PORTFOLIO MANAGEMENT A Unified Approach to Credit Limit Setting by Jeremy Taylor T his article examines the use of economic capital to set limits on credit exposure, extending the framework outlined in a June 2000 article by the same author, and discusses how to structure house limits on individual obligors and concentration limits on portfolio segments. The author finds that, at an aggregated level, stress-testing is more appropriate than portfolio-level limits. The unifying tool is correlation coefficients, which are as important as they are difficult to measure or estimate. T he discussions over reform of the BIS capital adequacy guidelines are shedding new light on the role of economic capital as a tool for managing credit and other banking risk. While the importance of capital as a cushion against losses, credit related or otherwise, has long been understood, the science of risk measurement—coupled with the collection and management of the required data—only recently has progressed to the point of contributing to the practice of portfolio risk management for the banking book. To the extent that an institution has gained understanding and confidence in the output of the credit portfolio model being used, the resulting capital allocations—by borrower, relationship, and various higher levels of customer aggregation—provide an objective measure of risk that can be compared across the portfolio, across different types of risk, and across time periods. Most banks have only recently gained this capability and are just beginning to realize its potential. Limit setting offers a relatively straightforward opportunity to leverage this tool in an area historically characterized by somewhat arbitrary assumptions and procedures. Moving Away from the Traditional Approach While computing limits based on book capital is not unusual, it tends to be done crudely as a fixed percentage of capital—enough to accommodate exposure at the time plus an allowance for growth—and without attention to the varying risk characteristics of individual borrowers within the segment. House limits (on individual obligors and on relationships) may have even less foundation in terms of a tie-in to capital. Ditto for country limits. By computing a limit in dollars of capital and then measuring utilization in economic capital terms, a limit structure can be devised that incorporates and reflects variability in default, recovery, and drawdown rates, as well as the all-important correlation characteristics. And borrowers can be grouped according to the underlying similarities in their loss patterns. All of this provides a more flexible, consistent, and defensible structure that can adjust automatically to changes in the size, composition, and risk profile of a portfolio over time. Why Are Limits Necessary? In a world in which credit losses recurred at the same x% rate each year without variation, there would © 2002 by RMA. Taylor is VP, Portfolio Risk Management, Union Bank of California. The views expressed in this article are his and not necessarily those of the bank. 56 The RMA Journal July/August 2002 A Unified Approach to Credit Limit Setting be no need for capital (forgetting about market and operating risk); losses would be covered by the expected loss (EL) premium over the cost of funds and accounted for by the loss-reserving process. But in reality, of course, the x% is only a long-run average rate around which there’s year-to-year variability. Capital is required to absorb those unexpected losses when they exceed their mean value. And when they greatly exceed it (for example, a severe recession), the question arises as to whether capital is adequate to ensure the institution’s solvency—and with what level of confidence? Uncertainty and variability in default rates and other components of credit loss is a concern for capital management. But of even greater concern is the covariability between loans, which can be thought of as the glue in concentrations. In a large portfolio without large individual exposures, covariances (that is, the off-diagonal terms in the variance-covariance matrix of expected defaults) greatly outweigh the effect of the variances. Limits are necessary to guard against large single or correlated credit events and the erosion of capital they produce. Limits promote granularity by restricting exposure to any single borrower and promote diversification by restricting exposure to any grouping of borrowers. As the banking industry evolves toward more active forms of loan portfolio management, limit setting can be thought of as a front-end, more passive tool— relative to the buying and selling, securitizing, and hedging of credit risk. As a way to measure and monitor concentrations, limits can provide guidance for the deployment of other, more active tools, such as those that identify where to build up exposure or where to sell off or hedge. Grooming the portfolio in this manner can help nudge it toward the elusive “efficiency frontier” of modern portfolio theory, minimizing portfolio risk at any given level of expected return. A Contrary View In theory, with a sufficiently fine-tuned RAROC (risk-adjusted return on capital) model and the pricing discipline to