procedure for the stepwise systems instead the dominance of the pooled system would be even larger. Banks determine credit scores and default probabilities exclusively based on annual accounts information. They do not add any qualitative information. The inclusion of qualitative factors is not a prerequisite for admittance to the internal ratings based approach of Basel II.11 Yet, many banks base their internal credit ratings on some qualitative components like management quality.12 Since we do not have any additional qualitative information in our dataset, we might underestimate system performance, particularly for small banks. We employ the area-under-curve statistic (AUC) and the Brier score to quantify rating system quality. The AUC measures the quality of ranking borrowers from high to low default risk. If low credit scores are defined to indicate high default probabilities, then all borrowers that actually defaulted in a learning sample should be assigned a relatively low credit score, and those that did not default a relatively high credit score. The AUC is only concerned with ranking, and does not assess the accuracy of 11 Basel Committee (2004), § 417. 12 Basel Committee (2000) surveyed large international banks. Most banks assign ratings using considerable judgmental elements. The relative importance of qualitative versus quantitative factors ranged from very minor to more than 60%. Günther and Griming (2000) report that 72 of 146 surveyed German banks use qualitative criteria for default prediction. 38 of 49 banks state that the quality of default prediction has been improved by the inclusion of qualitative factors. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 415 scores of the stepwise system are above the threshold. This procedure is similar to calculating the power of a test given a type-I error. Relative frequencies below 10% indicate the inferiority of the pooled system, while values above 10% indicate its superiority. In Frerichs and Wahrenburg (2003), this kind of representation is used to identify inferior internal credit rating systems. There, values of at least 50% are taken to indicate a clear underperformance of a given system relative to the pooled system. From Table 7, it can be seen that there is just one case for which the 50%-threshold is reached (Big cities, in-sample). Otherwise results in Table 7 reinforce many of our former findings. Regional specialization seems not to play a large role for the out-of-time quality of internal credit rating systems. In most cases, quality differences between the stepwise and the pooled system are quite alike for specialized and non-specialized banks. Only for out-of-time Brier scores, pooling seems to be even more beneficial for regionally specialized banks than for nonspecialized banks. If we use the stepwise system instead of the pooled system to derive quality thresholds, then there are a number of cases, in which the quality of the pooled system is actually significantly (with a relative frequency of at least 50%) worse than the stepwise system. These are the numbers printed in bold and separated by a semi-colon. Yet, as all these cases refer to in-sample observations they indicate structural differences, but not necessarily quality differences. To sum up this section's results, in-sample results favor non-pooling most probably due to overfitting. Out-of-time, results indicate that pooling is generally beneficial for small to medium-sized banks (1,875 and 3,750 observations) whether or not they are regionally specialized, while it is inferior for larger banks. Overall, differences in system quality are small. 2. Is Pooling Beneficial for Sectorally Specialized Banks? Tables 8 and 9 display simulation results for sectorally specialized banks. Table 8 is equivalent to Table 5 giving direct comparisons of system quality for the stepwise and the pooled system based on AUC values. As again the respective analysis based on Brier scores gives very similar results, we do not display it. Table 9 is equivalent to Table 7 reporting results from the comparison of AUC value and Brier score distributions for the stepwise and the pooled system. