Discover millions of ebooks, audiobooks, and so much more with a free trial

Only $11.99/month after trial. Cancel anytime.

Credit-Risk Modelling: Theoretical Foundations, Diagnostic Tools, Practical Examples, and Numerical Recipes in Python
Credit-Risk Modelling: Theoretical Foundations, Diagnostic Tools, Practical Examples, and Numerical Recipes in Python
Credit-Risk Modelling: Theoretical Foundations, Diagnostic Tools, Practical Examples, and Numerical Recipes in Python
Ebook1,375 pages11 hours

Credit-Risk Modelling: Theoretical Foundations, Diagnostic Tools, Practical Examples, and Numerical Recipes in Python

Rating: 0 out of 5 stars

()

Read preview

About this ebook

The risk of counterparty default in banking, insurance, institutional, and pension-fund portfolios is an area of ongoing and increasing importance for finance practitioners. It is, unfortunately, a topic with a high degree of technical complexity. Addressing this challenge, this book provides a comprehensive and attainable mathematical and statistical discussion of a broad range of existing default-risk models. Model description and derivation, however, is only part of the story. Through use of exhaustive practical examples and extensive code illustrations in the Python programming language, this work also explicitly shows the reader how these models are implemented. Bringing these complex approaches to life by combining the technical details with actual real-life Python code reduces the burden of model complexity and enhances accessibility to this decidedly specialized field of study. The entire work is also liberally supplemented with model-diagnostic, calibration, and parameter-estimation techniques to assist the quantitative analyst in day-to-day implementation as well as in mitigating model risk. Written by an active and experienced practitioner, it is an invaluable learning resource and reference text for financial-risk practitioners and an excellent source for advanced undergraduate and graduate students seeking to acquire knowledge of the key elements of this discipline.
LanguageEnglish
PublisherSpringer
Release dateOct 31, 2018
ISBN9783319946887
Credit-Risk Modelling: Theoretical Foundations, Diagnostic Tools, Practical Examples, and Numerical Recipes in Python

Related to Credit-Risk Modelling

Related ebooks

Business For You

View More

Related articles

Reviews for Credit-Risk Modelling

Rating: 0 out of 5 stars
0 ratings

0 ratings0 reviews

What did you think?

Tap to rate

Review must be at least 10 words

    Book preview

    Credit-Risk Modelling - David Jamieson Bolder

    © Springer International Publishing AG, part of Springer Nature 2018

    David Jamieson BolderCredit-Risk Modellinghttps://doi.org/10.1007/978-3-319-94688-7_1

    1. Getting Started

    David Jamieson Bolder¹ 

    (1)

    The World Bank, District of Columbia, Washington, DC, USA

    If I cannot do great things, I can do small things in a great way.

    (Martin Luther King, Jr.)

    Risk-management teams, in financial institutions around the world, are classically organized by functional responsibility. This implies distinct credit-risk and market-risk departments. Increasingly, we observe additional sub-groups responsible for operational risk, but the two solitudes of market and credit risk still remain more often the rule than the exception.

    Such a distinction is not necessarily a poor way to organize a risk-management function. Market and credit risk do have some important differences and, in many cases, require different skill sets. For the modern risk manager, however, there is a significant advantage to being familiar with these two key dimensions. A risk manager who looks at the world solely through his own, be it credit- or market-, risk lens, is more likely to miss something important. A complete risk manager appreciates both of these key dimensions of financial risk. This is particularly true given that the distinction between these two areas is not as clear cut as many pretend.

    There is already a large number of excellent works, however, in the area of credit-risk modelling. Why, the reader is entirely justified for asking, do we need another one? This book, written by a risk-management practitioner, takes an inherently practical approach to modelling credit risk. Much of the extant literature is either extremely mathematically complex and specialized or relatively general and vague on the technical details. In this work, the focus is on the tools and techniques required to actually perform a broad range of day-to-day, credit-risk computations. These include, but are not restricted to, estimating economic-capital computations, decomposing the contribution of each obligor to overall risk, comparing the results in the context of alternative underlying mathematical models, and identifying key model parameters. There are, thankfully, many works that may be used by the credit-risk modeller seeking more insight into the world of market risk.¹ Fewer works exist for the market-risk manager looking to familiarize herself with the practical details of the credit-risk world. This book seeks, in part, to rectify this situation.

    These ideas can all be found in other sources, in varying degrees of complexity, but are rarely combined together in a single source and developed from first principles. The unifying concept behind this work is the provision of a road-map to the quantitative analyst who is trying to assess the credit riskiness of a given asset portfolio. To this end, we will use a common portfolio example throughout the entire book. Moreover, in addition to the mathematical development and motivation associated with different modelling techniques, we will also provide tangible illustration of how to implement these ideas. This will take the form of concrete computer-code snippets from the Python programming language. The objective is to thereby shorten the distance between textbook and practical calculation. Python also offers many tools to visually supplement our analysis. Visualization of quantitative data is of the utmost importance and we seek to incorporate this aspect—inspired heavily by Tufte (2001)—in all aspects of the modelling discussion.

    Credit-risk modelling is a vast area of practitioner and academic research. The existing collection of models are both rich and complex in terms of technical detail and application. A single work, of manageable size, cannot hope to cover it all. Some focus is, therefore, necessary. We have, consequently, chosen to emphasize the risk-management dimension. While many of the models in this work can be extended to price credit-risky instruments, our focus, however, will be on the construction of empirical default-loss distributions under the physical probability measure. Once this distribution is in hand, it can be used to compute tail probabilities, risk measures such as VaR and expected shortfall, and to approximate the contribution of individual credit obligors to these measures.