go with it, concentrations would either be compensated for (by a rising risk premium required to reach the target RAROC, as concentration risk increases) or cut off (should potential customers refuse to pay that higher pricing, relative to what other, presumably less concentrated institutions may be charging). In practice, however, two factors build a strong case for having exposure limits: 1. Pricing discipline may wane in the face of competition and the pressure to protect relationships and to 2. leverage industry success. The result may well be accumulation of exposure to levels that do pose concentration risk. As it continues, pricing (which includes the risk premium built into the lending rate but also other credit and noncredit revenues in excess of costs) may not compensate for the incremental risk. Credit portfolio models may not be fully sensitive to the effect of build-ups in exposure. While correlations exert a powerful influence on credit risk capital, they’re also extremely difficult to estimate, and they are not always stable. In particular, they tend to climb in periods of stress, just when they matter most. In this environment, limits provide an important backstop to the portfolio-balancing impetus that a RAROC model alone can provide. The Guiding Principle of Limit Setting An economic capital-based approach to credit limit setting starts with a question: How much of its capital (or earnings) is an institution willing to put at risk to any one obligor, industry, country, and so forth? This approach is consistent with the way that VaR (value-atrisk) models approach the question of market risk and apply it to the setting of market limits. Viewed this way, a limit should be expressed in dollars of capital, as some share of available (book) capital, depending on various risk-related characteristics of a borrower or group of borrowers. Applying the same principles to calculating use of the limit suggests that exposure amounts be converted into equivalent dollars of economic capital to reflect the amount of the bank’s capital being put at risk as a result of that loan or group of loans. Using economic capital in this way offers several advantages over the more traditional approach of viewing concentrations as actual dollars of commitments or outstandings: • It represents “units of risk.” For example, $1 of risk grade 3 exposure is allocated less capital and, therefore, counts less heavily against a segment’s limit than $1 of risk grade 5 exposure. • It’s a single measure to capture exposure, including outstanding, undrawn, and other off-balance-sheet amounts. • It’s fully consistent with the RAROC methodology being deployed by banks for other risk management and performance measurement purposes. • It’s a measure that can be compared directly with a bank’s available capital. 57 A Unified Approach to Credit Limit Setting The Role of Correlations From a regulatory and a creditor perspective, a bank holds capital to protect against losses—more precisely, the unexpected losses (UL). The expected losses are anticipated and absorbed by the EL premium. In recognition of the uncertainty in EL (that is, the yearto-year variability in actual losses experienced), a risk premium is charged on top of the EL premium, which hopefully suffices to provide some targeted RAROC. The amount of capital allocated against a loan is a multiple of its UL, defined as one standard deviation of the loss distribution. That multiple is determined by the targeted agency rating; a higher rating means a lower probability of the bank experiencing losses that push it into insolvency, which, in turn, means holding more capital. UL reflects two things: 1. Variability of EL components—the default, recovery, and utilization rates. 2. Correlation of EL across borrowers. Figure 1 Effect of Loss Correlations Unexpected loss 2 1 0 Rising correlation: 2 > 1 > 0 Expected loss 2. Lower covariances: Diversifying to avoid concentration risk. Note that concentration risk really has two sources. Systematic risk factors—business cycle swings, for example—impact all or a large number of borrowers in a similar way. Granularity effects occur in those portfolios not fine-grained enough to have diversified away all idiosyncratic risk. In a typical larger-bank portfolio— diversified along industry, geographic, and other dimensions and without large individual exposures—risk is predominantly what’s systematic across borrowers. As laid out in Figure 2, the full stand-alone risk (UL) of a borrower can be decomposed into the idiosyncratic (or diversifiable) risk and the systematic risk. The former, in turn, can be split into the following: • The risk that has been diversified away (by filling the portfolio with relatively small, relatively dispersed exposures). • The risk that has not been diversified but could be. That still-diversifiable portion, together with the systematic risk, constitutes a borrower’s contributory risk. Figure 2 The Components of Unexpected Loss Diversified risk Diversifiable risk Systematic risk Figure 1 illustrates how, for a given level of EL, higher loss correlations translate into higher UL and, therefore, higher economic capital allocations. This is because higher correlations