Credit Correlation - A Guide

Credit Derivatives Strategy
London 12 March 2004
Credit Correlation: A Guide
First principles in defining correlation

Correlation in Credit requires special care
Credit Derivatives Strategy Lee McGinty*

(44-20) 7325-5482 lee.mcginty@jpmorgan.com
Eric Beinstein
Estimating default correlation from equity prices is simple and convenient
(1-212) 834-4211 eric.beinstein@jpmorgan.com
But using default correlation from credit spreads may CDO/Credit Derivatives Strategy have advantages Rishad Ahluw alia
(44-20) 7777-1045 rishad.ahluwalia@jpmorgan.com
We introduce Base Correlations: JPMorgans innovative approach to reporting correlation implied by tranche spreads
Key advantages over alternative approaches: A unique solution More meaningful results Enables pricing of off-the-run tranches
Derivatives Research Martin Watts

(44-20) 7777-3881 martin.r.watts@jpmorgan.com
Table of Contents
Credit Correlation: A Guide Overview Section 1: Correlation in the Credit Market Section 2: Estimating Default Correlation Section 3: Correlation Trading Basics Appendix 1: Implied Volatility in Equities versus Base Correlation in Credit Appendix 2: Correlation and CDO Rating Methodologies 1 2 3 13 21 30 33
The certifying analyst(s) is indicated by an asterisk (*). See last page of the report for analyst certification and important legal and regulatory disclosures.
http://mm.jpmorgan.com
Lee McGinty* (44-20) 7325-5482 Eric Beinstein (1-212) 834-4211 Rishad Ahluwalia (44-20) 7777-1045 Martin Watts (44-20) 7777-3881
Credit Derivatives Strategy London 12 March 2004
Overview
Correlation has been used for many years in finance to quantify a relationship or a dependence between two or more variables. In some markets, such as equities and interest rates, the use of correlation is well-developed and statistically sound. However, correlation is being used in new areas such as the credit market, where the underlying assumptions required for a calculation to be valid are easily violated. Consequently, any scenario analysis, risk management, or model price based on these correlation assumptions may be misleading. Thats not to say that correlation analysis and trading in the credit market is invalid, rather that as market participants, we think it prudent to reacquaint ourselves with some of the basic assumptions. To this end, our research note is divided into three sections. Section 1 defines and explains correlation in general and introduces it to the credit market specifically, although not just in derivative pricing models. This section explains in non-technical terms where it can be safely used and, more importantly, where it cannot. Significantly, Section 1 also introduces some rules of thumb for using correlation through analogies from other asset classes, where appropriate, and touches on other methods for understanding relationships between variables, such as copula functions. Section 2 introduces and compares methods for estimating default correlation from historical defaults, equity returns, asset returns, and credit spreads. We use Ford and General Motors as an example where one can get different correlations from these methodologies (which are often confused with one another), and open the discussion as to whether it might make sense to start using credit spreads (rather than equity returns) to estimate default correlation as the CDS market develops. Finally, Section 3 discusses some practical implications of using correlation for trading and managing credit risk, focusing on implied correlation in standardised synthetic CDOs. This section also introduces a simple model for reporting so-called Base Correlations in tranches, the Large Pool Model, and explains some of the problems associated with using the alternative approach which has been suggested by some partiesCompound Correlation. This piece is designed to cover the background needed to properly analyze correlation in the credit market. There is some discussion on trading and pricing in Section 3, but a much bigger focus on the fundamental issues that we believe have been, to some extent, glossed over in the market place. We expect to publish a more comprehensive piece on pricing and hedging correlation in structured credit after this explanatory note.
Section 1: Correlation in the Credit Market

Introduction
Correlation has been used for many years in finance. It is central to the Markowitz portfolio theory first published in the 1950s and has been extensively used in equity markets to understand the behaviour of portfolios and indices. Correlation has also been used in derivative products, where the payoff depends on more than one underlying instrument (for example, interest rate spread or equity basket options). However, there is a danger that techniques, which are perfectly reasonable and valid for other asset classes, are being transplanted into the credit market, without a full understanding of the implications.
What is Correlation?
Ignoring mathematics for a moment, correlation is just the strength of a relationship between two or more variables. In this original intuitive sense, no mention is made as to the kind of relationship. For example, its well understood that a car decreases in value as it ages, so its fair to say that, for a given car, age and value are correlated. However, a cars value depreciates rapidly over the first few years (particularly on the day it becomes no longer new), then much more slowly over the next few years.
Chart 1: Car Value as a function of Age
Illustrative value as percentage of price when new (y axis), age in years (x axis)
100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% 1
Source: JPMorgan.
12
16
20
This is an example of a non-linear correlation (Chart 1); a linear analysis of this relationship may well report little or no correlation. This linear versus non-linear distinction is important to remember when discussing correlation in any data set, whether in cars or credits. For the purposes of this research note, we will assume correlation means linear correlation. There are other techniques such as the Spearman or Kendall rank correlation coefficient, which is often used as a method to overcome some of the deficiencies of the standard linear correlation calculation (it uses the statistical rank of the variables when the distribution of the original variables may make the standard linear correlation coefficient misleading), but we will not consider this method further as it is infrequently used.
The most common method of calculating correlation is the linear correlation coefficient (alternatively, the product moment correlation coefficient, or the Pearson correlation coefficient). Mathematically, linear correlation is often represented by the Greek letter (rho), and given by this formula:
(X
n i =1
X Yi Y
2 n
)(
)
2
(X
n i =1
) (Y Y )
i i =1
Where X and Y are variables (the two assets being compared), and hence Xi, Yi each observation, and observations.
X and Y the mean of each variable. Each variable has n
Intuitively, one can say that the correlation coefficient determines how much the values of two variables are proportional to one another. The correlation value is scaled into a range between +1 (perfect positive correlation) to 1 (perfect negative correlation) and hence does not depend on the specific measurement units used (for example, the correlation between height and weight will be the same irrespective of whether inches and pounds or metres and kilograms are used). Proportional means linearly related, that is how much can the relationship be approximated by a straight line. Interpreting Correlation There have been several attempts in academic literature to understand correlation coefficients, but it is by no means straightforward to come up with an intuitive description. At the simplest level, correlations close to +1 will be very closely related, with observations close to a straight line. Correlations near zero will be fairly random, and correlations near 1 will be highly correlated, but the line of the relationship will have a negative slope. However, this doesnt help answer the sensible question of what does a correlation of say 50% mean? One suggestion has been to put correlations into bands, with less than 20% meaning no or negligible correlation, through to 80% or more to mean high correlation (A. Franzblau, 1958). Another scholar suggests that as a rule of thumb, we can say that correlations of less than 30% indicate little if any relationship between the variables (Hinkle, Wiersma & Jurs, 1988). Other possible intuitive ways to think about correlations involve manipulating the variables first. If we convert the variables to z-scores (how many standard deviations they are from their means), then the correlation is the slope of the regression line (Chart 2). Alternatively, if we square the correlation then we have a slightly different measure (the r squared, or coefficient of determination), which tells us how much of the variation in the first variable is explained by variation in the second variable. That is, for a 50% correlation, we get 50% 50% = 25% of the variance of the first variable is explained by the variance of the second.
Chart 2: Scatter charts of different linear correlations