    Another area that perhaps makes this work somewhat unique is its incorporation of a dimension of growing importance in the world of risk management: model risk. The risk of model misspecification, selection of inappropriate parameters, or implementation errors is significant with credit-risk models. Indeed, it is typically judged to be significantly higher than in the market-risk setting. Unlike the high-frequency movements of market-risk factors, default is a rare event. This paucity of default data creates challenges for model selection, calibration, and estimation; back-testing one’s model, in most cases, is literally impossible. Written by a practitioner, who has spent much of his career on the market-risk side of the business, this book has a keen appreciation for the data shortcomings in the world of high-quality credit obligors. As a consequence, it consistently seeks to explore modelling alternatives, the use of diagnostic tools, and the direct incorporation of the model-risk dimension into the credit-risk modelling framework. The aim is to thereby enhance the usefulness of the discussion for risk-management practitioners.

    In this first chapter, we highlight and motivate the key questions addressed, in a risk-management setting, by credit-risk models. This exercise provides useful insight into what makes credit-risk modelling such a challenging exercise. A comprehensive comparison with the market-risk problem also helps to identify the unique elements involved on the credit-risk side. We also introduce our sample portfolio and provide an illustrative, and highly graphical, demonstration of what we seek to accomplish. As a first step, however, we need to provide a greater degree of precision regarding how these specific issues will be addressed in the latter chapters. That is, we begin by detailing the perspective adopted in this book.

    1.1 Alternative Perspectives

    As a starting point, it is important to have a clear idea of what precisely one means by credit risk. Credit risk stems, ultimately, from a counterparty, or obligor, defaulting on their obligation. It can be an outright default or a deterioration of an obligor’s ability to pay, which ultimately makes default more probable. The former situation is termed default risk, whereas the latter is typically referred to as migration risk. Our principal focus in the following chapters is upon the default event, but we will see that this perspective is readily generalized to the migration setting. It is, in fact, precisely in this latter case that the credit perspective begins to overlap with market-risk analysis.

    Default or migration events, which are the key elements of credit risk, lead to financial losses. The scope of the loss can range from a small fraction to the entirety of one’s investment. This type of risk can arise with almost any asset that requires ongoing future payment. Typically, however, we think of loans, bonds, swaps, or deposits. The notion of default, however, can be broadly applied to credit-card obligations, student loans, mortgages, car loans, or complex financial derivative contracts. One can even use—by defining default as the occurrence of a rare, non-financial, risk event such as a fire, a car accident, or an earthquake—credit-risk modelling techniques in assessing insurance claims.² Thus, default risk is inherent in two key pillars of the finance industry: banking and insurance.

    1.1.1 Pricing or Risk-Management?

    An assessment of the uncertainty surrounding the payment of future asset claims is thus central to informing important decisions about the structure of financial claims. Credit-risk models act as an important tool in performing such an assessment. There are, at least, two broad applications of credit-risk models:

    1.

    pricing credit risk; and

    2.

    measuring the riskiness of credit exposures.

    The first element seeks to estimate the price at which one would be willing—in the face of uncertainty regarding a counterpart’s ability to pay—to purchase an asset.³ It looks to determine, on average, what is the cost of credit riskiness associated with a given position. This price is usually treated as a credit charge, typically as a spread over the risk-free rate, on the loan obligation. Equivalently, this amounts to a discount on the value of the asset. Pricing naturally leads to the notion of mean, or expected, default loss.⁴

    Riskiness is not so much about what happens on average, but rather what is the worst case event that might be experienced. In risky physical activities, such as sky-diving and race-car driving, worst-case events are typically quite gruesome. In a financial setting, however, we are worried about how much money one might lose. It is insufficient, however, to simply say: we could lose our entire investment. While entirely true, this is missing a key ingredient. We need to consider the interaction of a range of possible losses and the associated probability of incurring these losses. In the combination of outcomes and likelihood of negative events lies the assessment of risk. Marrying these two ideas together naturally leads us to statistics and the notion of a default-loss distribution.

    The distribution of default losses is the object of central interest for credit-risk modellers. Pricing and risk-management, therefore, both share this interest. In the former case, the central aspect is of more interest, whereas the latter focuses on its tails. Figure 1.1 describes this situation. Let us denote the default loss as the random variable, L, for a given portfolio, over a given time interval. The default-loss density, f L(), summarizes the relative probability across a range of possible loss outcomes, denoted . The form of this, admittedly schematic, default density function is quite telling. It places a large amount of the probability mass on small levels of loss and a correspondingly modest amount of probability on large losses. This is very typical of credit-loss distributions. In a phrase, the probability of default is typically quite low, but when it occurs, losses can be extensive.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig1_HTML.png

    Fig. 1.1

    A (stylized) default-loss distribution: Here is a popular schematic that reveals a number of key points about the default-loss distribution.

    Figure 1.1 also highlights a few aspects of the default-loss distribution, which provide additional precision on the distinction between pricing and risk-management. The average default loss, denoted as $$\mathbb {E}(L)$$ , is the principal input for the determination of the price of credit risk. Conversely, the value-at-risk, or VaR, is one possible—and quite useful—measure of the tail of the distribution. It is used extensively by risk managers to assess extreme, or worst-case, outcomes. It would be an oversimplification, however, to state that valuation experts care solely about the central part of the default-loss distribution, whereas risk managers focus only on its tail. Both perspectives routinely employ multiple aspects of the statistical description of default. Figure 1.2 provides an, again highly schematic, description of how credit-risk is priced. It holds that the price of credit risk is the expected default plus some additional terms. One additional term is a reflection of administrative costs, which fall outside of the realm of modelling, but another term seeks to incorporate tail information.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig2_HTML.png

    Fig. 1.2

    Pricing credit risk: For pricing credit risk, both the expectation and the (expensive) capital requirement are typically involved.