increase the likelihood and severity of big losses when a large number of related borrowers get into difficulty concurrently—the opposite of a correlation-free constant x% loss rate every year. It’s the periodic spikes in loss experience due to borrower correlations that will produce a higher UL and a more skewed right tail on the loss distribution. And it’s because of the possibility of correlated credit events imposing occasionally magnified losses on the portfolio that credit models suggest the need for capital at several times UL. Credit risk can be mitigated (and economic capital reduced) in two ways: 1. Lower variances: Taking on transactions with lower inherent (stand-alone) risk. 58 The RMA Journal July/August 2002 S = Stand-alone risk C = Contributory risk Measuring Correlations Correlations—the key piece of this exercise—perform double duty. As described earlier, they drive contributory risk and, therefore, the calculation of economic capital. But they also play a critical role in constructing a limit system. Starting with an individual obligor and going all the way to the full loan portfolio, differences in loss correlation will define and parameterize the groupings of borrowers to which each limit will apply. As important a role as they play, however, correlations are also the toughest part of credit portfolio modeling. This stems from the unhelpful fact that default correlations aren’t observable—certainly not at the obligor level. Logic tells us that the default and loss probability on a loan to United Airlines is correlated with the likelihood of Delta defaulting, but we have no A Unified Approach to Credit Limit Setting default history (for either company, let alone both) to go by. Segment-level default data (for example, how the average default rate for airline carriers changes in response to higher average defaults in some other industry segment or in the economy as a whole) can provide some indication. But it’s not particularly robust in terms of the quantity and quality of data available. Nor is it the whole picture. Conceptually, the default correlation between two borrowers, 1 and 2, in industries A and B, respectively, is given by: D1,D2 A,B A B where A and B are the intrasegment default correlation coefficients (for example, the average probability of any two borrowers in A defaulting together) and A, B is the intersegment correlation (that is, the relationship between changes in the average default rate for A vs. B). Segment-level data will address the latter, but the second two terms on the right side require the companylevel, intrasegment effects—the unobservable piece. Instead, the modeling and estimation of default correlations derive from asset correlations, which look at the co-movement in asset values between firms, or between one firm and the portfolio as a whole. In the context of a Merton-type, options-theoretic approach, fluctuations in firms’ asset value produce default only to the extent that it dips below the liability-determined default point. This is why default correlations are small (mostly in the low to mid single digits, in percentage terms) relative to asset correlations. Moving from asset correlations to default correlations requires factoring in the default probabilities. For two firms, 1 and 2, this means a default correlation, D1, D2 , as follows: D1,D2 D1,D2 D1 D2 E(D1,D2) E (D1) E(D2) D1 D2 E(D1,D2) DP1 DP2 (DP1 (1 DP1)) (DP2 (1 DP2)) where D is default, DP is default probability, and E (D1, D2) is the joint probability distribution, modeled from the joint distribution of asset values. For zeroasset correlation, E (D1, D2) = E (D1) E (D2)—that is, the joint probability is simply the product of the stand-alone probabilities, like the 25% probability of getting two heads in two coin tosses, which results in zero default correlation. As asset correlation rises, and/or as DP rises, so does default correlation. But there’s another wrinkle here. Asset correlations (indeed, asset values) aren’t observable, either. Instead, the usual starting point is equity correlations, which are readily obtainable and which serve as a reasonable proxy for asset correlations but with two important exceptions: highly leveraged firms, and low credit-quality firms.1 In both cases, relying on equity correlations will result in asset correlations being understated. In summary: We need default correlations but can’t observe them. So we fall back on asset correlations— but we can’t observe them, either. So we start with equity correlations, being mindful that they’re an imperfect proxy for asset correlations. But there’s something even easier. Sources of Data A correlation engine is an integral piece of any credit portfolio/RAROC model. An off-the-shelf model, like KMV’s or CreditMetrics’, comes with the required capability to process the customer-level data from a bank’s portfolio. For example, KMV’s GcorrTM (Global Correlation Model) is a multi-factor model for determining pair-wise asset correlations, for translation into default correlations using KMV’s expected default frequencies (EDF). Reliably measured correlations are critical for the allocation of economic capital—that is, having a measure that’s sensitive to