Two normally distributed random numbers, X (x axis) and Y (y axis)
0% correlation 4 2 0 -5 -2 -4 0 5 -5 50% correlation 4 2 0 -2 -4 0 5 -5 90% correlation 4 2 0 -2 -4 0 5 -5 100% correlation 4 2 0 -2 -4 0 5
Source: JPMorgan.
Correlation Appropriateness Even armed with the above information, its still possible to misuse correlation. One could take any two data sets, say equity price histories of a regional US bank and a Japanese retailer, and run them through a correlation calculation. The calculation always returns a number, but that doesnt mean its always valid to examine the correlation. (One of the most common errors is to just do a correlation analysis and draw conclusions from the results without considering whether there is any causal relationship between the data.). Lets say the above example generates a correlation of 70%. Its a big leap solely from this data to believe that theres a strong relationship between the two entities. A correlation measure doesnt tell us anything about one variable causing changes in another. This correlation of 70% may have come about from nothing more than pure coincidence, and before even making those calculations we should have considered whether theres a causal relationship. Perhaps there is one, but it should be investigated first. Whilst it is possible to apply the correlation calculation to any data set, the correlation measure is only defined if the variables are normally distributed. Normal distributions are the familiar bell-shaped curves symmetric around a mean that are used to analyse biological, socioeconomic, and some financial data. A mean and a standard deviation (Chart 3) is enough to describe normally distributed data.
Chart 3: Standard Normal Distribution
Standard deviation (x axis), probability (y axis)
-4
-3
-2
-1
0 mean
- 1 standard dev iation
+ 1 standard dev iation
Source: JPMorgan.
Correlation in Credit
Unfortunately, typical credit returns do not fit a normal distributiona critical point, as it complicates things when running a correlation analysis in credit instruments, be they bonds, credit default swaps, or the like. This is one of the reasons why the original CreditMetrics model and most other accepted ones use the correlation of asset returns (or as a proxy, equity returns), which are closer to being normally distributed. Also, this is the reason why CDO tranches have implied asset return correlations embedded in their prices. Sections 2 and 3 explore these issues in much more detail, but first, we compare equity and credit return distributions as a primer. Equity Return Distributions Chart 4 shows actual daily percentage equity returns of a bellwether US company General Motorsover the last 5 years, and a fitted normal distribution. The returns are not exactly normally distributed, but theyre fairly close, and importantly, theyre symmetric around a mean. While its possible that an equity price can fall to zero in times of distress, its also possible for it to double or triple in price, resulting in a reasonably symmetrical distribution of returns over a period of time. A mean and a standard deviation are generally sufficient to understand the behaviour of GMs equity returns, and generally speaking, equity market participants are typically more interested in the middle part of the distribution (close to the mean) to predict future price behaviour with some degree of confidence.
Chart 4: Equity Return Distribution
Daily percent change (x axis), frequency (y axis)
-8 -7 -6 -5 -4 -3 -2 -1
Equity Return Distribution
Normal Distribution
Source: JPMorgan.
Credit Return Distributions Unlike in equities, the returns from a typical credit instrument will not be normally distributedthey are inherently skewed. While both credit and equity prices can fall dramatically in the case of a default, typically a credit instrument has upside limited by a known maximum payout of the principal amount plus interest.
For example, consider a corporate bond. Even though the probability of default is relatively low, the fall in price from the current state to recovery value could be very large, producing a relatively large probability-weighted loss. However, the maximum upside is limitedespecially in the case of a highly rated bond. This produces a skewed or fat tailed distribution shape for the returns of typical corporate bonds. This is also true of credit default swap (CDS) returns, as there is a limit on how low the spread can trade, but if the underlying name is close to default, the spread can widen out to several thousand basis points. Note that this is also true for the typical portfolio of corporate bonds or CDS as even a very small correlation between defaults will produce a significantly skewed return distribution. It is often suggested that a large enough portfolio of credits will start to look normal, or perhaps normal enough. We argue that even for a portfolio consisting of all liquid bonds (an index) this doesnt occur. To see this visually, we could test for the normality of credit spreads to get a sense of the distribution of credit returns. Chart 5 below plots actual daily spread volatility versus spread level for a large number of credits. If spread changes for this sample were fairly close to being normally distributed, we wouldnt get the pattern in Chart 5, but observations within a horizontal band indicating an equal likelihood of spread movements for any absolute level of spreads. It is cleaner to test for normality through spreads, but we can also conclude that credit returns are neither normal nor log normal as there is a simple linear relationship between spread changes and returns. For more insight into distributions in the credit market as well as more details on testing for normality, please see Credit VolatilityA Primer by Lee McGinty and Martin Watts, June 2003.
Chart 5: Scatter chart of absolute credit spread changes
Absolute daily spread change in bp (x axis), level of spreads (y axis)
100 90 80 70 60 50 40 30 20 10 0 0%
Source: JPMorgan.
2%
4%
6%
8%
10%
So, while it may be reasonable to use a normal distribution to estimate equity returns, it is not suitable for returns on corporate bonds or other credit instruments such as credit default swaps. An estimate of the returns from a credit instrument using just a mean and a standard deviation could produce very misleading results as the returns are so skewed. Estimating a joint distribution So far, weve only tested the closeness of an individual equity or credit distribution to a normal distribution, which is relatively straightforwardwe only tested a single name. For a correlation calculation to be valid, we also need to check whether the joint distribution is normal, which is more difficult; the actual process is less well
7
defined, and the number of checks is much greater as all combinations of the variables must be analysed. While it may be difficult to perform all of these tests, at least some worst case relationships should be analysed. When we are talking about many variables, we use the terms joint distribution to refer to the process by which the variables behave together and marginal distribution to refer to any of the individual distributions. The easiest way to conceptualise a joint distribution is to consider only two variables. Plot a scatter diagram of the two variables and look at the shape of the resulting cluster of dots. In general terms, if the cluster of dots has an elliptical shape, then the variables have a joint normal distribution and its valid to use correlation as a measure of dependence between the two variables. As the correlation between these two variables increases, the dots will converge and become a single line (the ellipse in Chart 7 becomes a line). However, if the shape is significantly different from an ellipse, especially with a high concentration of dots in any one area (i.e. clustering), then it may not be safe to use a single correlation number. Note that these are higher correlations than typically seen in the credit market, but they are increased to more clearly show the relationships.
Chart 6: Joint normal scatter charts, correlation = 80%
Source: JPMorgan.
Chart 7: Joint normal scatter charts, correlation = 95%

Source: JPMorgan.
Now consider a joint relationship that is not normal. Using an example of joint credit default probabilities, Chart 8 shows the same joint normal distribution as above, with a correlation of 80%. Chart 9 also has a correlation of 80%, but a distribution thats more typical for creditit is clear visually that the relationship between the two variables is quite different. Look at the top-right corner of Chart 9: theres a more concentrated pattern for higher values than in Chart 8. However, there are also more outliers and a wider dispersion in the rest of Chart 9 than in Chart 8 (Chart 9 has a star rather than elliptical shape). The type of relationship illustrated in Chart 9 corresponds to anecdotal evidence that default correlations between names is low in a normal market environment, but if the probability of default increases for one name, then it is likely to increase for the others as well, such as during a recession (a common pattern in the credit market is for a star or teardrop shaped pattern of defaults).
Chart 8: Joint normal credit distribution (illustrative), correlation = 80%
Probability of default for Credit X (x axis), probability of default for Credit Y (y axis)
Source: JPMorgan.
Chart 9: Typical joint distribution for credit (illustrative), correlation = 80%

Probability of default for Credit X (x axis), probability of default for Credit Y (y axis) correlation 3x higher than joint normal
Source: JPMorgan.
The area outside the dotted rectangle in Chart 8 and Chart 9 contain data observations assumed to be in usual market conditionsaround 90% of the total observations. The correlation between the two sets of variables outside these dotted areas is the same in both charts. The dotted box is where one or both of the variables experience a large move, i.e. the probability of default has increased. Here, the correlation of the
9
data in the top-right corner in Chart 9 is three times higher than in Chart 8 (the dots are more clustered or squeezed-in at the extreme). We can see that there is an increased likelihood of both names defaulting as the probabilities of default both increase at the same time. The correlation between the two variables in both charts is the same if we consider the whole data sample at once (i.e. the average correlation), or in other words, correlation is not a unique measure of the relationship. However, if we concentrate on the part of the distribution where the two names are close to default, which in general is the most important area, we find that the single, average correlation number can be quite misleading.
Avoiding Correlation Breakdown