    Financial institutions, as we’ve discussed, take default risk in many ways. As a first step, it needs to be properly priced so that proper compensation for risk is provided. These institutions also set aside some additional funds—which act as a kind of reserve or rainy-day fund—to cover any unexpected losses arising from default outcomes. Figure 1.1, after all, clearly indicates that the distance between average and extreme outcomes can be quite substantial. These additional funds are often referred to as economic capital. There are different ways to estimate economic capital, but it generally amounts to the difference between a worst-case, or tail, risk measure and the expected loss. It is thus occasionally referred to as unexpected losses.⁵ Regulators of financial institutions are very concerned about the magnitude of one’s economic capital and the methodology required for its computation.

    How does this idea impact the pricing of credit risk? Setting aside funds in the amount of one’s economic capital is expensive. It implies that these funds cannot be used to make additional loans and generate profits for the firm. Many practical pricing models, therefore, add an explicit charge to compensate the lender for the costs of holding economic capital. A natural way to estimate this expense is to multiply the approximated economic-capital amount with the firm’s cost of capital, or hurdle rate. In this way, the computation of a price for credit risk uses multiple aspects of the default-loss distribution. In a phrase, therefore, the price of credit risk associated with a security can depend on both the central and extreme aspects of the default-loss distribution.

    1.1.2 Minding our $$\mathbb {P}$$ ’s and $$\mathbb {Q}$$ ’s

    One of the major insights of modern mathematical finance is that the price of an instrument is well described by the discounted expected future value of its pay-outs. The classic example is the celebrated Black and Scholes (1973) option-pricing formula. This is not, however, an arbitrary expected value, but rather the expectation taken with respect to the equivalent martingale measure, $$\mathbb {Q}$$ , induced using an appropriate numeraire asset.⁶ When we select the money-market account as our numeraire asset, this is popularly referred to as the risk-neutral probability measure. This result is termed the fundamental theorem of asset pricing. While this can get quite mathematically involved—the interested reader is referred to Harrison and Kreps (1979) and Harrison and Pliska (1981) for the foundations—the intuition is quite appealing. Discounting future cash-flows is difficult, because the discount rate depends on their riskiness. The assumption of lack of arbitrage implies the existence of an equivalent martingale measure, $$\mathbb {Q}$$ , which takes risk considerations out of the equation. Expected pay-outs under this measure can thus be discounted using the risk-free rate.

    The probability measure thus plays a critical role in mathematical finance. In practitioner and regulatory settings, such as BIS (2001, 2004, 2005), the probability measure is rarely mentioned. It is thus relatively easy to get confused. It is nevertheless, if not explicitly mentioned, always present. A bit of clarification is, therefore, useful.

    Valuation analysts spend the vast majority of their time working under the $$\mathbb {Q}$$ measure to price various types of financial instruments. Credit-risky securities are no exception. In credit-risk pricing work, like say Bielecki and Rutkowski (2002), the equivalent martingale measure plays a starring role. For risk managers, however, the situation is rather different. Risk managers are less interested in understanding the asset prices at a given point in time, but instead are focused on how the value of instruments move dynamically over time. While the $$\mathbb {Q}$$ measure is required to price each instrument at each static point in time, it provides little insight into the intertemporal dynamics of these valuations. This is captured by the true, or physical, probability measure, which is generally denoted as $$\mathbb {P}$$ . This is often referred to as the real-world probability measure.

    This topic is extremely technical and subtle; this work does not intend to add to this important and fundamental area. Duffie (1996) is, for interested readers, an excellent place to start to understand the complexities of asset pricing. We seek only to raise the important point that, in the large, pricing applications employ the equivalent martingale, or risk-neutral, measure, $$\mathbb {Q}$$ . Risk managers, conversely, typically use the physical measure, $$\mathbb {P}$$ . This is, of course, something of a simplification, but nevertheless represents a reasonable rule of thumb.

    The key point is that the differences between these two measures are intimately related to aggregate market attitudes toward risk. Valuation works best when risk preferences can be extracted from the computation. Risk managers, by the very definition of their task, are not able to ignore risk attitudes. It is at the heart of what they do. Risk-management, to be relevant and useful, has no choice but to operate under the physical probability measure.⁷ Given our risk-management focus, this book will also predominately focus on the real-world probability measure. Any deviation from this perspective will be explicitly mentioned and justified.

    1.1.3 Instruments or Portfolios?

    To this point, it is unclear whether we refer to a specific credit-risky security or a portfolio of such instruments. The simple answer is that the default-loss distribution found in Fig. 1.1 could equally apply to either a single instrument or a portfolio. The choice is, however, not without importance. Typically, we imagine that the default or migration risk associated with a single instrument is determined by the creditworthiness of a sole entity. Portfolio default risk, conversely, is determined by the interdependence of the credit worthiness of multiple entities. More technically, the former involves a single marginal distribution, whereas the latter encompasses a full-blown, joint-default distribution. We can all agree that joint distributions of collections of random variables are, on the whole, more involved than any of the individual marginal distributions.

    Once again, this is something of an oversimplification. Some complex instruments—such as collateralized debt obligations or asset-backed securities—do actually depend on the creditworthiness of large number of distinct obligors. Such instruments are, in fact, actually portfolios of other simpler securities. The message, however, remains the same. Understanding, and modelling, the credit risk of a portfolio is a more involved undertaking than dealing with a single security.

    In this work, we will tend to focus on portfolios rather than individual securities. This viewpoint allows us to place more emphasis on the modelling of default and migration events. It places the credit-loss distribution at the forefront of our discussion. This is not to say that the characteristics of individual securities in one’s portfolio do not matter. They matter a great deal. For our purposes, however, it will be sufficient to understand the size of the exposure associated with each obligor, the unconditional probability of its default or credit migration, and the magnitude of the loss in the event of a default outcome. We will, in some cases, require additional information—such as the tenor of the individual security—but information such as day-count conventions, coupon rates, payment indices, or implied volatilities will not not explicitly considered.