any buildups in concentration within a portfolio and to the complex interplay between those borrowers and the rest of the portfolio. Correlation data fulfills a second function—guiding the grouping of borrowers into like segments, based on principles laid out in the next section. How tightly glued (that is, how closely correlated) those borrowers are will be shown to guide the appropriate limit setting for any such grouping. For this purpose, less precise data will suffice—getting them correct in relative terms, with roughly correct differentials between them. For that, there may be intersegment default correlation data, drawing on historical bond default rates, for example.2 But again, we quickly run into the void of intrasegment correlations, the missing half. For those with access to a bells-and-whistles credit model like KMV’s Portfolio ManagerTM or CreditMetricsTM, introduced in 1997 by J.P. Morgan, the required intra- and intersegment correlations can be obtained by generating the required subsets of pairwise correlations—that is, against firms in the same industry (or other segmentation), then against firms not 59 A Unified Approach to Credit Limit Setting in the same industry. It bears emphasizing that this is likely to be the only manageable way to access the full set of data required for all of the pieces of the limit-setting framework outlined in the following sections. Defining Concentration Segments Segments (most obviously by industry and geography, though others are possible) should be constructed according to the loss correlations3 between borrowers. Within a segment, correlations should be relatively high, even if something less than perfect—that is, < 1. Between borrowers in different segments, they should be relatively low (even if > 0). Carving up the portfolio into segments should not be dictated by how large or important a group of borrowers is but rather by how closely correlated they are.4 If there’s too much glue binding a group of borrowers, then that’s the relevant grouping for determining correlated loss potential and how much capital to put at risk lending there. This principle suggests that some borrowers may fall into more than one segment (for example, a computer retailer, which presents both technology and retail industry risk), while others (for whom the systematic risk component is less important) may not fall into any. House limits apply to borrowers and/or relationships. For example, a borrower is effectively a concentration segment with = 1, while a relationship is one where < 1 (that is, not necessarily the case that one borrower in a relationship defaulting means that all the rest will) but is considerably higher than within a typical industry segment. Setting Concentration Limits Let’s start at the segment level, somewhere between borrowers and relationships at one end of the spectrum (Figure 3) and the total portfolio at the other. We’ve constructed our segments as just described. How do we set limits on their exposure? First, we allocate a certain percentage of book capital to each segment (we’ll return shortly to look at how that percentage might be set). That percentage can then be adjusted according to each segment’s risk-related characteristics—for example, its strategic importance, its short-term prognosis, and, perhaps, its riskadjusted profitability. That gives a limit expressed in dollars of capital at risk. In order to calculate how much of its limit a segment is utilizing, we can bring in the economic capital allocation, which measures how much actual risk is attributable to the bank’s lending to that segment. 60 The RMA Journal July/August 2002 As we move from borrower to relationship to segment to portfolio, the within-group correlation coefficient will steadily fall, implying lower concentration risk and, therefore, the ability to take on increased exposure without putting the bank’s capital at undue risk. Exposure limits should be scaled accordingly—that is, the limit should rise with the decline in (the square root of) . A crude approach, in which all concentration segments (all industry groupings, for example) start with the same x% of book capital, is certainly simpler to do and less demanding data-wise. For that approach, intuition and economic logic may suffice for going through the exercise of grouping borrowers into segments. But with the benefit of the required segment-level data, this can be done more flexibly, more sensitively, by giving lower- segments a higher base-capital allocation. This relationship is depicted in Figure 3. Figure 3 A Risk Continuum House limit: Relationship House limit: Obligor Major segment limit Subsegment limit: Industry, country... As (intra-group) correlation Portfolio limit? ... ... limit amount Correlation data can allow the base capital allocation for each segment’s limit calculation to be appropriately scaled up from a starting point. Which starting point? The obligor level (with its = 1), most obviously, though it could also be a concentration segment, perhaps the one with highest intrasegment correlation. But how do we anchor that starting point? It could be by working backwards—allocate a base capital amount that produces a reasonable base