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 416 Hergen Frerichs As in the former section, the quality of the pooled system increases relative to the stepwise system as bank size decreases. In-sample, the pooled system calibrated on the complete economy is mostly inferior to the stepwise system. Yet, if the pooled system is calibrated on a sectoral sub-sample (manufacturing or trade), then it is actually superior to the stepwise system for small and medium-sized samples (1,875 and 3,750 observations). This indicates that deriving pooled credit scoring functions depending on sectors might improve system quality. Out-of-time, this result is reinforced, but to a considerably smaller extent. In-sample, there are again some industries that seem to differ structurally from the overall economy as the stepwise system is clearly superior to the pooled system (trade, 7,500 observations; wholesale trade, 3,750, and construction, 1,875). Yet, only for construction this result also holds out-of-time. As this result is particularly strong, it clearly indicates the construction industry differs considerably from other industries, and credit portfolios that are concentrated in this industry do not benefit from pooling. Out-of-time, pooling is mostly beneficial for small and medium-sized credit portfolios (1,875 and 3,750 observations) and particularly for portfolios specialized in metal production and metal products, in mechanical engineering, and in automobile production. Results in Table 9 reinforce those in Table 8. It is noteworthy that out- of-time banks in the sectors metal production and metal products, and automobile production using the stepwise system would be regarded as using an inadequate internal credit rating system by the standards set up in Frerichs and Wahrenburg (2003). For the construction industry the opposite holds. The pooled system clearly underperforms the stepwise system. 3. Is Pooling Beneficial for Sectorally-Regionally Specialized Banks? Tables 10 and 11 display simulation results for banks that are specialized either in manufacturing or in trade and that are additionally specialized in one of four regions. The pooled credit scoring function is either based on manufacturing or trading companies, respectively, from the whole economy. Due to decreasing sample sizes, we are only able to simulate small and medium-sized portfolios (1,875 and 3,750 observations) in theses segments. Again, we leave out the Brier score analysis as it is very similar to the AUC value analysis. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 417 While in-sample results are mixed - pooling dominates non-pooling in quite a few cases - out-of-time results are again clearly in favour of pooling, which is in accordance with our former results for small and medium-sized portfolios. Pooling seems to be particularly beneficial to credit portfolios specialized in trade. Results in Table 11 show the same picture. Out-of-time, the stepwise systems of banks specializing in manufacturing and the north-east, or in trading and the north- or south-east, will be identified as using inadequate rating systems based on Frerichs and Wahrenburg (2003). Overall, it seems that any specialization advantages banks might have are considerably smaller than the added value that results from a sectoral pooled credit scoring function. IV. Conclusion Based on a large financial statements and default database of Deutsche Bundesbank, we use a simulation approach to answer the research question whether pooling of financial statements and default data improves the quality of internal credit rating systems of regionally, sectorally, or sectorally and