    It is not that these elements are not employed in credit-risk analysis. They do, in fact, play an important role. There is an entire literature allocated to determining the worst-case exposure that one might have in the event of a default event. The ideas of potential future exposure (PFE) and expected positive exposure (EPE) arise in this setting. This is, in many ways, the flip side of the typical market-risk problem. Market-risk models generally seek to identify the worst-case loss—for a given portfolio over a given time interval with a particular level of confidence—that one might experience due to market-risk movements. In a credit-risk setting, the greater the market value of an instrument, the larger its exposure to default, and the worse the ultimate default-loss outcome. PFE and EPE models, therefore, are conceptually in the business of estimating the best-case VaR measures for one’s portfolio. While these are an integral part of credit-risk modelling, they do not describe the heart of the undertaking. Indeed, for the most part, they fall into the category of market-risk modelling.

    1.1.4 The Time Dimension

    Time is always a complicating factor. Risk-management usually involves computing high-frequency estimates—daily, weekly, or monthly—of the risk associated with one’s current portfolio over a pre-defined time horizon. For market-risk practitioners, the time horizon is typically quite short, usually ranging from a few days to about a month. In the credit-risk world, the approach is similar, although the time horizon is generally somewhat longer—one-year is probably the most popular, but multiple years are not uncommon.

    This raises an important question: what kind of assumptions does one make about the evolution of the portfolio over a given horizon? The shortness of the market-risk perspective makes it reasonably defensible to assume a static portfolio. When the horizon is multiple years, however, this is less easy to defend. In the credit-risk setting, there are, at least, two aspects to this challenge: changes in the portfolio and movements in the creditworthiness of the individual obligors.

    It is common practice, in both market- and credit-risk settings, to assume that the portfolio does not change over the horizon of analysis. In some cases, this leads to silly results. If one’s horizon is two years, but some instruments possess shorter tenors, it is often clearly unrealistic to treat the portfolio as unchanged. Moreover, if the portfolio is managed relative to a benchmark, it should be expected to change frequently to reflect benchmark movements. The problem, however, is that trying to describe—in a sensible, tractable, and realistic way—these structural or tactical portfolio movements is very difficult. It basically amounts to forecasting future changes in your portfolio. In many cases, this may prove to be virtually impossible.

    To the extent that one is trying to measure the amount of risk—either from a credit or market perspective—it may not even be desirable to model the dynamic evolution of one’s portfolio. Ongoing risk monitoring and management often employ a snapshot approach to portfolio surveillance. Computing the credit-risk, over a given horizon, based on a static portfolio assumption allows one to compare today’s portfolio to tomorrow’s and to last week’s. A static, one-period perspective is thus, in many ways, the risk manager’s default setting. In the following chapter, for the most part, we will also adopt this viewpoint.

    This is not to say that the dynamic perspective is uninteresting. Static long-term credit analysis, as suggested, often assumes that the credit worthiness of the individual obligors is also constant. This is a strong assumption and can be relaxed by permitting credit migration. This turns out to be a reasonably straightforward extension of the static, one-period approach. Incorporating losses, and gains, associated with non-default migration, however, is a bit more work. We will, in the upcoming chapters, discuss the essence of this idea and provide appropriate references for moving further in this direction. Our main focus will be, to be clear, on static, one-period default models. As will soon become evident, this is already a sufficiently complicated undertaking.

    1.1.5 Type of Credit-Risk Model

    When embarking upon, even a cursory, overview of the credit-risk modelling literature, one will be confronted with two competing approaches: the so-called structural and reduced-form methodologies. The difference is relatively easy to describe. Structural models are descriptive; they say something about the reason behind the default event. In this setting, default is endogenous to the model. This is consistent with the way the term structural is used in other modelling frameworks. It denotes the imposition of specific model behaviour as suggested by theory.

    Reduced-form models take an alternative perspective. These methods treat the default event as exogenous. That is, there is no structural description of the underlying cause of default. It simply occurs with a particular probability. The driving idea is that the default event is modelled in an empirical sense, rather than informing it from a structural or theoretical perspective.

    A simple example might help the reader understand the differences between these two models. Imagine that your boss is trying to construct a model of your, perhaps difficult-to-predict, morning arrival time. A reduced-form model might simply involve collecting some data and describing your office entry with a classic Poisson-arrival technique. This is a purely statistical approach. A structural approach, conversely, would seek to describe the main reason for your tardiness. The structurally minded boss might model your arrival as conditional upon the weather, the time of year, the state of public transportation, or the day of the week. Clearly, there is some room for overlap, but there is a basic difference. Reduced-form models are generally agnostic about the underlying reasons and are informed by data, whereas structural model seek to impose certain relationships often imposed by theory or logic.

    There are advantages and disadvantages to each approach and, as such, it is not particularly useful to describe one method as superior to the other. In the following chapters, we will examine both methodologies. This provides a useful insight into the differences between the techniques. Comparing and contrasting alternative model assumptions is also a powerful tool in managing and mitigating model risk. As famously stated by the statistician George Box, all models are wrong, but some of them are useful. This, slightly depressing mantra for quantitative analysts, suggests that we should welcome competing model approaches. It allows us to test the robustness of a result to their underlying, and differing, assumptions. Moreover, an awareness of the fallibility of our models, along with understanding of (hopefully) realistic alternatives, are the first defence against over-reliance on a specific modelling technique and the associated model risk such a reliance creates.

    1.1.6 Clarifying Our Perspective

    As we’ve seen in the previous discussion, there is a wide variety of possible perspectives one might reasonably adopt for credit-risk modelling. Some focus is, therefore, required. We may, for example, choose among:

    Application

    Pricing or risk-management;

    Granularity

    Instrument or portfolio level;

    Time

    Dynamic or static perspective; or

    Model Type

    Structural or reduced-form approaches.

    To summarize, in a concise manner, the choices taken in this book, we will consider static structural and reduced-form credit-risk models in a portfolio setting from a predominately risk-management perspective. Although our principal application is risk management, most of the ideas, with a few changes such as choice of probability measure, are readily applied to the pricing context.