limit, relative to existing exposure and comfort levels. As a refinement of that, the institution could look for internal agreement on the maximum loss that it’s prepared to take on any one borrower or credit event. House Limits Suppose a bank’s concentration segments exhibited an average within-segment default correlation of 4%. The relationship between house limits (HL) and concentration limits (CL) should then be approximately as A Unified Approach to Credit Limit Setting follows: HL = 0.2 CL , where 0.2 = 0.04 Some further downward adjustment of that fraction may be warranted as a granularity adjustment—that is, to recognize the typical size dispersion of exposures within a concentration segment (for example, more than 5 [ = 1/0.2] times as many borrowers). This would express house limits in dollars of capital, as was done earlier for concentration limits. But it may be more practical to convert and communicate them in dollars of commitments, as a guide to lenders and line management. If month-end or quarter-end analysis reveals a segment to be over its concentration limit, the institution can inform the line of whatever measures then come into effect (as discussed below). But it will rely on individual lenders to monitor and comply with house limits as they go, which is more easily done in commitment dollars than economic capital dollars. Converting a limit figure into dollars of commitments requires dividing it by an appropriate economic capital allocation factor (that is, capital per dollar of commitments, from the bank’s RAROC model output), differentiated by major risk criteria, such as risk grade and line of business. For example: If a limit for a risk grade 5 borrower in a particular area were $10 million of economic capital and the capital allocation per dollar of commitments were $0.05, then the house limit would be $200 million in commitment terms. Whereas concentrations can be computed and tested on a regular (at least quarterly) basis, it may make sense to have house limits set and maintained for longer periods—say, annually. Among other things, this reduces the problem of having individual borrowers fall out of compliance (that is, go over house limit) without a change in their commitment amount. Relationship Limits Relationship limits are similar to house limits, but they are aggregated across obligors to the relationship level. The need for relationship limits in addition to house limits is an open question. Furthermore, there are (even bigger) data issues that come into play—for example, how do we measure the likelihood of all borrowers within a relationship going into default together? Intuitively, the number is high, certainly higher than for any concentration segment, though how high will depend importantly on an institution’s aggregation rules for defining what constitutes a relationship. This is not easily quantified; neither is the underlying correlation within a relationship likely to be constant across a bank. One important example: For commercial real estate, relationship (that is, sponsor) limits may be higher than for general commercial relationships. This is because the dispersion of individual projects with different tenants and more autonomous legal structures means less likelihood of one defaulting project triggering others within the same relationship. The upshot: If separate relationship and obligor limits are deemed necessary, an assumption about the degree of commonality within a typical commercial relationship is required to scale up the relationship limit from the obligor limit. Commercial real estate appears different enough to warrant separate treatment (that is, a separate scaling factor). Country Limits The theoretical framework outlined here could certainly be applied mutatis mutandi to setting and monitoring country limits. Indeed, the earlier discussion of concentration limits was couched in terms that are general enough to apply to any portfolio segmentation, including a geographic one like countries or regions. In practice, though, there may well be data limitations that make RAROC determination more complicated and less reliable for international customers. Just as important, the exercise of setting country limit amounts is likely to be overwhelmed by the subjective factors—in particular, the strategic or comfort factors. They’re intended to capture the institution’s commitment to, experience in, and comfort with a given country. These considerations are likely to vary significantly from one country to the next, much more so than the industry-toindustry differences that a bank’s senior management might identify. Meanwhile, the data required to draw comparisons of intra-country correlations is likely to inhibit that part of the exercise. Portfolio Limits versus Portfolio Stress-Testing Continuing in the same direction (that is, to the right end of the Figure 3 continuum), it’s possible to go to the extreme of setting a portfolio-level limit on the total lending book, tied to the overall level of capital in the same way that the various other limits were. The fact that correlation across the whole portfolio would be that much lower than in any single segment (which a full credit model would substantiate