Correlation Breakdown is a term used to describe what happens when the main assumptions necessary to be able to calculate a correlation coefficient (is the data normally distributed, is there a possibility for causality, is the relationship linear, etc.) are not met to such an extent that the correlation calculation is invalid. It is valuable to understand something of the behaviour of the data series before calculating and using a correlation number to try to avoid correlation breakdown, or at least to understand the limitations of results. The following are suggestions for testing for correlation breakdown and adapting the standard correlation calculation to better reflect the characteristics of data in the credit markets. When testing for correlation breakdown, the first stage is to consider whether there is any possibility for causality between variables. An example would be the Japanese retailer/US regional bank calculation in the beginningthat high correlation number of 70% may have come about from pure chance. Where feasible, check the structure of the data by looking at a scatter chartfor example, among the extreme names in a portfolio. As explained earlier, correlation is only a valid measure of dependency if the scatter chart has an elliptical shape. Correlation would be a poor measure of dependency when the shape of the scatter plot is more star or teardrop shapedrefer to Chart 9 (the star-shaped pattern which may make linear correlation misleading). Also, it can be useful to split the data period into 3 or 4 sub periods and use different symbols for each sub period to highlight behaviour changes. The data sample used should be as representative as possible. A good sample should contain enough data to be statistically valid, typically a minimum of 6 months of daily data or longer if using weekly or monthly data, though not too much so that the data is no longer relevant to the current trading environment. The key point is that the correlation being estimated is actually the future correlation, so we should use a historical sample that is representative of how the market is likely to trade going forward. In addition to calculating correlation over the whole of the data period, we can calculate a rolling correlation (in much the same way as a historical volatility) to see how the correlation varies over time. The correlation can then be compared with the two variables to see if there is any relationship (i.e. if correlation tends to increase as the underlying variables increase). This is valuable as the correlation calculation takes no account of when the individual data occurs, or, in other words, it jumbles up all the data throughout the whole time period. By calculating a rolling correlation, this splits the sample into smaller pieces.
10
Finally, try estimating correlation for just bad or big moves. This can be done by either just extracting the bad moves from the data sample to produce a conditional correlation, or looking at the rolling correlation over the worst sub time period.
Other methods - Copulas

If none of the above provide a reasonable estimate of a relationship being investigated, then another method which has become popular is to use a copula function. Simply, a copula is a way to model relationships between variables by splitting the definition of their marginal distributions (i.e. is it a normal, student-t, chi-squared etc.), from their joint relationships. The copula function is the part that defines the dependency between the variables, and this can vary from the very simple (independence or dependence copula) to the more complex (non-linear gumbel or student copula). While a copula can provide an approach more suited to the (generally non-linear) credit market, the first decision that needs to be made is which one to use, as there are many different types, each with different characteristics. Additionally, each copula will have parameters that modify the relationship between the variables. These parameters can be difficult to calibrate when there is very little reliable market information. A full explanation of copulas is beyond the scope of this note but below we outline two typical ones and how they perform compared to the multivariate relationships shown earlier. We have already used a simple copula, the normal or Gaussian copula in Chart 6 to Chart 8 earlier. Recently its become common to describe the CreditMetrics market standard for credit portfolio correlation modelling and pricing of standardised CDO tranches (i.e. DJ Tranched TRAC-X) as this Gaussian copula. A Gaussian copula defines the dependency relationship between asset returns in a credit portfolio as multivariate normal, and that this normally distributed joint relationship is independent of the actual distribution of the individual asset returns (although theyre assumed to be normalsee Section 2). So, in theory, though it would be rather unusual in practice, its possible to have two random variables with non-normal marginal distributions (say a Student-t distribution), and then use a Gaussian copula to define the relationship between these two variables. A possible alternative step in structured credit pricing could be the use of a different copula to define the relationship between asset returns. As an example, a Student copula is often suggested as an alternative, as this increases the dependency between the underlying variables at extreme values (relative to a Normal or Gaussian copula) so this would seem to fit well with the intuition that defaults tend to happen together, such as during adverse economic conditions. The star-shaped joint scatter plot shown earlier in Chart 9 is produced by a Student copula function.
Conclusion
Correlation can be a useful process for analysing markets, but in the credit market especially it has its limits. The shape of the individual credit return distributions and relationships between the distributions can vary significantly from a normal distribution, probably more so than in other asset classes. Typically, most credit derivative pricing models have found ways around this problem. For example, the CreditMetrics model uses the correlation between asset returns rather than credit
11
market instruments as its correlation input (see next section). This is possible because asset returns are closely related to equity returns, which tend to be more normally distributed. Even so, an implied correlation skew is still observed across different tranches in a CDO (see Section 3), which suggests that asset returns are, to some extent, also not normal. We think the message of this section is that for any analysis of dependencies in the credit market, it is important to first know exactly what are the underlying variables that are being analysed. Second, the behaviour of this relationship should be understood before using linear correlation as the measure of the dependency.
12
Section 2: Estimating Default Correlation

Introduction
As observed in Section 1, modelling default correlation is complicated by the fact that there is actually very little empirical data on the real relationship between defaults. Credit portfolio modelling inextricably involves estimating an underlying process which is not easily observable, which makes it quite different to modelling in other asset classes. At the risk of repeating this point too many times, it is important to be clear when using the term correlation to understand exactly what the variables are whose relationship is being investigated. The correlation between defaults and the correlation of asset returns is often confused or thought to be equivalentwhich is generally not the case. This confusion arises because while an important factor in the pricing of tranched credit is the correlation between defaults, the input required for models such as the CreditMetrics model is actually the correlation of asset returns. Asset return correlation can then be converted to default correlation within the model. It is worth mentioning at this point that strictly speaking a default is a binary event a company has either defaulted or it has not. In order to estimate the correlation between defaults, we can either look at a large sample of actual defaults within certain time intervals, or investigate the relationship between the probability of default. It should be noted however that correlation of default probability is not the same as correlation of default. To illustrate this, consider the following example of two companies in the same business. Their probabilities of default are likely to be closely related over time. However, if one business defaults, it is plausible that due to the decreased competition, the trading conditions for the other business will actually improve, so it could be argued that the actual default correlation would be low. Different bankruptcy regimes may have an impact here. This outcome is most likely in Europe, where bankruptcy often results in liquidation not reorganization, but in the US, Chapter 11 allows companies in default to operate as a going concern while restructuring their operations and balance sheet. While in bankruptcy, companies do not pay interest on their debts and are allowed to reject above market contracts. Essentially, these things serve to reduce the cost base of the bankrupt company, which enables them to compete in the market place more effectively. This in turn creates competitive pressure, which increases the likelihood of default among other market participants. Hence, we cannot assume that the correlation between default probabilities is exactly equivalent to the correlation between actual defaults. Our main focus in this section will be on the correlation of default probabilities. There are several ways to consider historical default probabilities and hence calculate default probability correlations: (1) looking at historical defaults, (2) using information from equity prices (via a so-called Merton model), and (3) using information from credit spreads. Here we consider each in turn and attempt to calculate comparable correlation numbers. We will generally use the example of Ford and General Motors.
13
Our message in this section is that estimating default correlation from equity returns, the current market standard, is simple and convenient, but using credit spreads as an alternative source of data may have advantages.
Default Correlation From Historical Defaults

First of all, is there actually any evidence of correlation between company defaults in real life? As each firm is unique and defaults are caused by many different events, it could be argued that there is little relationship between individual defaults. This can be checked, approximately at least, by looking at the history of actual defaults. In theory, this is the most obvious source of the relationship between defaults for modelling purposes, though, as we will see, it is not generally used in practice. Chart 10 below shows the percentage of defaults per year since 1970 from Moodys Investors Service, across all countries. This suggests that there is some correlation between defaults as they tend to cluster. If there was no correlation, they would be more evenly distributed over the historical time period. The chart also suggests that there is some correlation with the business cycle.
Chart 10: Global Defaults, 1970 year-end 2003
default rate percent per year (LHS), annual percent change in US Industrial Production (inverted RHS) 5% 4% 3% 2% 1% 0% 1970 1974 1978 1982 1986 1990 1994 1998 2002
Source: Moodys Investors Service, JPMorgan.
-10% -5% 0% 5% 10% 15%
As far as quantifying the actual historical default correlation, the only information we can obtain from this data is an approximate measure of probability of default and default correlation over all companies over this total period. The problem with this approach is that defaults are very rare events, so in order to obtain a statistically significant sample of data points, a very long history is required. This then poses the question of whether the relationships produced from the data would be relevant to the corporate world today. For example, if we consider a single sector of the market, we see a similar picture. Chart 11 shows defaults in the healthcare industry over the same period. Again, we see a strong default clustering effect, though this was caused more by a change in the reimbursement levels that US government programs provided to healthcare providers (specifically, nursing homes) than just the general business cycle.
14
Chart 11: Global Healthcare Defaults, 1970 year-end 2003

defaults per year
12 10 8 6 4 2 0
Gov 't legislation
1970 1974 1978 1982 1986 1990 1994 1998 2002