    First and foremost, however, our focus in this work is on alternative techniques to describe the default-loss distribution and associated summary statistics for portfolios of financial assets. This is the essence of all credit-risk modelling approaches and requires us to think carefully about the very nature of credit risk and how it can be effectively modelled with mathematical and statistical methods. Much can be learned about the challenges and techniques associated with addressing this question by examining the market-risk perspective. This is precisely our focus in the following section.

    1.2 A Useful Dichotomy

    The value of asset portfolios rise and fall with the passage of time. These changes in value can occur for a variety of reasons. It is useful, however, to decompose these movements into two (very) broad categories:

    1.

    general market movements making certain groups of assets more or less valuable; and

    2.

    idiosyncratic movements in specific assets that lead to relative changes in their individual value.

    This disaggregation is not new; it has been a key element in the finance literature for decades. Indeed, this is the intuition behind the celebrated capital asset-pricing model (CAPM). Much has been written on the CAPM; a good start, for the reader interested in more specifics, would include Sharpe (1963, 1964); Treynor (1961) and Lintner (1965). The CAPM is not used explicitly in credit-risk models, but it appears to motivate some of the fundamental ideas.

    The distinction between general, or systematic, and specific, or idiosyncratic, risks is an immensely useful breakdown. Let us consider a few examples. When the general level of interest rates rises in an economy, it affects many assets. This is a general, systematic movement. The value of fixed-income assets fall and, depending on the context, it may lead to a positive upward shift in equity prices. If, conversely, a particular company has internal difficulties, for reasons that are unique to the firm, it typically has a more limited impact. The value of its debt and equity claims may fall, but we would not expect a concurrent impact on the value of other assets in the economy. This is a specific, or idiosyncratic, type of risk.

    These different types of risk have different implications for how we model their associated portfolio-value movements. Moreover, in a very broad sense, these two perspectives capture the distinction between market and credit risk. To maintain that market risk arises from systematic factors and credit risk from idiosyncratic elements is sweeping generalization. It does not always hold. Some aspects of market risk can be quite specific to industries or regions, whereas credit risk can arise from general macroeconomic trends or events. There is, however, some truth to this statement and it offers a useful first-order approach to organizing the differences between these two key risk dimensions.

    Let us start by considering the market-risk setting in a bit more detail. Market risk describes the portfolio P&L implications associated with movements in underlying market-risk factors. A specific financial instrument can be influenced to different degrees by a broad range of possible risk factors. The same risk factor, however, can impact multiple classes of financial instruments. Movements in these factors can lead to gains or losses in an asset portfolio. Ang (2013) offers a clever and helpful analogy of the link between asset classes and risk factors as strongly resembling the connection between food and nutrients. Portfolios (meals) are constructed from combinations of asset classes (foods). Diet experts do not focus on foods, but rather on their underlying nutrients. Nutrients are what really matter in our diets and, analogously, risk factors are what really drives asset returns. As a financial risk-expert, one is expected to delve into what really determines price movements in one’s portfolios: risk factors.

    In a market-risk setting, these underlying market risk factors typically include interest rates, exchange rates, equity and commodity prices, market liquidity, volatility, break-even inflation and credit spreads. Broadly speaking, therefore, these market risk factors relate to general, or systematic, quantities. As indicated, it is not always quite this simple. Credit spreads tend to move in unison, for example, but there is often also an idiosyncratic element. Moreover, while a security’s value may depend on a general risk factor, its factor loading, or sensitivity to this factor, may be defined in a unique way.

    The credit-risk domain is, in many ways, more difficult to manage. Credit risk, as has been indicated, describes the inability of a credit counterparty to meet their obligation. The ability to pay is, in a general sense, influenced by a rather different array of factors including, but not restricted to: indebtedness and capital structure, success of business model, management ability, competition, ongoing litigation, and willingness to pay. While these criteria appear to be predominately specific to the credit obligor, many of them can, however, be influenced by general economic conditions. As such, they also possess, to some extent, a general or systematic element.

    It would thus be a mistake to suggest that market risk depends on a underlying factor structure, whereas credit risk does not. The use of observable or latent risk factors, or state variables, is an integral aspect of modern finance. They arise in term-structure, risk-management, macroeconomics, and strategic analysis. They are simply too convenient and useful—in terms of dimension reduction and the ensuing mathematical parsimony—to not employ them in multiple settings. Both credit and market-risk computations depend on underlying risk factors. The point, however, is that the nature of these factors are often quite different in a credit-risk setting.

    It is also an oversimplification to state that credit-risk factors are uniformly idiosyncratic in nature, but an important part of credit risk does indeed stem from specific factors. Figure 1.3 attempts, in a schematic and illustrative manner, to summarize this key message. Market risk is predominately systematic in nature, whereas much of credit risk arises from specific, idiosyncratic factors. The intersection of these two types of risk is not empty. There is an overlap of market and credit risk. It is sufficiently small that it is expedient to treat them separately; hence, the separation of risk-management functions into credit- and market-risk departments. It is nonetheless sufficiently large that it cannot be completely ignored.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig3_HTML.png

    Fig. 1.3

    Market vs. credit risk: Credit and market risk are not completely independent, but they are typically treated separately.

    1.2.1 Modelling Implications

    Underlying driving risk factors are not the only difference between credit and market risk. In the market-risk setting, we often think of value changes to instruments and portfolios as occurring incrementally in a gradual fashion. Risk managers are, of course, worried about large jumps in market-risk factors leading to catastrophic losses. Such events certainly do occur, but not often. To this end, market-risk managers frequently employ stress-testing techniques to assess this type of risk.