regionally specialized banks if these factors are disregarded in the pooling function. This research question is important as the economic success of any commercial bank, whether it is obliged to implement Basel II or not, depends to some extent on the quality of its internal credit rating system. Banks that will be obliged to implement the IRB approach of Basel-II have an additional incentive to think about pooling as they need regulatory approval for their rating system. The study's primary result is that pooling improves the quality of internal credit rating systems, particularly of small banks, even if pooling does not take into account regional and sectoral factors although banks are regionally or sectorally specialized. For some specialized banks the results are extraordinarily strong, while for banks specialized in the construction industry pooling does not seem to be of value. When evaluating results, it needs to be taken into account that we strengthen the point of pooling by setting up the pooling functions either based on the complete Deutsche Bundesbank database or on a large part of it. If two banks decide to pool their data, then the quality improvements are likely to be smaller. Also, the low default rate in the Deutsche Bundesbank database might affect results as discussed in the introduction. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 418 Hergen Frerichs For future research, it would be interesting to increase sample sizes by including small companies in order to mitigate data problems and allow simulating larger credit portfolios. Off-course, it would also be of considerable interest to estimate the effects of pooling on real data such as that of the savings or cooperative bank sector. Table 1 Financial Ratios Used as Independent Variables in Credit Scorin16 Financial variables are taken from Niehaus (1987), Hiils (1995), Deutsche Bundesbank (1999), and Altman (1968) (cf. footnote to table). The column 'Hypothesis' indicates whether the value of the financial variable is expected to be generally lower or higher, respectively, for insolvent (I) observations than for solvent (S) observations. Variable Ratio Hypothesis VI operating profit (before taxes) / revenues I < S V2 EBITDA (excl. extraordinary items) / revenues I < S V3 earnings before financial expenses / total assets I < S V4 operating profit (before taxes and financial expenses) / total assets I < S V5 EBITDA (excl. extraordinary items) / total assets I < S V6 (EBITDA (excl. extraordinary items) + financial expenses) / total assets I<S V7 EBITDA (incl. extraordinary items) / total assets I<S V8 (revenues - expenses for raw materials and supplies - amortization of fixed assets - other operating expenses) / total assets I<S V9 EBITDA (incl. extraordinary items) / revenues I < S V10 EBITDA (excl. extraordinary items) / total debt I < S VII EBITDA (incl. extraordinary items) / total debt I<S V12 EBITDA (excl. extraordinary items) / (total debt - cash) I < S V13 EBITDA (incl. extraordinary items) / (total debt - cash) I<S V14 EBITDA (excl. extraordinary items) / (total debt - cash - securities - trade receivables) I < S Continue page 419 Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 419 Table 1: Continued Variable Ratio Hypothesis V15 EBITDA (incl. extraordinary items) / (total debt - cash - securities - trade receivables) I<S V16 EBITDA (excl. extraordinary items) / short-term debt I < S V17 EBITDA (incl. extraordinary items) / short-term debt I<S V18 (short-term debt * 360) / revenues I>S V19 (trade payables + liabilities from accepted bills) * 360 / revenues I>S V20 (cash + securities + trade receivables) / short-term debt I<S V21 working assets / short-term debt I<S V22 (working assets - short-term debt) / total assets I < S V23 (working assets - short-term debt) / revenues I < S V24 (cash + securities + trade receivables - short-term debt) / (operating expenses - amortization of fixed assets) I<S V25 adjusted equity capital / total assets I<S V26 (equity capital + total earnings) / total assets I<S V27 adjusted equity capital / total debt I < S V28 (equity capital + total earnings) / total debt I<S V29 short-term debt / total assets I > S V30 short-term bank debt / total debt I>S V31 (adjusted equity capital + pension provisions + long-term debt) / long-term assets I < S V32 adjusted equity capital / (total assets - cash - properties) I<S V33 adjusted equity capital / (fixed assets - properties) I < S V34 revenues / total assets I < S V35 (debt from accepted bills + trade payables) * 12 / expenses for raw materials and supplies I > S V36 trade receivables * 12 / revenues I>S Continue page 420 Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 420 Hergen Frerichs Table 1: Continued Variable Ratio Hypothesis V37 finished goods * 12 / revenues I>S V38 raw materials and supplies * 12 / expenses for raw materials and supplies I>S V39 amortization / (fixed assets + reductions of fixed assets + amortization) I<S V40 investments / (fixed assets + reductions of fixed assets + amortization - investments) I < S V41 investments / amortization I<S V42 (adjusted equity capital + provisions/2) / total assets I<S V43 (trade payables + debt from accepted bills + bank debt) / (total debt - received advance payments) I > S V44 (trade receivables + inventories) / revenues I > S V45 (adjusted equity capital + pension provisions) / total assets I < S V46 earnings before taxes on income and interest paid / total assets I<S V47 earnings before taxes on income / adjusted equity capital I < S V48 net interest result / revenues I < S V49 retained earnings / total assets I < S 16 Variables V1-V41 are taken from (.Niehaus (19879, p. 75-76). The variable 21 of (Niehaus (19879) is not sufficiently defined so that we do not use it. V42-44 are from Huls (1995), p. 241, Table 22 (V42 = K_122, V43 = K_68A, V44 = K_85). The variable K_08EP cannot be calculated because we do not have data on the change in pension provisions, but V5 is very similar. The variable K_35 = VI9, and the variable K_79 = V34. V45-48 are from Deutsche Bundesbank (1999), p. 55 (V45 = Equity/pension provision ratio, V46 = Return on total capital employed, V47 = Return on equity, V48 = Net interest rate). The capital recovery rate cannot be calculated because it is not sufficiently defined. The equity ratio equals V26. V49 is taken from Altman (1968). Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 421 Û cu cu -+-> Co B o w fi co A S o o fi o '5jd to Cl <3 ce 35 ^ « 3 Ti M 15 g 2 Î « « Ö ä " ü 'S cu 5 M OJ O) rC ^ T3 H 5 . o O Oi rs 11 £ o X CO CO fi co 'SH co g .y 3 -M S S n CO " Q) ö ß ce •i-t in $_, II I 1 2 e tuo g > <v ^ in <—« w O M A> r/T ^ C/3 +-> o fi O) CO —' Jh <v <V -t-> " Sf § Q) g g w b «H O CO o 5 O w „ tuo ^ b fi s a ¿o CO di fH rd QJ CO S ( a - o 3 w) S ö S S £ « o 5 -S s W 'tH £ £ ° CO G 'S 'S « <u PU 3 5 û S » ü H Oh O 1/1 -M ^ ^ fi fi Ü S :cO «Ö & s ö CO rfi W a co E W PQ fi 0) fi cO di PQ .5 ® m a a; o « -3 o co <D ttg cO 2 ifl vi A a> o w "S 3 cO fi cO CO T3 G PQ O) X ? S 2 Ü J ö _ rfi 0) TJ o ^ u co ^ CO o b rj <v O £ w qj £ « i0 ft *h J î-i g ^ cO ^ a. <Z) CO ö •S Qi O ö 'd u i JS > d ^ 1 S £ 25 g ... CO tuo co £ >> s § ^ s -4-> cO t! co <v 3 CO £ I rt fi > cu co -S PQ cO .. PQ "S cO cO~ W '5 5 > s co O PQ CO cO bo W g e ^ g g B £ f-i m x „ o 3 s •S S S î CO fi K ^ - o .S PQ O -m X « tuo . öo .; M w -T QJ co S 'g) u Ö CO ^ § ^ X -M M _g o >Ï ^ o .2 S e co cO CO TJ tuo fi S -s Q « O co -M o o co >5 P S .s Q a ÎH O co o O cO -M o fi o 'Sd Q) PÎ CO CO 00 LO o W) s Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 422 Hergen Frerichs CO S S .2 g « Ci ^ Ifl (1) O w Ö fi H 'S S V O 3 O Ö & ä « 6 * Cd ö * " •2 2 cd -m Ö cd o o S S ^ "g ß 5 S s tì w o ^ •rt co <1 '-H w 7100 ? SS csi E le 1 ìì » rû O Xi 0) V, > cu a Z B ö S fî s § ^ cd a; fH ^ -2 ^ o w ai £ w