Source: Moodys Investors Service, JPMorgan.
In conclusion, if we collect enough data to obtain a statistically significant number of default observations, the resulting correlation will be an average over several business cycles (and possibly several business environments as well). So, while historical default correlation is in theory the input that should be used for pricing tranche products and CDOs, it is rarely used because of the lack of any specificsfor instance we cannot obtain any information about the default correlation between Ford and GM as they have not defaulted! The closest thing would be to investigate defaults in the same sector, or for companies with similar ratings. This is typically different to modelling in other asset classesfor example the volatility input required for pricing interest rate options can be estimated from the historical volatility of the forward rate.
Default Correlation from Equity Returns

It is not news to suggest that credit and equity are related. Robert Merton first published his framework modelling the relationship in 19741. The Merton model (or family of models) is now commonplace in the market and is based on the observation that an equity holder is long a call option on the assets of a firm, and a bond holder is short a put on the assets. Put-call parity therefore links credit and equity values. If you consider the assets of a company to be a random variable, then the company will default at some threshold (basically where the assets arent enough to repay the debt) see Chart 12. The main parameter that is difficult to determine is the relative amount of credit and equity that make up the assetsknown as the leverage. Derivations of the Merton model can be used to relate equity volatility and correlation through asset volatility and correlation to default probability and correlation. The CreditMetrics model is based on the Merton framework. Please refer to the CreditMetrics technical document2 for a full explanation of the process, but the actual correlation input required by CreditMetrics is the correlation of the companies asset returns. As a time series of asset values are not readily available, equity correlations
1 2
On the Pricing of Corporate Debt: The Risk Structure of Interest Rates, 1974 CreditMetricsTM Technical Document, 1997.
15
are used as a proxy, and RiskMetrics argues that these are the same, given certain assumptions3.
Chart 12: Asset Return Distribution
Probability (y axis), asset return (x axis)
Firm defaults
Source: RiskMetrics.
The advantage of using equity data is that there is a good history of prices with a distribution that is close to normal, though a model (with associated assumptions) is required to convert the equity returns correlation to a default correlation. Returning to our two example companies, Chart 13 below shows the equity prices for Ford and GM over the last four years.
Chart 13: Ford and GM Equity Price History
100 90 80 70 60 50 40 30 20 10 0 Mar-00 Sep-00 Mar-01 Sep-01 Mar-02 Sep-02 Mar-03 Sep-03 Ford
Source: JPMorgan.
GM
Chart 14 then shows the rolling one year correlation between the equity returns, with an average of around 60% over this period, though exhibiting an upward trend.
Equity and Asset Correlations in CreditMetrics, 2002.
16
Chart 14: Rolling 1 yr Equity Return Correlation between Ford and GM

90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Oct-00 Mar-01 Aug-01 Jan-02 Jun-02 Nov-02 Apr-03 Sep-03 Feb-04
Source: JPMorgan.
As stated at the beginning of this section, equity returns are generally used as a proxy for asset returns. Here we will very briefly check this assumption for three pairs of companies. This is only intended to be an approximate analysis due to the lack of data and differences in accounting treatments of debt levels across different companies. However, even with these caveats, our examples show that asset return correlation can be significantly different to equity returns correlation. We have chosen pairs of similar U.S. companies as our example. As debt levels can only be obtained from company reports, we are restricted to a quarterly time series at best. Using the relationship that equity = assets liabilities, we added company debt reported every 3 months to the equity value observed in the market at the same time to obtain a measure of enterprise value. This produces an estimate of how the assets of the company varied over this period, and the correlation of these asset returns can then be compared to the quarterly equity return correlation over the same period (Table 1).
Table 1: Comparisons between Asset and Equity return Correlations (quarterly, 5 yrs)
Company pair and sector Safeway vs Albertsons (food retail) Caterpillar vs Deere & Co (construction) Ford vs General Motors (autos) Dow Chemical vs Du Pont (chemicals)
Source: JPMorgan.
Equity Return Correlation
Asset Return Correlation
35 % 54 % 41 % 77 %
-4 % 34 % 7% 36 %
This table shows that choosing which return series to correlate can make a difference, with asset return correlations tending to be lower than equity return correlations. For example in the auto sector the asset return correlation between Ford and GM is as much as 30% lowerwhile GM and Fords equity values behaved similarly over the last five years, GMs debt increased while Fords remained relatively constant.
17
Hence, its not surprising that we get such a large difference between correlations calculated from equity and asset returns for Ford and GM. We caveat our analysis with the obvious point that this is a very small sample of companies. This serves as a reminder that there are limitations to Merton models, both practical (credit is a less liquid market than that of equities) and theoretical (it assumes that default and insolvency are the same thing), but the market is happy to accept that there is a relationship between credit and equity. The main limitation is that the link between equity, debt, and assets (leverage) is not constant, difficult to measure and judge, highly correlated (in the intuitive not the linear sense) to the business cycle, and can be changed up or down by management decisions (the so-called management option). Specifically, for regulated financial businesses and companies like Ford and GM that have large captive finance subsidiaries (FMCC and GMAC), Merton models are at their least reliable. This is broadly because the balance between debt and equity is so skewed. The same phenomenon can be observed with the spread for these companies calculated from CreditGrades, RiskMetrics equity to credit spread model.
Default Correlation from Credit Spreads

The price of a corporate bond comprises two parts: an interest rate component and a credit component. If the interest rate component is stripped out, then the remaining fluctuations in the bonds price should contain primarily information about the credit worthiness of the issuing company. From this we can estimate the probability of default, although there are some problems with this method. First, the method for extracting the probability of default from the bonds price requires the loss given default (recovery rate) which is not generally well known in advance of a default. Second, the bond price could change due to technical reasons unrelated to the credit worthiness of the companyfor instance the price will be influenced by liquidity factors for that particular issue. There are also logistical problems, such as individual bonds maturing at different times in the issuers life (so a history of probabilities of defaults may have to be made up of several bonds issued by the same company), and additional complications from different rankings of the bonds. Most of these logistical problems can, in theory, be resolved by using credit default swaps (CDS) rather than corporate bonds. The par CDS spread quoted in the market can be converted into a probability of default (again, assuming a certain recovery rate), but the main drawbackuntil recentlyof using CDS spread data was that there was insufficient data to create a reliable correlation estimate. However, now there is a reasonable history of around three years of data for the most liquid names. Admittedly, this only covers part of a business cyclewhich as we saw earlier in Chart 10 can affect the correlation between defaults quite strongly. Also, even though the probability of default obtained from a CDS may be cleaner than that obtained from a corporate bond, there are still risk and liquidity premia incorporated in CDS spreads. However, we think it would be instructive to compare the default correlation estimates from the available CDS data and compare it to the correlation derived from equity return data for our two example names.
18
Chart 15: Annual probability of default for Ford and GM, calculated form 5yr mid CDS spreads
16% 14% 12% 10% 8% 6% 4% 2% 0% Mar-00 Sep-00 Mar-01 Sep-01 Mar-02 Sep-02 Mar-03 Sep-03 Mar-04 Ford
Source: JPMorgan.
GM
Chart 15 above shows the probability of default for Ford and GM, obtained from the CDS market. This is the annualised probability of default for each company, calculated from the 5 year CDS spreads (the most liquid CDS maturity) and hence can be considered the markets average annual probability estimate for each company defaulting over the next 5 years. In order to compare this probability of default data with the equity and asset returns time series produced earlier, we need to convert the data shown earlier in Chart 15 to an asset return correlation. The way to do this is to use the CreditMetrics methodology in reverse to obtain the default threshold from the known probability of default. The output of this reversing process is in Chart 16, which depicts the asset correlation obtained from Ford and GM CDS spreads. Chart 16 shows a very different pattern to the actual equity return correlations in Chart 14. Although the average over this period is similar, around 60%, the correlation has declined over time, unlike in the equity return correlation in Chart 14, where its increasing.
Chart 16: Rolling 1 yr Asset Return Correlation, calculated from Ford and GM 5-yr mid CDS spreads
90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Nov-00 Apr-01 Sep-01 Feb-02 Jul-02 Dec-02 May-03 Oct-03 Mar-04
Source: JPMorgan. 19
Conclusion
So, which of these data series (historical defaults, equity returns, asset returns, and credit spreads) produces the better estimate of default probability? Actual historical default information is too coarse to be used in CDO tranche pricing. This leaves asset/equity returns, or probabilities of defaults obtained from the CDS/bond markets. Currently, the market consensus is that equity inputs into the CreditMetrics methodology is the most practical approach for models typically used, although we believe it is necessary to be mindful of the potential change in leverage of the company over time if possible. Increasing liquidity in the CDS market may lend itself to the creation of alternative models which value tranches from default correlation calculated from credit spreads. This is significant as CDS trades are generally used to delta hedge tranched credit trades, and in most other derivatives markets, price is determined by hedging cost. Theoretically, as tranched credit products (such as CDOs backed by CDS) are priced using equity correlations and hedged using CDS, there could be a mismatch in the delta hedging of these products over their life, caused by the difference between the default correlations derived from equity and CDS data. A final point to remember is that while the history of actual company defaults does give us a real estimate of the correlation between defaults, the other methods we have investigated in this section produce a correlation between some measure of probability of default or credit worthiness. This measure of credit worthiness may come directly from the CDS market or indirectly from asset / equity correlations (via a model), but they are not strictly a correlation between defaults.
20
Section 3: Correlation Trading Basics