    Most market-risk computations, implicitly assuming continuous and gradual risk-factor dynamics, make use of a rather different modelling approach. In particular, it is common to assume that market-risk factors follow a stochastic process with continuous sample paths. Geometric Brownian motion is one of the most common choices. An examination of many of the key market-risk sources—for example, Mina and Xiao (2001); Morgan/Reuters (1996), and Mina (2005)—reveals the fundamental aspect of this assumption. This amounts to describing the infinitesimal dynamics with a stochastic differential equation of the following form,

    $$\displaystyle \begin{aligned} \begin{array}{rcl} {} \text{Change in Factor} &\displaystyle =&\displaystyle \text{Drift} + \text{Diffusion},\\dX_{t} &\displaystyle =&\displaystyle f(X_t)dt + g(X_t) dW_t, \end{array} \end{aligned} $$

    (1.1)

    where X t is defined on $$(\varOmega ,\mathcal {F},\mathbb {P})$$ , f(X t) and g(X t) are measurable functions and {W t, t ≥ 0} is a standard Wiener process. A key feature of the Wiener process is that,

    $$\displaystyle \begin{aligned} \begin{array}{rcl} {} W_t - W_s \sim \mathcal{N}(0,t-s), \end{array} \end{aligned} $$

    (1.2)

    for all t s.⁸ In other words, the increments of a Wiener process are independent, and identically normally distributed with a variance that is directly proportional to the size of the time step.

    What is the rationale behind this choice? The path of a Brownian motion is noisy, bumpy, but ultimately continuous. Moreover, Brownian motion is inherently unpredictable. The same can be said, most of the time, for the movement of market risk factors. Changes in exchange rates, interest rates, and credit spreads (typically) occur gradually over time—they are also very difficult to predict. Brownian motions, with the appropriate parametrization, are also readily correlated. This is highly convenient for market-risk factors, which are generally expected to possess some degree of dependence or interaction. The implicit assumption embedded in the use of this approach is that market uncertainty is reasonably well described by a collection of correlated Brownian motions.

    Is this a reasonable assumption? Figure 1.4 tries to motivate this choice. It displays two graphics: one displaying the actual daily changes in the USD-JPY exchange rate, a market-risk factor, and another simulated from a theoretical Brownian motion. Which is the real, observed risk factor? Ultimately, it does not matter and will remain a secret, since visually these two processes appear essentially the same. Does this mean that market-risk factors are equivalent to geometric Brownian motions? No, not at all. It does, however, motivate why financial modellers opted, and continue to opt, for geometric Brownian motion as an approximation to reality. There is a significant amount of evidence indicating that the inherent Gaussianity and continuity of geometric Brownian motion are not perfectly suited for financial markets. This has led to numerous extensions and improvements—such as through the addition of stochastic-volatility, imposition of more complex transition densities with heavier tails, or jump-diffusion models—but the choice still has significant conceptual merit.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig4_HTML.png

    Fig. 1.4

    Brownian motion in action: Examine the underlying risk factors. Which is a real, observed market-risk factor and which is a simulated path from a Brownian motion?

    1.2.2 Rare Events

    Default risk is also unpredictable. It rarely occurs, however, in a continuous and gradual manner. There is typically some element of surprise in a default outcome. Moreover, defaults are, for the most part, rare events. Precisely how rare and surprising a default event is considered, depends, of course, on the credit quality of the underlying obligor. The default of a highly creditworthy and financially solid counterparty would be considered surprising. As a consequence, such an event would typically be assigned a low probability. Default of a highly risky credit counterpart with poor creditworthiness, conversely, would probably not be considered such a rare and surprising event.

    Although the degree of rarity and surprise depends on the credit quality of the obligor, the nature of the default event is inherently different than in the market-risk setting. Changes in a portfolio’s value occur constantly, as market-risk factors move, and may have either positive or negative profit-and-loss implications. Default is a binary event. It either happens or it doesn’t. Once it happens, default is also typically a permanent state. After arrival in default, exit is, in general, a slow and messy process. This is a fundamentally different type of risk.

    This does not necessarily imply that continuous sample-path stochastic processes cannot play a role in the credit-risk setting. The canonical credit-risk reference, and genesis of much of the current state of credit modelling, Merton (1974), is built on a foundation of geometric Brownian motion. In this approach, a firm’s asset-value process, {A t, t ≥ 0}, evolves over time according to geometric Brownian motion. K is the lower bound, or threshold, denoting the value of the firm’s outstanding liabilities. If, or when, A t K for the first time, default is triggered. After this point, the firm is considered to be in default. Figure 1.5 graphically illustrates this situation.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig5_HTML.png

    Fig. 1.5

    Brownian motion in credit risk: Geometric Brownian motion can, and often does, arise in the credit-risk setting. It plays, however, a rather different role.

    This is, in technical vernacular, referred to as a first-passage time. That is, in the Merton (1974) model, the first time that the firm’s asset value falls below its liabilities, default occurs. While this remarkably useful intuition supports much of credit-risk modelling, a few points are worth mentioning. First of all, it is two rather different things to assume that over a predefined time interval:

    1.

    our investment’s value changes with the evolution of a stochastic process; and

    2.

    we experience a significant loss on our investment if the same stochastic process falls below a particular threshold.

    The nature of the financial risk—and, by extension, the modelling tools used to estimate this risk—is different in these two cases. We should expect links between the credit- and market-risk dimensions, but there will also be important contrasts.

    The second critical point is that, from a default perspective, the process may never cross the barrier over a particular time interval; or, if it does, it occurs only hundreds of years in the future. This brings us back to the binary aspect of credit risk. Default events are rare. Market-risk movements, by way of contrast, occur virtually at every instant in time. For some classes of assets—such as credit cards and consumer loans—default losses may, of course, occur with a high degree of regularity. Defaults may also be observed more frequently with car loans and mortgages, but since they are backed with assets, the magnitude of the losses are typically less important. In a more typical investment setting—corporate loans, derivative contracts, insurance policies, and high-grade investments—default typically occurs with much less frequency.⁹ At a one-year horizon, for example, the probability of default for investment-grade debt approaches zero. Indeed, it is still only a few percent for non-junk bonds below the invest-grade line. Default can, therefore—from a modelling perspective, at least—be considered to be a rare, low-probability event. In these cases, it is usually considered to be a surprise.