o cd rC C4H w o <v <D 3 % rz cd w 5 s H rS co >> CT> «5 N a ^ £2 Ö ai cd S 2 Ö ä ü u <u & 2 « iö S ••SI o -M +J 03 ft -m a> c/3 w cd o •g z <D CD Il TJ fa TJ P cd cd oj So Ö rh <v u C ^ •n cd -Oft cd XÌ O a o < M " m ISP cd g Ü <u T3 S ö £ s » S S O ^ cd -S fö tuo Ö H tuo £ ö P "ö .. o a> cj Ü « £ cd o O cd Gl, a _ •H ä .2 cd OS W os cd fa o ^ <u S T3 cd o ^ o T3 g cd G Ö b ° Ö O 3 <D P H ö rt cd S & * "m O a ° a rt a; o g M O « Ci T3 w cd ^ 5 .a 1996 share of German total Insolvencies o o 1—1 CSI 1—1 Oi 1—1 00 CM CM CM co CM 00 1996 share of German total GVA o o 1—1 CSI CSI 1—1 1—1 LO 1—1 CD co CO o 1999 validation sample Default rate in % 0.74 o co o 0.58 0.58 0.44 2.05 0.57 1.23 0.50 1999 validation sample % of total o o rH o IO o co o CM CD IO iH i> co rH os 1999 validation sample # defaults CO CO 1—1 O) CD CM C— CM 00 rH CM os 00 H CM rH 1999 validation sample % of total o o 1-1 CD os co LO CM O 1—1 LO 00 00 CO iH 1999 validation sample sqo # 18,671 8,625 7,233 4,671 1,810 1,025 1,582 1,469 2,392 1994-1998 training sample Default rate in % 0.58 0.63 0.42 0.43 0.26 1.50 0.51 0.96 0.46 1994-1998 training sample % of total 001 LO 00 CM 00 1—1 crH 00 co T—1 o rH 1994-1998 training sample # defaults CO rlO CSI o> CSI oo LO 1—1 CM O rH IO CM 00 OS OS IO 1994-1998 training sample % of total o o 1—1 t> 00 co ^ CM 01 t> os 00 00 iH 1994-1998 training sample sqo # 98,910 46,539 37,467 23,665 9,526 6,514 8,574 8,062 12,786 Sector Overall Manufacturing Trade Wholesale T. Automobile T. Construction Metal Prod. Mech. Engin. Automob. Prod. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 423 «M o -o g tfl 2 ^ ÖJD H M Ï S § ë S ÎH co cd GO <D s fi co co co •r1 cO ä w m 3 .. o 5 M g Ö W g •S CD £ S M -g X .SP 1 PQ £ cO 2 -g «I O s 03 fi < CO tö -ö 5 a « >> Xt Vi G O '3d § ^ ^ o co w S3 PQ il ^ aj biò co 5 rH ^ >> fi o X cO CO fi o X CO • ~ co P ffi a " co I fc CO ,2 « • - w S fi & 2 S S S ^ 55 a) Xi ^ w OT rfi o o ö W I 2 £ * I tuo ÎH fi o .s <D > T5 CO TJ fi cO X co <D fi PQ 0) .fi S <D Q a> rfi H tt S g CO Si O X ^ cO biD CO Ö n a; S 0 PH fi ¡H eu w 1 I tuo $H 3 I g o 5 -S 3 O 0> O fi ^ o o J-l S X ^ s w 2 lo ö « ,2 <u +* 5 g o o ^ s tüD CO .s t eu ÎH W o £ o o co XI Oh w g I eu fi M o <u s ff. 2 -B X fi fi QJ -S 3 fi 0) 2 § m HH J" . . ^ 3 S CO S_i W « •f fi pö • " m 1 « TÎ O fl S t—I AU oT a o 'S U » cO lai S fi S H a> W £ § co ^ o I1 EPS g x fi X § T! co «3 o» CO X PQ 1999 validation sample Default rate in % 0.74 0.55 0.73 1.28 1.17 0.42 0.66 0.88 0.47 1999 validation sample % of total o o 1-1 i—i CT> rH o i—i CT) t> c- i—i CTI CO 1999 validation sample # defaults 00 CO i—H CT> i—i CO CSI <N CO rH CT) ^ CSI CSI 00 1999 validation sample % of total o o 1—1 00 1—1 CT) rH 01 CO II o CSI [> CT5 1999 validation sample sqo # 18,671 3,453 3,580 1,879 1,110 2,146 3,644 1,356 1,717 1994-1998 training sample Default rate in % 0.58 0.60 0.46 1.06 0.57 0.46 0.33 0.60 0.50 1994-1998 training sample % of total 001 o <N IO r—1 00 rH co CT» i—i rH i> 00 1994-1998 training sample # defaults CO t> LO m i—t rH CD OO o rH m CO CO IO o CO i—i 1994-1998 training sample % of total o o 1—1 CT) i-l CT) i-l 01 CO CSI i—l 00 1-1 t> CT) 1994-1998 training sample sqo # 98,910 19,017 18,608 9,816 6,126 11,573 18,045 6,813 8,877 Sector - region Overall Manufacturing South North-West South-East North-East Trade South North-West South-East North-East Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 424 Hergen Frerichs o 0) iO co ' o £ U P < c o CO ft £ o U C8 9 M C eo CQ TS 0> #N 13 tO - CD v 1 M S >> es C © 'Sjd 3 ÖJD 'S o Pm bX) 'S o Oh t: o a J2 CO fi iTt '"J o ^ SH Ä o fi Cd £ 0) O -+-1 CO co .JÜ <u ^ fc fco CO ö 3 1 co w 43 g - o ^ * s s* .Q CO cd CO « -d 0) CO £ S to 0> fr Z ÜJD ^ •S CTJ Ö CO CÖ ^ II ä * CD $h CT1 +-> CU <D Jh O M—I ^ QJ <V > e Ö o s § o c G O 0 0 CO 5 i» ö H CO . CO <V 5-1 fn Z-S co s 's o ° T3 <u Ö Ö CO O 0 ° N <V <D ^ CO CO ° <8 gj - CO £ o 43 co 3 CO SH 43 O PU CO <D co a £ Ol o - Ö a § Ö ^ O <D ' 43 H 5 <V lO I> CO 'W O * o lO t> <V B i <4H 0 1 +-> o 1,500 o o ©^ co" lO t> TO r-T 42 o lO i> O co" <V OL* Ö c CO CO 1 ¡3 i—i 7,500 o o CD i-l 0J 0 «g tuo PS CO ^ Ö 3 £ CO < ^ P Ö o '5b <D « csT 1-1 o in 1—1 o CO 1-1 iH in CS1 c^ i-l CM CM i—l ^ CO CM irt 00 o i—l r> CD t-in CD CM 1—1 1—1 CD lO 1-1 CD rH rH 00 1—1 i—l —' s—' ' ^—' IO <M I> CO m CO CD m CD CD ^—s /—v CD CO 1—1 CD c\T AT 00 CD CO CM rH AT 1T5 CM CM i—l CD m CM CD i—l «0 o in in t> 00 O rH CD O 00 CO 1-1 t^ ^ CD CO CD 00 o ,—s ^—s p ^—s t—1 CD CD CM CO ¿s O i cT i 1—1 m 00 Od Oi CO CO CO CO 1—1 ^ 1—1 00 H cT o T—r CO t-CO CM ^ CM CD CS] cT ^ TH \0 xo \0 \0 x? o^ o^ o^ o^ o^ o CO m o Ci CM ^ CD CM CO CD CM 1—i i—i t-H CM i—l ^ i—l -M w W w CO CO CO w W V lo X! N -M <D f-i w öß > O O o O CO O CO CO 5? W s CM 1-H CO 1-H CO o" in [> CM CM i—l CM rH IO o SD 00 i> _ CO i—l CO CM t> CD rH 1—1 i—1 m i—l ' ••—' s—' 00 CM i-i i-l CD CD CD _ CM [> CO CM 1—1 CD ^ CD O P „—s i—l CO Oi co" CD AT ^ ^ 00 CO CO ^ ^ 00 00 CM 1-1 i—l 00 CM o cT i—T OI o ic 1-1 i-l i-H p o o CM IFT \0 xO vO vP o^ o^ O^ CO m O Ol CD CM 00 CD i—l 1—1 csi i-H * * * * * CO CO CO CO CO w w 43 43 43 in iH o O o o CO J5 CO Z Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 431 References Altman, Edward I.: "Financial ratios, discriminant analysis and the prediction of corporate bankruptcy", Journal of Finance 23 (September 1968), 568-609. - Basel Committee on Banking Supervision: "International conversion of capital measurement and capital standards: A revised framework" (June 2004), Basel. - Basel Committee on Banking Supervision: "Range of practice in banks' internal ratings systems" (January 2000), Basel. - Bloch, Daniel A.: "Evaluating predictions of events with binary outcomes: an appraisal of the Brier score and some of its close relatives", Technical Report No. 135 (May 1990), Stanford University, Division of Biostatistics. - Blochwitz, Stefan/Liebig, Thilo/Nyberg, Mikael: "Benchmarking Deutsche Bundesbank's default risk model, the KMV Private Firm Model and common financial ratios for German corporations", Working paper Deutsche Bundesbank, KMV (November 2000). - Carey, Mark: "Credit risk in private debt portfolios", Journal of Finance 53, 4 (August 1998), 1363-1387. - DeLong, Elizabeth R./DeLong, David M./Clarke-Pearson, Daniel L.: "Comparing the areas under two or more correlated receiver operating characteristic curves: a non-parametric approach", Biometrics 44 (September 1988), 837-845. - Deutsche Bundesbank: "The methodological basis of the Deutsche Bundesbank's corporate balance sheet statistics", Deutsche Bundesbank Monthly Report (October 1998). - Deutsche Bundesbank: "The Bundesbank's method of assessing the creditworthiness of business enterprises", Deutsche Bundesbank Monthly Report (January 1999). - Diebold, Francis X./Rudebusch, Glenn D.: "Scoring the leading indicators", Journal of Business 62 (July 1989), 369-391. - Engelmann, Bernd/Hayden, Evelyn/Tasche, Dirk: "Testing rating accuracy", Risk (January 2003), 82-86. - Frerichs, Hergen/Wahrenburg, Mark: "Evaluating internal credit rating systems depending on bank size", Working Paper Series: Finance and Accounting, No. 115, University of Frankfurt. - Günther, Thomas/Grüning, Michael: "Einsatz von Insolvenzprognoseverfahren bei der Kreditwürdigkeitsprüfung im Firmenkundenbereich", Die Betriebswirtschaft 1 (2000), 39-59 (Use of processes for default prediction for corporate borrowers, in German with English abstract). - Hüls, Dagmar: Früherkennung insolvenzgefährdeter Unternehmen. Düsseldorf: IDW-Verlag, 1995 (Early identification of companies with high default risk, in German). - Lopez, Jose A.: "Evaluating the predictive accuracy of volatility models", Journal of Forecasting 20 (2001), 87-109. - Lopez, Jose A.: "Regulatory evaluation of valueat-risk models", Journal of Risk 1 (June 1999), 37-64. - Niehaus, Hans-J.: Früherkennung von Unternehmenskrisen. Düsseldorf: IDW-Verlag. 