Introduction
So far, our discussion of correlation has been fairly abstract. The reason default correlation matters so much in credit is that it directly impacts the pricing of complex derivativesin credit this means leveraged exposures to credit portfolios, for example CDO tranches. If we model the relationships between default probabilities incorrectly, then we risk pricing such instruments incorrectly. In this section, we will first investigate the impact of correlation assumptions on tranche pricing. We will then introduce JPMorgans framework for calculating correlations embedded in tranche pricesBase Correlationsand finish by considering some of the advantages of this approach over other methodologies. This piece is not intended to be a detailed discussion of tranche pricing, but in essence tranches are generally priced by first modelling the pattern of losses, which for a typical credit portfolio looks something like Chart 17, the probability of certain losses occuring. The average of this distribution is the expected loss. It is then a relatively simple matter to apply the relevant write-down structure and value these losses for a relevant tranche. For example, a generic mezzanine tranche from a synthetic CDO with a linear writedown might have a lower attachment point (K-) of 3% and an upper attachment point (K+) of 6%. This tranche experiences losses in the 3-6% range of the underlying portfolio losses, as there is a linear write down structure between K- and K+ (Chart 17).
Chart 17: Credit Loss Distribution with 3-6% Tranche
Probability (y axis), portfolio loss (x axis)
0 3% 6%
Source: JPMorgan.
50
100
Different assumptions of the relationship between credits and correlation between default probabilities change the shape of the modelled loss distribution. All else being equal, that is the individual default probabilities and recovery rates remain the same, then the loss distribution will change as the probabilities are redistributed around the expected loss point. See Chart 18 for a few different loss profiles based on different correlation inputs.
21
Chart 18: Credit Loss Distributions under 0, 25, and 50% Correlations
Probability (y axis), portfolio loss (x axis)
ex pected loss
20 0%
40
60 25%
80 50%
100
Source: JPMorgan.
The price of a tranche is therefore directly related to the probability of portfolio losses occurring relative to the attachment points; this is directly impacted by the correlation assumptions. With the recent developments in the standardisation of credit derivatives indices and tranched documentation, the synthetic CDO market has the potential to move away from a market where prices are determined by models to a market where there is sufficient supply and demand to enable efficient price discoveryand hence prices will be determined by industry wide consensus rather than one model. Rather than use either equity or credit spread based models to look at forecasts from historical analyses, what we will increasingly want to do is to look at implied correlations from quoted prices to determine relative value.
Correlation Trading
Correlation trading has been spurred by the development of standardised CDO structures such as DJ Tranched TRAC-X. It is essentially buying or selling a tranche of a synthetic corporate CDO, with the view that the realised correlation over the holding period will be different to that embedded in the instrument. Pure correlation trades are likely to isolate as many other market factors as possible by delta hedging. The obvious (but not straightforward) question therefore is what correlation is embedded in the price?. In this section, we will consider some alternative answers to this question, as well as introduce JPMorgans approach, Base Correlations, and our reasons why we think this meets the needs of the investment community. In the same way that the volatility embedded in an option is known as the implied volatility, a natural extension is to talk about implied correlation. JPMorgans approach to calculating implied correlations from standardised tranche spreads has two main differences from some other approaches. Firstly, we standardise on a simple model (and implementation of that model, which we describe in the next section) that has few inputs and can be freely distributed. Secondly, we calculate correlations on bootstrapped equity tranches (that is, tranches with one attachment point at 0%).
22
Homogeneous Large Pool Gaussian Copula Model

The model we use to imply correlations in tranches is known as the homogeneous large pool gaussian copula (the large pool model, or HLPGC), which is a simplified version of the Gaussian copula widely used in the market. This model is not new, its a simple methodology thats almost identical to the original CreditMetrics model (Gupton et al, 1997). It is a simplified form of earlier one-factor models (Vasicek, 1987) and is described in numerous places, for example Schonbucher (Credit Derivatives Pricing Models 2003). Here we will discuss this model in the context of what is a tradable and transparent credit structure, DJ Tranched TRAC-X, but in theory this analysis could be applied in an approximate manner to similar static CDOs (with purely linear writedowns and portfolios that can be marked-to-market). The model is based on three major assumptions. First, default of a reference entity is triggered when its asset value falls below a barrier (see Section 2 for a more complete description of this Merton/CreditMetrics-based asset correlation framework). Second, asset value of the portfolio is driven by a common, standard normally distributed factor M, which is often referred to as the Market, and can be taken to imply the state of the overall business cycle. Finally, the portfolio consists of a very large number of credits of uniform size, which effectively cancels the effect of a single names performance on tranche loss and is why the portfolio can be considered homogenous. We believe that the fundamental benefits of the large pool model are transparency and replicabilitywe can provide our specific implementation of the model. The model also has the advantage that it requires few inputs only the average level of market spreads and average recovery rate (which we define as 40%), rather than individual spreads for all of the credits in the portfolio, which would be impossible for a user to reproduce at any instant. The downside of course, is that the model does not consider single name blow ups correctly. This manifests itself in two main ways: one, the model cannot differentiate between a single name widening by 10,000bp and 100 names widening by 100bp, and two, there is a discontinuity as credit spreads widen towards default. The model is unlikely to produce spreads consistent with market observations in these scenarios. We believe because of the infrequent occurrence of big single name moves that this limitation is more than made up for by its simplicity and transparency. The concept of using Base Correlations does not require this model, but note that by definition, any model that can differentiate between single name effects must also require entry of all single name spreads to imply correlation, which is not practical in OTC markets. The large pool model is provided as an aid both to understand pricing concepts, as well as to help in correlation quoting. No representation is made as to its accuracy or efficiency. It is not the model JPMorgan uses internally for its valuation and risk management.
23
Why have we chosen to use this model? Using the Large Pool model has advantages over using an apparently more sophisticated Gaussian Copula because of a need for simplicity and transparency, in our view. Even though these standardised copula models are in principle open and transparent, each requires multiple inputs, which are difficult to agree on instantaneously such as the spread and recovery rate of each of the 100 names in the DJ TRAC-X portfolio. It is therefore very difficult for an investor to take these quoted implied correlations and calculate the appropriate tranche spreads.
Implied Correlation Methodologies