    It is important to appreciate this point, because it has two important implications. First, the structure of credit-risk models will differ in fundamental ways from the models employed in a market-risk setting. This idea is illustrated, in a stylized manner, by Fig. 1.6. Market risk is, generally speaking, the combination of numerous, symmetric and systematic risk-factor movements, whereas credit risk is associated with low-probability events of large magnitude. The result is two rather different loss distributions. Alternative tools, concepts, and methods will, therefore, be required in each of the two settings. Conceptually, however, the task is the same: combining profit-and-loss outcomes along with an assessment of their relative likelihoods to estimate the magnitude of total risk.

    ../images/458636_1_En_1_Chapter/458636_1_En_1_Fig6_HTML.png

    Fig. 1.6

    A stylized comparison: Due to their fundamental differences, profit-and-loss outcomes and their associated likelihoods are combined in alternative ways in the market- and credit-risk settings.

    The second aspect has a larger impact on the analyst’s life. The rarity of default events implies a shortage of useful data for the estimation and calibration of model parameters. A multitude of market risk-factor movements are observed every single trading day. Conversely, only a handful of high-quality credit defaults are observed over the course of a decade. This creates a number of challenges—many of which are addressed in the coming chapters—and represents a core reality for the credit-risk modeller.

    1.3 Seeing the Forest

    Our principal perspective is that of a risk manager. In this context, we will seek to perform credit Value-at-Risk, expected shortfall or, almost equivalently, economical-capital calculations for an asset portfolio. Whenever one selects a specific mathematical model to compute metrics or perform important analysis to take key decisions, it should not be done lightly. Analysis is required to assess the wisdom of the specific choice. At a minimum, it requires the analyst, and his superiors, to examine the key aspects of the modelling framework and to opine on the reasonableness of the overall approach. This is easier said than done. An examination of the credit-risk literature reveals a broad range of competing models and a high-degree of complexity.

    To make matters a bit worse, in most credit-risk settings, empirical questions cannot be easily answered using statistical techniques. The principal reason, as already mentioned, is a dearth of default-event data. The limited amount of available data will not take us extremely far. Its principal use is in the estimation of model parameters and we will demonstrate how, in many cases, this can be very challenging.

    Compare this to the market-risk world. Market-risk managers compute quantile measures, such as Value-at-Risk, of portfolio profit-and-loss movements on a daily basis. They may then compare these estimates to actual observed valuation shocks. If they are patient enough to collect a few years of data, there are a battery of useful statistical tests that can be usefully employed to assess the reasonableness of one’s estimates. This exercise, known as back-testing, is an invaluable part of an analyst’s toolkit.¹⁰ It provides comfort and perspective in selecting a specific model; back-testing can also play an important objective role in model selection and calibration. Credit-risk analysts, due to the shortage of default data, cannot typically use back-testing techniques. Credit-risk modellers must thus learn to manage a correspondingly higher degree of subjectivity.

    Critical assessment, and a general aversion to subjectivity, is nevertheless an essential part of the model-selection process. We seek to build this directly into our discussion rather than try to layer in onto our thinking at the end. We thus employ alternative techniques and metrics to assess model appropriateness. In particular, our analysis will highlight key modelling decisions, discuss alternatives, and perform analysis to motivate and challenge each choice.

    The analyst’s model-selection challenge is thus an important motivating factor behind the organization of this book. There is, of course, a significant amount of material to cover to help the quantitative analyst, risk manager, or graduate student to become more acquainted with this area and help them take defensible choices. To facilitate this, the following chapters are organized into three distinct themes: modelling frameworks, diagnostic tools, and estimation techniques. Table 1.1 summarizes the organization of the following chapters.

    Table 1.1

    Topic overview: This table illustrates topics addressed in the following chapters and technical appendices.

    1.3.1 Modelling Frameworks

    Our thematic categorization is intended to inform three distinct questions. The first question, associated with the initial theme, is: which credit-risk model should one use? There is, naturally, no single correct answer for all applications. It will depend upon one’s situation and needs. We will try to help by allocating four chapters to the consideration of three alternative modelling frameworks including:

    1.

    an independent-default model;

    2.

    a dependent-default reduced-form approach; and

    3.

    a variety of alternatives within the structural setting.

    There are numerous interlinkages between these models, which will help the analyst better understand the collection of extant models. We will not cover every possible modelling approach, but hope to provide a sufficiently broad overview so that further models will be easily accessed and classified.

    There are a number of common elements, albeit handled in different ways, among these three modelling approaches. The first aspect is default dependence. Dependence between default events is a critical aspect of this discussion. Dependence of risk-factor movements is almost always an essential aspect of financial risk management, but in many ways its influence is even larger for the credit-risk modeller than in the market-risk setting. The reason is that the addition of default-dependence, induced in alternative ways, has an enormous influence on the ultimate default-loss distribution. Understanding the relative advantages and disadvantages of one’s model choice, therefore, basically reduces to confronting issues of default dependence.

    We also consider how the heterogeneity of exposures and default probabilities interact with the choice of model. In general, as the portfolio increases in size and homogeneity, the better behaved the model, the easier the estimation effort, and the stronger the performance of analytic approximations. This is an important consideration for the many organizations facing, for underlying structural business reasons, portfolios with concentrated and heterogeneous exposure and credit-quality profiles. Diversification thus inevitably also plays a role in the credit-risk setting.

    It is also natural to ask how many factors are required, or are desirable, in a credit-risk model. The choice ranges from one to many and represents the classical trade-off between parsimony and goodness of fit. Goodness of fit is a tricky criterion for model selection in this setting, however, given the lack of data for calibration. There is some inevitable circularity, since with various parametric choices, it is typically possible to calibrate a one-factor model to yield broadly similar results to its multi-factor equivalent. Logical considerations based on modelling advantages and disadvantages thus play an important role in the ultimate model selection.