1987 (Early identification of company crises, in German). - Shumway, Tyler: "Forecasting bankruptcy more accurately: a simple hazard model", Journal of Business 74 (2001), 101-124. - Sobehart, Jorge R./Keenan, Sean C J Stein, Roger M.: "Benchmarking quantitative default risk models: a validation methodology", Moody's Investors Service, Global Credit Research (March 2000). - Winkler, Robert L.: "Evaluating probabilities: asymmetric scoring rules", Management Science 40 (November 1994), 1395- 1405. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 432 Hergen Frerichs Summary Does Pooling of Financial Statements and Default Data across Specialized Banks Improve Internal Credit Rating Systems? Under the new Basel capital accord, banks will have the opportunity to estimate default probabilities for regulatory capital calculation. In the European Union, most banks will implement the internal ratings based approach, as most corporate customers are not externally rated. Based on data from Deutsche Bundesbank and using a simulation approach, this paper addresses the issue whether pooling of data improves rating system quality even if participating banks are regionally or sectorally specialized and the pooling does not take these factors into account. The primary result is that even under these circumstances pooling is beneficial for most small and medium-sized banks independent of their specialization. (JEL G2, G21, G28, C52) Zusammenfassung Verbessert die Zusammenführung von Bilanz- und Ausfalldaten spezialisierter Banken die Qualität interner Kreditratingsysteme? Mit der Inkraftsetzung der Basel-II-Regeln werden Banken die Möglichkeit haben, Ausfallwahrscheinlichkeiten für die Berechnung des regulatorischen Kapitals selbst zu schätzen. In der Europäischen Union werden die meisten Banken den auf internen Ratings basierten Ansatz umsetzen, da die meisten Firmenkunden keine externen Ratings besitzen. Basierend auf Daten der Deutschen Bundesbank und mithilfe eines Simulationsansatzes adressiert dieser Beitrag die Fragestellung, ob die Zusammenführung von Daten mehrerer Banken die Qualität der Ratingsysteme einzelner Banken verbessert, selbst wenn diese regional oder sektoral spezialisiert sind und die Zusammenführung unter Nichtberücksichtigung regionaler und sektoraler Faktoren erfolgt. Das Hauptergebnis ist, dass die Zusammenführung von Daten selbst unter diesen Bedingungen vorteilhaft für die meisten kleinen und mittelgroßen Banken ist, und zwar unabhängig von ihrer Spezialisierung. Résumé La mise en commun de données de bilan et de données de pertes de banques spécialisées améliore-t-elle la qualité des systèmes internes de rating de crédit ? L'entrée en vigueur de la réglementation Bâle II permettra aux banques d'évaluer elles-mêmes les probabilités de pertes pour calculer le capital régulatoire. Dans l'Union Européenne, la plupart des banques utiliseront les ratings internes car la plupart des entreprises clientes ne possèdent pas de ratings externes. Sur base de données de la Deutsche Bundesbank et à l'aide d'une simulation, l'auteur de cet article examine si la mise en commun de données de plusieurs ban- Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41 Does Pooling Improve Internal Credit Rating Systems? 433 ques améliore la qualité des systèmes de rating des différentes banques, même si celles-ci sont spécialisées au niveau régional ou sectoriel et même si cette mise en commun de données se réalise au-delà des limites régionales et sectorielles. Il en conclut principalement que la mise en commun de données est avantageuse pour la plupart des petites banques et des banques moyennes, indépendamment de leur spécialisation. Kredit und Kapital 3/2005 OPEN ACCESS | Licensed under CC BY 4.0 | https://creativecommons.org/about/cclicenses/ DOI https://doi.org/10.3790/ccm.38.3.401 | Generated on 2023-01-16 13:22:41