Compound Correlation Some market participants are already publishing implied correlations, assuming that an implied volatility approach from equities will work directly. What has generally been done is to agree on a model (usually a Gaussian copula), which takes all single name spreads and a single asset correlation as input and produces the tranche spread as an output. A simple iteration process can then take the traded spread and produce an implied correlation. At least in theorybelow we discuss some concerns we have with a compound implied correlation approach.
We define compound implied correlation as the single correlation that gives the value of a complete tranche of a synthetic corporate CDO (a CDO with a purely linear writedown and a portfolio of CDS).
Base Correlation Base Correlations are defined as the correlation inputs required for a series of equity tranches that give the tranche values consistent with quoted spreads, using the standardised large pool model and the inputs in Table 2. Equivalently, we can quote these same Base Correlations for the upper attachment point of these tranches.
Table 2: Standardised Assumptions for Calculating Base Correlation
Recovery Rate Market Spread Discount Rate
Source: JPMorgan
40 % Mid level of equivalent quoted DJ TRAC-X Unfunded Swap 0%
In order to obtain the correlations for each first loss tranche (known as Base Correlations), we need to create a bootstrapping process. For example, consider tranche attachment points K1, K2, K3 , (3%, 6%, 9% . for DJ Tranched TRAC-X) the process for calculating the expected loss (EL) for each first loss tranche is as follows: EL[0,K1] (already traded in the market) EL[0,K2] = EL[0,K1] + EL[K1,K2] EL[0,K3] = EL[0,K2] + EL[K2,K3] This is known as a bootstrapping process and is also used in creating a zero curve from market swap rates. The principle of bootstrapping is that the process must start with a known starting point, which is then used to obtain the next step in the sequence, and so on in order. For example, the implied correlation for the first loss tranche 0-3% traded in the market can be obtained directly from its spread. Then the
24
expected loss of the 0-6% (non-traded) tranche is obtained by adding the expected loss of the 0-3% first loss tranche to the 3-6% mezzanine tranche. Once we have expected losses for the sequence of first loss tranches, we can iterate to find the single Base Correlation for each tranche.
Why is Base Correlation an Improvement?

We fully anticipate that there will still be some discussion on the relative merits of different methods of implying correlation from tranche prices its not clear that one method is right. However, we present the key reasons we believe that Base Correlations offer many advantages over compound correlations:
1. Base Correlation is well defined Tranches of credit portfolios can be thought of as options on portfolio losses; specifically, a call spread with the different strikes at each attachment point. Chart 19 shows the payoff profile for the 3-6% tranche through an option diagram, the familiar hockey stick diagrams for call and put options. A long credit position through this tranche is similar to being short one call option struck at 3% and long another struck at 6%, with the underlying asset represented by the portfolios loss. In the same way that it would not make sense to quote an average implied volatility for an equity option spread, we believe it is inappropriate to calculate a single implied correlation for a tranche.
Chart 19: 3-6% Mezzanine Tranche Payout Profile
Protection payout (y axis), portfolio loss (x axis)
3%
Source: JPMorgan
6%
Quoting a single implied correlation is especially problematic for mezzanine tranches. Chart 20 plots spread for the 3-6% tranche of DJ TRAC-X Europe for various correlations. In each case it can be seen that for the traded spread of 227bp, there are two possible correlations, one in the 10-15% range, and another around 80%. Moreover, if the spread on this tranche were over 335bp, there would be no correlation that gives this spread, and hence no solution. We have repeated this analysis for a Gaussian copula implementation that uses single name spreads to demonstrate that the problem of multiple solutions is not a function of the use of the large pool model.
25
Chart 20: Relationship between Compound Correlation and Spread for Mezzanine Tranche
DJ TRAC-X Europe 3-6%Tranche spread (y axis), Implied Correlation (x axis), as of February 16 2004 400 350 300 250 200 150 100 50 0 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% Large Pool Model
Source: JPMorgan.
no solution above 335 bp
two solutions for 227 bp
Single name Specific Model
A Base Correlation however will only have one solution. This is because an equity tranche spread is a monotonic4 function of correlation5. The behaviour of mezzanine tranches can also be explained by considering that they can be split up into two first loss tranches, one long protection and the other short protection. This creates one position of long correlation sensitivity (at one of the constituent first loss tranche attachment points) and another position of short correlation sensitivity (at the other first loss tranche attachment point). The degree to which these two correlation sensitivities offset each other will depend on the current market conditions at the time. It is these two opposing correlation sensitivities that create the problems outlined above for the mezzanine tranche.
4 5
Monotonic means always increasing or always decreasing. Another way to consider this relationship is with an option analogy: A single put or call option has a payoff similar to a first loss tranche. As the volatility used to price the option increases, the price of the option will increase monotonically. The situation is the same with the first loss trancheas the correlation increases the expected loss of the tranche will decrease monotonically. This guarantees that there is a unique solution when searching for either the implied volatility or correlation.
26
2. Base Correlation defines a more meaningful skew We show the Base Correlation skew in Chart 21. Whilst at first glance Base Correlation values may appear very high compared to compound correlations (Chart 22), this skew actually reflects the difference between real world pricing and the models assumptions.
Chart 21: Base Correlation Skew
Base Correlation (y axis), upper attachment (x axis)
Chart 22: Compound Correlation Skew

Compound correlation (y axis), attachments (x axis)
70% 60% 50% 40% 30% 20% 10% 0% 3

Source: JPMorgan
70% 60% 50% 40% 30% 20% 10% 0%
12
22
0-3
3-6
6-9
9-12
12-22
The reason for this is that CreditMetrics type models place very low probabilities on losses for very senior tranchesfor example, the 22-100% tranche. A CreditMetrics model-driven value for this part of the capital structure is, effectively, zero basis points. In reality, nobody would be prepared to take this risk for nothing, and the market charges a few basis points. Because the 22-100% tranche has even a small spread of a few basis points, this skews the implied correlation substantially for the remainder of the capital structure. To see this clearly, Chart 23 shows fair spread for various correlations on the 0-22% tranche, scaled to reflect the portfolio size. Note that all correlations less than about 50% produce a fair spread exactly the same as the entire portfolio, which was trading at 42bp, suggesting there is no extra value in the 22-100% part. In reality, the market prices the 0-22% tranche at 38bps (bootstrapped from standardised tranches), which implies a correlation up to the 60% range (Chart 23).
27
Chart 23: Hypothetical 0-22% Tranche

Tranche spread (y axis), Base Correlation (x axis) 45 40 35 30 25 20 15 10 5 0 0% 10% 20% 30% 40% 50% 60% 70% 80% 90%
Source: JPMorgan
0-22% tranche spread of 38bp
Another way to look at this apparently steep skew, is to use an equity option analogy. Let us use the S&P 500 index, and assume the index level is currently 1100. At the money volatilities are approximately 17.5% for one year maturities. We consider a deeply out of the money put, with a strike of say 350 index points. With a volatility of 17.5%, an option model gives an option price of nearly zero for this put. Of course nobody would sell such risk for zero, but even if we use a tiny price such as one index point, the implied volatility shoots up to over 50%. We regard the Base Correlation skew as a fairer representation of relative correlation levels because compound correlation is really a combination of two points.
3. Base Correlations can be used to value other tranches An important application of the Base Correlation framework is that it allows market participants to quickly and consistently determine value for tranches other than those that trade actively. As is clear from the explanation of the difference between base and compound correlations above, any sort of interpolation process from the compound skew shown in Chart 22 is compromised by the fact that these compound correlations are a function of two Base Correlations. An analogy would be in the interest rate swap markets, where a set of zero rates is bootstrapped from market swap rates in order to value off-market swaps. Using a Base Correlation framework, we can use the market standard liquid tranches to calibrate the model for Base Correlation inputs, and then to interpolate from these Base Correlations to value an off-the-run tranche with the same collateral pool.
As an example, what if an investor wanted to value a tranche based on non-standard attachment points of DJ Tranched TRAC-X Europe? Previously, the only way would be to ask for theoretical prices, but now there is a consistent and well-defined method to value such tranches. Chart 24 below uses this methodology to back-out implied correlations for an example off-the-run 4-7% tranche; the markers on the curve are the standardised attachment points (i.e. 3, 6, 9, 12, and 22%), and the intersections of the lines are the interpolated 4 and 7% attachment points of our new tranche.
28
Chart 24: Off-the-run 4-7% Tranched TRAC-X Europe, as of February 19 2004