and quantify) means a correspondingly higher limit than for any segment within it. However, there are practical problems here, too. In particular: Whose limit is it? If a portfolio-level limit were calculated and the bank were getting close to it, it article continues on page 72 61 A Unified Approach to Credit Limit Setting continued from page 61 becomes an unwieldy management tool. Decisions regarding overall portfolio size, growth, and composition are part of board-level strategy, not the short-term grooming that limit-setting produces. That said, however, the exercise of comparing the institution’s total capital to the credit and other risks (expressed in economic capital terms) that lurk in its balance sheet and operations is an important one. Rather than approaching it from the viewpoint of setting aggregate limits on the institution’s assets based on its available capital, the exercise can be turned around to one of stress-testing—taking the existing portfolio and the stresses to which it might reasonably be subjected and assessing the adequacy of bank capital to absorb the resulting losses.5 Stress-testing is familiar enough at the obligor level as an assessment of the firm’s ability to continue to service debt under stress conditions. At the segment level, it aims to determine relative vulnerabilities across the portfolio and the stress-scenario impact on loss reserves. Our concern here is the portfolio (or macro) level. In setting concentration limits, the portfolio was segmented into buckets of relatively homogeneous risk. Concentration risk stems from overexposure to shocks emanating from and largely confined to individual industries or regions. Stress-testing, on the other hand, recognizes that, in addition to those segment-specific shocks, there are systematic risk factors (like interest rate changes, recessions, equity shocks) that affect all segments, though to varying degrees. These are why > 0 between segments. Portfolio-level stress-testing is a substitute for having limits in place to protect against those macro-level shocks. It looks at the incremental losses and the resulting erosion of bank capital (defined to include loss reserves) under stress conditions deemed “reasonably possible” by historical norms. Limit Compliance and Breach Management Crafting a new limit structure designed to signal unhealthy buildups in concentration won’t do much good without some teeth. On the spectrum from soft to hard limits, shown in Figure 4, there are probably benefits to seeking a middle strategy somewhere between “raising the limit” and an “absolute ban on new business.” The former undermines the integrity of the process, while the latter places too much faith in it; there’s a need for some flexibility, given the data limitations already discussed, the drawbacks of the SIC codes for classifying industries, and other such “real world” 72 The RMA Journal July / August 2002 intrusions. Response to the breach of a segment’s limit may therefore include a tightening of underwriting guidelines and/or credit approval procedures or perhaps temporary approval of the over-limit amount until it can be sold down or hedged. Figure 4 Limit Compliance Alternative Breach Management Strategies SOFT LIMIT Liquidity-based constraints (only sectoral review: loans that are saleable can be •install higher booked)...perhaps pricing hurdle with a temporal •revisit limit-setting assumptions dimension (i.e., a •revisit underwrit- temporarily higher limit, where liqing guidelines uidity justifies) Raise Use breach as limit opportunity for Absolute ban on new business HARD LIMIT Conclusion The development and acceptance of economic capital as a risk measurement tool are just beginning to impact bank portfolio management practices. Credit limit-setting is an area in which traditional approaches have tended not to give due attention to the key differences between borrowers or segments in the underlying risk drivers. By incorporating the variances and covariances in loss patterns, economic capital provides a comprehensive measure of risk, one that can be directly compared with an appropriate calculation of the amount of capital an institution is willing to put at risk in lending to a particular borrower or group of borrowers. This can make the limit-setting process more defensible and more disciplined, so as to steer the institution away from the concentration risk that a more focused and specialized approach to lending can cause. ❐ Contact Taylor by e-mail at [email protected] Notes 1 For a good discussion of this, see "Measuring Credit Correlations: Equity Correlations Are Not Enough," B. Zeng and J. Zhang, unpublished KMV paper, January 2002. 2 For example, Michael Niemira of Bank of Tokyo-Mitsubishi in New York has done this analysis on high-yield bond defaults, using Fitch data. 3 We'll assume independence in recovery and drawdown rates and focus on default (vs. loss) correlations. 4 Of course, a larger group of borrowers (e.g., multi-industry, or multi-region) is likely to be more heterogeneous (i.e., to have lower intra-group correlation). 5 See "Stress-Testing a Commercial Loan Portfolio," J. Taylor, Commercial Lending Review, March 2002.
© Copyright 2026 Paperzz