    The advantages of a one-factor approach relate principally to the availability of a broad range of analytic formulae for its solution and computation of ancillary model information. On the negative side, many one-factor models may just be too simple to describe more complex credit portfolios. Multi-factor approaches may, in certain settings, provide a higher degree of realism and conceptual flexibility. The ability to categorize default-risk and dependence by, for example, industrial or regional dimensions is useful and valuable. Indeed, specific organization of one’s factor structure provides a powerful lens through which one can consider, discuss, and take decisions on one’s risk profile.

    More factors imply, however, a higher degree of complexity. Typically, for example, model computations are performed through stochastic simulation. This can be either complex, computationally expensive or both. Moreover, the larger the number of factors, the broader the range of possible model parameters. These parameters must be either estimated from historical data, calibrated to current market data, or selected through expert judgement. In virtually all cases, this is a non-trivial undertaking. We will work, in the coming chapters, predominately with one-factor models due to their greater ease of exposition and illustration. We will, however, also explore multi-factor implementations to provide critical perspective. Often, the results generalize quite naturally, but there are also unique challenges arising only in multi-factor situations. Where appropriate, these will be highlighted and discussed.

    The marginal risk-factor distributions in these models may take alternative forms. Irrespective of the number of model factors selected and their distributions, a joint distribution must be imposed to proceed toward economic-capital calculations. The simplest choice would be to assume independence, which would permit straightforward construction of the joint default distribution. This is equivalent to an independent-default approach where each default event is treated as a Bernoulli trial implying, in the limit, a Gaussian loss distribution.¹¹ Much can be gleaned from this approach, but ultimately it is necessary to add default dependence. As dependence is introduced, even in the simplest case of a Gaussian joint distribution for the underlying risk factor (or factors), the loss distribution is no longer Gaussian. In fact, it is highly skewed and heavily fat-tailed.

    The specific distributional choice has, not incidentally, important implications for the size and dependence of outcomes in the tail of the default-loss distribution. Tail dependence describes the probability, as we move very far out into the tail of a distribution, that a given counterparty defaults conditional on the fact another counterparty has already defaulted. This is an essential feature in a credit-risk model, particularly since one routinely uses the 99.50th or 99.97th quantile of the loss distribution to make economic-capital decisions.

    Chapters 2 to 5 will thus explore a variety of credit-risk models and, in each case, consider dependence, distributional, factor, and portfolio homogeneity considerations. The objective is to provide a useful frame of reference for selecting and defending a specific modelling choice.

    1.3.2 Diagnostic Tools

    Once a specific credit-risk model has been selected and implemented, the work of the conscientious analyst is far from over. Models occasionally break down or, in some circumstances, they can provide conflicting or non-intuitive results. This seems, from personal experience, to typically happen at the worst possible time. The question addressed by the second theme is: how can we mitigate this risk? Sanity checks and diagnostic tools are helpful and necessary in resolving breakdowns or explaining incongruities in model outputs. Moreover, it is also essential to understand and communicate modelling results to other stakeholders in one’s organization. Understanding and communication are intimately linked. Without a strong intuition of why and how one’s model works, it is a tall order to communicate its outputs to one’s colleagues or superiors. These two perspectives argue strongly for the inclusion of model diagnostics in one’s modelling framework. We find this sufficiently important to allocate three chapters to this notion.

    Credit-risk modelling is not performed in a vacuum. Most institutions performing credit-risk management operate under the control of regulatory regimes. Chapter 6 explores the regulatory guidance from the Basel Committee on Banking Supervision’s (BCBS) Accords with an eye to link the main ideas to the modelling discussion found in Chaps. 2 to 5. While the ideas may seem relatively simple, they form a critical backdrop for other credit-risk modelling activities. The Basel Internal-Ratings Based approach is constructed based on an ongoing careful review of existing models. A thorough understanding of the regulatory framework will help one better understand, and compare, existing choices to the decisions made by the BCBS. It will also help analysts answer inevitable questions from senior management on the differences between the internal and regulatory model results. These issues are the focus of Chap. 6.

    Credit-risk economic-capital is, up to a constant, a quantile of the credit-loss distribution. This is a useful measurement, but others, such as the expected shortfall, are also possible. In isolation, these measures are nonetheless silent on the contribution of each obligor, or perhaps region or industry, to the total risk. Decomposing a specific risk measure into individual contributions along a predetermined dimension is referred to as risk attribution. It is reasonably well known that, in general, the risk-metric contribution from a given obligor is the product of the counterparty (or regional or industry) weight and its corresponding marginal-risk value.¹² This marginal quantity is only available in analytic form, however, under certain model assumptions. If one deviates from these assumptions, it may still be computed, but one must resort to simulation methods.

    Despite these challenges, risk-metric decompositions are an invaluable tool. They are critical for trouble-shooting model results by permitting high-dimensional comparison to previous periods. They also aid in understanding how the results vary as model inputs—such as parameter values, default probabilities, and exposures—are adjusted. Finally, they are indispensable for communication of results to other stakeholders. Their regular computation and use should ultimately permit better decision making. In short, computation of individual obligor contributions to one’s risk metrics may be the most powerful diagnostic at the modeller’s disposition. Chapter 7 addresses this topic in detail.

    It is almost inevitable that, in the course of one’s credit-risk model implementation, one will make use of Monte Carlo methods. Credit-risk metrics typically involve taking a value from the very extreme of the default-loss distribution, such as the 99.9th percentile. When this rare event is estimated using stochastic simulation, it is natural and healthy to ask how many simulations are necessary to provide a reasonable and robust approximation. Such an assessment requires significant analysis and diagnostic tools. In particular, the following quantities prove instrumental in answering this

    Enjoying the preview?
    Page 1 of 1