Base Correlation (y axis), upper attachment (x axis, per cent)
70% 60% 50% 40% 30% 20% 10% 0% 0
Source: JPMorgan
10
15
20
25
Base Correlations can therefore be used to produce consistent implied correlations (and spreads) for non standardised tranches of a credit portfolio, which greatly increases the risk/return horizon for investors seeking a reality check on quoted prices of tranches with non-standard attachment points. For instance, this enables us to value super senior tranches in a consistent way. Whilst different writedown structures, different collateral pools, and the lack of readily observable market prices for certain types of collateral mean that this approach cannot be directly applied to other (more complex) types of CDOs, it is nevertheless an interesting starting point for discussions around the market correlation view.
Conclusion
Base Correlations calculated with a large pool model provide market participants with a measure of implied correlation in tranches which has the advantage of simplicity, replicability, a unique solution, realistic skews, and the ability to price offthe-run tranches. We believe that these advantages over a compound implied correlation will make the tranched credit market more transparent and enable more investors to start treating correlation as an asset class.
29
Appendix 1: Implied Volatility in Equities versus Base Correlation in Credit

For readers familiar with equity and interest rate options and the role implied volatility plays in those markets, the following may help explain how we view the future development of Base Correlations in the credit market. Those not familiar can safely skip this appendix. The fact that equity and interest rate options markets are over 20 years old and are now quite mature is good news for those of us analysing the effect of default correlation on tranches of CDOs. Or at least it will be if we can demonstrate that the concepts are similar enough so that we can either use one to explain the other, or to predict the future growth of implied correlation. At first glance, the concept of implied volatility looks similar to how we are starting to think of correlation in tranches. There is a model for pricing, and a non-observable forecast unknown. The concepts are indeed very similar, but in understanding how correlation is likely to develop, we must understand the differences, both fundamentally and in terms of the stage of the development.
Equity Implied Volatility

In equities, there is a well understood and accepted model for pricing optionsBlack Scholes6. The inputs into this model are current stock price, the level of interest rates, dividend yield, strike price, and stock price volatility. As the first four are readily observable, volatility can be thought of as the only differentiator between different prices from different dealers. Options in these markets are often exchange traded and very liquid, hence there is typically a market observable price, and therefore an implied volatility. This implied volatility can be thought of as the market consensus view of the forecast volatility of the stock. In theory, if you think the market has got it wrong, then you can trade the option and delta7 hedge, realizing the difference between the realized and implied volatility, and generating profits (or losses if youre wrong). In practice, there are complications. First, Black Scholes doesnt reflect the real world perfectly, it makes an assumption about the distribution of the returns of the underlying stock or indexthat they are lognormal. This is not quite the case, so dealers tend to quote slightly different implied volatilities for each strike. Chart 4 in Section 1 illustrated the distribution of stock returns from General Motors, indicating the difference between equity returns and a perfectly normal distribution. Second, supply and demand will skew option prices. There are natural demands from end users for certain kinds of optionsi.e. greater demand for out-of-the-money put options by investors seeking downside protection or hedges. Third, while we tend to assume that the other inputs are readily observable, this isnt quite true. It is possible for different dealers to have the same view on future volatility, to use the same
6 7
"The Pricing of Options and Corporate Liabilities", Journal of Political Economy 81, 1973. Delta is the sensitivity of the option price to a small change in the underlying equity price, so delta hedging is buying or selling a ratio of the equity to isolate a view on volatility.
30
model, but to have different views on the future dividend stream of a company. In this case each would quote different option prices and different implied volatilities. All of the above contributes to a skew or smile in implied volatility for many traded financial instruments. This is shown below by Chart 25, which plots the implied volatility for each strike for call options on the Dow Jones Stoxx 50.
Chart 25: Implied Volatility Skew for Dow Jones Stoxx 50, as of February 4, 2004
Implied Volatility (y axis), Option Contract Strike (x axis)
40 35 30 25 20 15 10 5 0 2450 2500 2550 2600 2650 2700 2750 2800 2850
Source: Bloomberg.
That said, in equities, dealers and investor have got comfortable with the fact that implied volatility, while not perfect, mostly incorporates market consensus of future volatility (with some adjustment).
Credit Implied Correlation

In terms of the definition we will choose to use, one can consider implied correlation in the same terms as implied volatility. That is, implied volatility is the volatility that, given a known model and known inputs, produces the current price, and we define implied correlation as the single pairwise correlation that, given a known model and known inputs, produces the current price. In each case, we need to be careful about not drawing the immediate basic conclusionsthat the implied measure is simply the markets consensus view of future volatility or correlation. There are model inaccuracies, unknown inputs and supply and demand affecting pricing dynamics. We believe that these differences are likely to be more important in the credit market. In this research note, we suggest using a large pool model. Given a known recovery rate assumption (40%) and credit spreads for a given portfolio, the only unknown in this model is the average pairwise default correlation. This model is simple, transparent, well understood, and is not realistically the model that sophisticated houses use to determine their prices (like Black Scholes). We define implied correlation as the single pairwise correlation between asset returns that when included as in input into this model, produces the market spread being quoted. We combine tranches and calculate implied, or Base Correlations for
31
equity tranches only (that is tranches with an attachment point at 0%). Again, in theory if you take other unknowns as read, you can buy (sell) correlation through these tranches, and dynamically manage the single name risk. Over a large set of observations, you will then profit if default correlation is higher (lower) than implied in the original price. Again, there are several problems with this scenario. First, most market makers are using more sophisticated models (such as the more advanced copulas provided elsewhere in the note), so the errors between this large pool model and the traders model can be significant. Second, the underlying market of single name (or even DJ TRAC-X) CDS is a relatively illiquid OTC market, so dynamic delta hedging is imprecise and costly. Third, the balance of supply and demand can skew prices irrespective of other factors, although its fair to say that the depth and liquidity of the synthetic credit market pales into insignificance when compared to equity or interest rate markets. Finally, the assumption that other facts are insignificant or observably agreed is less true in this market: recovery rates are the largest such factor, but there are many more. As it becomes easier and more transparent to trade recovery rates, then they too will become an observable factor.
Conclusion
So while we can use the term Base Correlation with a similar precise meaning to implied volatilitythe input into a model that leads to the same model prices as the quoted levelwe must be a little careful in interpreting that number. For equities the approximations are small and we can usually assume that implied volatility is the consensus view of future volatility, but in credit, there are still many other unknown areas included in an implied correlation measure. Nevertheless we will still use this measurewe believe it will be an important indicator for the credit market. We are aware of its limitations, but for those familiar with equity or interest rate option markets, this offers some insight into what Base Correlation means, as well as the possible future development of this measure.
32
Appendix 2: Correlation and CDO Rating Methodologies

For readers interested in how Moodys, S&P, and Fitch view and measure correlation in the credit market, this appendix summarizes their approaches to rating CDOs backed by corporate debt. Those seeking a more complete analysis should refer to the rating agency documents themselves.
Moodys Investors Service

In their CDO ratings framework, Moodys assumes that firms within similar industries and geographical regions are more likely to default or survive together, and quantifies this relationship with a Diversity Score for a given portfolio8. The Diversity Score calculation involves grouping a CDO portfolio into a matrix of different industries and adjusting for the par amount of the names within each industry group. The final result is that an actual, correlated portfolio has been reduced to an idealized, uncorrelated portfolio. Chart 26 demonstrates this relationship between industry concentration and portfolio diversity. For instance, 20 on the x-axis means that there are 5 equally-sized industries each comprising 20% of the portfolio, implying that there are between 25 and 30 uncorrelated components within this portfolio. We note that Moodys accounts for systemic risk (i.e. nonidiosyncratic risk arising from general macroeconomic conditions) not through the Diversity Score, but by stressing CDO portfolios with default and loss severity rates that are higher than past experience.
Chart 26: Moodys Approach to Correlation
Moodys Diversity Score (y axis), industry concentration (x axis) 60 50 40 30 20 10 0 20 15 10 5
Source: Moodys Investors Service Note: we assume that all names are equally-weighted and that the portfolio solely consists of corporate debt (i.e. no Structured Finance or nontraditional asset classes).
Rating Cash Flow Transactions Backed by Corporate Debt 1995 Update, 1995, Moodys Investors Service.
33
Fitch Ratings
In Fitchs factor model9 for CDOs equity returns are used to proxy asset correlation and the model can also account for idiosyncratic risk. Fitchs correlation matrix starts from the Dow Jones Global Universe, which is then broken down into industry and region codes. Based on its findings, Fitch notes that the correlation between industries ranged from 13 to 22%, while the correlation between regions was less on average, around 6 to 20%.
Standard and Poors

S&Ps approach uses a pairwise correlation matrix which is constructed from industry, country, and regional correlations, and is mainly used to rate CDOs10. Although S&P has indicated that they may revise their guidelines, for corporates in the same country they estimate intra-industry correlation at 30%, but assume no correlation (i.e. perfect independence) between industries (inter-industry).
10
Global Rating Criteria for Collateralised Debt Obligations, 2003, Fitch Ratings. CDO Evaluator Applies Correlation and Monte Carlo Simulation to Determine Portfolio Quality, March 2003, S&P.
34
35
Lee McGinty (44-20) 7325-5482 lee.mcginty@jpmorgan.com
"Dow Jones" and TRAC-XSM" are service marks of Dow Jones & Company, Inc. or its licensors and have been licensed for use for certain purposes by JPMorgan. Tranched DJ TRAC-X is not sponsored, endorsed, sold or promoted by Dow Jones, and Dow Jones makes no representation regarding the advisability of investing in such product(s). Analyst Certification The analyst(s) denoted by an * hereby certifies that: (1) all of the views expressed in this research accurately reflect his or her personal views about any and all of the subject securities or issuers; and (2) no part of any of the analysts compensation was, is, or will be directly or indirectly related to the specific recommendations or views expressed by the analyst(s) in this research. Copyright 2004 J.P. Morgan Chase & Co. All rights reserved. JPMorgan is the marketing name for J.P. Morgan Chase & Co., and its subsidiaries and affiliates worldwide. J.P. Morgan Securities Inc. is a member of NYSE and SIPC. JPMorgan Chase Bank is a member of FDIC. J.P. Morgan Futures Inc., is a member of the NFA. J.P. Morgan Securities Ltd. (JPMSL), J.P. Morgan Europe Limited and J.P. Morgan plc are authorized by the FSA and JPMSL is a member of the LSE. J.P. Morgan Equities Limited is a member of the Johannesburg Securities Exchange and is regulated by the FSB. J.P. Morgan Securities (Asia Pacific) Limited (CE number AAJ321) is regulated by the Hong Kong Monetary Authority. J.P. Morgan Securities Singapore Private Limited is a member of Singapore Exchange Securities Trading Limited and is regulated by the Monetary Authority of Singapore (MAS). J.P. Morgan Securities Asia Private Limited is regulated by the MAS and the Financial Services Agency in Japan. J.P.Morgan Australia Limited (ABN 52 002 888 011/AFS Licence No: 238188) (JPMSAL) is a licensed securities dealer. Additional information is available upon request. Information herein is believed to be reliable but JPMorgan does not warrant its completeness or accuracy. Opinions and estimates constitute our judgment and are subject to change without notice. Past performance is not indicative of future results. The investments and strategies discussed here may not be suitable for all investors; if you have any doubts you should consult your investment advisor. The investments discussed may fluctuate in price or value. Changes in rates of exchange may have an adverse effect on the value of investments. This material is not intended as an offer or solicitation for the purchase or sale of any financial instrument. JPMorgan and/or its affiliates and employees may hold a position, may undertake or have already undertaken an own account transaction or act as market maker in the financial instruments of any issuer discussed herein or any related financial instruments, or act as underwriter, placement agent, advisor or lender to such issuer. Clients should contact analysts at and execute transactions through a JPMorgan entity in their home jurisdiction unless governing law permits otherwise. This report may have been edited or contributed to from time to time by affiliates of J.P. Morgan Securities (Far East) Ltd, Seoul branch. This report should not be distributed to others or replicated in any form without prior consent of JP Morgan. This report has been issued, in the U.K. only to persons of a kind described in Article 19 (5), 38, 47 and 49 of the Financial Services and Markets Act 2000 (Financial Promotion) Order 2001 (all such persons being referred to as relevant persons). This document must not be acted on or relied on by persons who are not relevant persons. Any investment or investment activity to which this document relates is only available to relevant persons and will be engaged in only with relevant persons. In other European Economic Area countries, the report has been issued to persons regarded as professional investors (or equivalent) in their home jurisdiction. Australia: This material is issued and distributed by JPMSAL in Australia to wholesale clients only. JPMSAL does not issue or distribute this material to retail clients. The recipient of this material must not distribute it to any third party or outside Australia without the prior written consent of JPMSAL. For the purposes of this paragraph the terms wholesale client and retail client have the meanings given to them in section 761G of the Corporations Act 2001.

Credit Correlation - A Guide

Uploaded by

Copyright:

Available Formats

Credit Correlation - A Guide

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Credit Correlation - A Guide

Uploaded by

Copyright:

Available Formats

Credit Derivatives Strategy

London 12 March 2004

Credit Correlation: A Guide

First principles in defining correlation

Credit Derivatives Strategy Lee McGinty*

Estimating default correlation from equity prices is simple and convenient

(1-212) 834-4211 eric.beinstein@jpmorgan.com

Derivatives Research Martin Watts

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Section 1: Correlation in the Credit Market

Credit Derivatives Strategy London 12 March 2004

X and Y the mean of each variable. Each variable has n

Credit Derivatives Strategy London 12 March 2004

Chart 2: Scatter charts of different linear correlations

- 1 standard dev iation

+ 1 standard dev iation

Credit Derivatives Strategy London 12 March 2004

Equity Return Distribution

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Chart 7: Joint normal scatter charts, correlation = 95%

Credit Derivatives Strategy London 12 March 2004

Chart 9: Typical joint distribution for credit (illustrative), correlation = 80%

Credit Derivatives Strategy London 12 March 2004

Avoiding Correlation Breakdown

Credit Derivatives Strategy London 12 March 2004

Other methods - Copulas

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Section 2: Estimating Default Correlation

Credit Derivatives Strategy London 12 March 2004

Default Correlation From Historical Defaults

-10% -5% 0% 5% 10% 15%

Credit Derivatives Strategy London 12 March 2004

Chart 11: Global Healthcare Defaults, 1970 year-end 2003

Gov 't legislation

1970 1974 1978 1982 1986 1990 1994 1998 2002

Default Correlation from Equity Returns

Credit Derivatives Strategy London 12 March 2004

Equity and Asset Correlations in CreditMetrics, 2002.

Credit Derivatives Strategy London 12 March 2004

Chart 14: Rolling 1 yr Equity Return Correlation between Ford and GM

Equity Return Correlation

Asset Return Correlation

Credit Derivatives Strategy London 12 March 2004

Default Correlation from Credit Spreads

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Section 3: Correlation Trading Basics

Credit Derivatives Strategy London 12 March 2004

Credit Derivatives Strategy London 12 March 2004

Homogeneous Large Pool Gaussian Copula Model

Credit Derivatives Strategy London 12 March 2004

Implied Correlation Methodologies

40 % Mid level of equivalent quoted DJ TRAC-X Unfunded Swap 0%

Credit Derivatives Strategy London 12 March 2004

Why is Base Correlation an Improvement?

Credit Derivatives Strategy London 12 March 2004