RATINGS ARBITRAGE AND STRUCTURED PRODUCTS

John Hull and Alan White

University of Toronto

First Version: April 2010

This version: June 2010

ABSTRACT

This paper studies the criteria used by rating agencies when they rate structured products. We
assume that some investors assign a value to a product that is monotonic in the credit rating. This
leads to a necessary condition for there to be no arbitrage. The criterion used by S&P and Fitch
does not satisfy the condition while that used by Moody’s does.

1 
 
 RATINGS ARBITRAGE AND STRUCTURED PRODUCTS

1. Introduction

The traditional business of rating agencies is the rating of corporate and sovereign bonds. During
the 2000 to 2007 period another part of their business, the rating of structured products, grew
very fast, so much so that by the end of this period it was contributing close to half of their
revenues. This paper considers whether the rating agencies were using the right criteria for rating
structured products.

Rating agencies provide copious statistics on the default experience of bonds with different
ratings. These statistics provide support for the view that ratings provide a rank ordering of the
probability of default. (For example, a company rated AAA has a default probability of about
0.1% over the following five years; for a company rated AA, this default probability is about
0.2%; for a company rated A, it is about 0.5%; and so on.) As a result, bond’s credit rating is
often perceived by market participants as providing an estimate of the bond’s probability of
default.

Given this interpretation of bond ratings by the market, it was natural for rating agencies to use
probability of default as a criterion when assigning ratings to structured products. This is the
approach adopted by S&P and Fitch. If the probability of a AAA-rated bond experiencing losses
during a five-year period is 0.1%, they require the probability of loss for a five-year structured
product rated AAA experiencing losses to be about the same.

Moody’s criterion for rating structured products is the (undiscounted) expected loss as a percent
of principal. If a Baa-rated bond has an expected loss equal to 1.2% of principal over a five-year
period then a five-year structured product that is rated Baa by Moody’s should have a similar
expected loss percentage.

2. A No-Arbitrage Condition for a Credit Quality Measure

Suppose that q is a measure of the credit quality of an asset that is subject to default risk with the
property that q increases as the credit quality decreases. As mentioned, the measure used by S&P

2 
 
and Fitch for structured products is the probability that the loss over the life of the asset will be
greater than zero while that used by Moody’s is percentage expected loss over the life of the
asset. The asset for which a credit quality measure is calculated can be a single instrument such
as a bond. Alternatively, it can be a portfolio of instruments. The credit quality of a portfolio can
therefore be measured either in terms of the single value of q corresponding to the portfolio or in
terms of the probability distribution of q’s for the constituent assets. Denote the distribution of q
for the assets in a portfolio by F, so that F(q) is the fraction of the portfolio with credit quality
less than or equal to q. We will refer to F as the “credit quality distribution” for the portfolio.

We now define a concept which we refer to as “credit quality dominance.” Suppose that FX and
FY are the credit quality distributions for two portfolios, X and Y. Portfolio Y dominates Portfolio
X if FY(q) ≥ FX(q) for all q with strict inequality for a least one value of q. To use stochastic
dominance terminology, Portfolio Y has credit quality dominance over Portfolio X with respect to
a credit quality measure if there is strong first order stochastic dominance between the credit
quality measures for the two portfolios.

To illustrate the concept of credit quality dominance, suppose that there are three types of assets
with q-values of 1, 2, and 3, respectively. Consider the three portfolios in Table 1. Portfolio B
dominates Portfolio C because it has a bigger percentage of assets for which q=1 and the
percentage of assets for which q is 2 or less is the same for both portfolios. Also, Portfolio A
dominates Portfolio C because they have the same percentage of assets for which q = 1, but
Portfolio A has more assets for which q is 2 or less. There is no dominance between portfolio A
and Portfolio B.

We assume that some investors use the credit quality measure, q, as a guide to valuing a
portfolio. More specifically, we assume that the value some investors assign to a product per
dollar of principal is monotonically decreasing in q. This means that, when Portfolio Y dominates
Portfolio X under the credit quality measure, these investors are prepared to pay more for
Portfolio Y per unit of principal. To prove this formally we note that the value of Portfolio X per
unit of principal is
u

∫
V X = V ( q ) dFX (q )
l

3 
 
where V(q) is the value per dollar of assets assigned to an asset with credit quality q, FX is the
credit quality distribution for X, l is a lower bound for q and u is an upper bound. It follows that
u

∫
V X = V (u ) − V ′(q ) FX (q )dq
l

A similar result holds for the value of portfolio Y, VY so that

u
VY − V X = − V ′(q )[FY (q ) − FX (q )]dq
∫
l

By assumption FY(q) ≥ FX(q) for all q, FY(q) > FX(q) for some q, and V´(q) < 0.1 It follows that VY
> VX.

We define a restructuring of a portfolio as a method by which all the cash flows generated by the
assets in the portfolio are redistributed to create a new portfolio of assets. A credit quality
arbitrage occurs when a portfolio can be restructured into a new portfolio that has a higher value
for at least some market participants. We have just shown that credit quality dominance leads to
credit quality arbitrage. It follows that:

A necessary condition for a credit quality measure to be arbitrage-free is that, for

every Portfolio X and every Portfolio Y that can be restructured from X, there be no
credit quality dominance between X and Y.

3. A Violation of the No-Arbitrage Condition

Probability of loss does not satisfy the no-arbitrage condition. To show this, define Portfolio X as
any portfolio that may be subject to losses due to default. (For example, Portfolio X could be a
single bond or a portfolio of bonds.) Define Portfolio Y as a portfolio consisting of two securities
(or tranches). The first security is responsible for all losses on Portfolio X up to 50% of the
principal of portfolio X; the second security is responsible for the remaining losses on Portfolio
X. Portfolio Y is a portfolio that can be costlessly created from Portfolio X. The probability of
loss for the first security of Portfolio Y is the same as the probability of loss for Portfolio X. In
general, the second security in Portfolio Y has a lower probability of loss than the Portfolio X.2

                                                            
1
 Strictly speaking, we require FY(q) > FX(q) for a range of q that does not have measure zero. In practice, this will
always be the case if FY(q) > FX(q) for a particular q. 
2
This is always true providing there is some chance of losses less than 50% of the principal amount.

4 
 
As a result, the necessary condition for no arbitrage is violated. Part of portfolio Y has the same
q-measure as Portfolio X; the rest of the portfolio has a lower q-measure. If probability of loss is
the quality measure used for X and Y, then Y will always be more valuable than X to some
investors even though X can be costlessly converted into Y.

Now consider a third Portfolio, Z, which consists of three tranches responsible for losses in the
ranges 0 to 25%, 25% to 50%, and 50% to 100%. Using arguments similar to those used in
comparing Y and X we can show that when probability of loss is used as a criterion Portfolio Z
dominates Portfolio Y even though Portfolio Y can be costlessly converted into Portfolio Z.

During the 2000 to 2007 period, tranches were created from subprime mortgages in the way just
described.3 It is easy to see how the probability of loss criterion encourages financial institutions
to create multiple tranches from portfolios of loans. As more tranches are created, the violation
of the no-arbitrage condition becomes greater.

The credit quality measure used by Moody’s, percentage expected loss (EL), does satisfy the no-
arbitrage condition. To show this, suppose that Portfolio X can be restructured into Portfolio Y.
When the quality measure is proportional EL, the percentage expected loss on Portfolio X is
u

∫ qdF
l
X (q )

This is the mean of q for Portfolio X. Similarly the percentage expected loss on portfolio Y is the
mean of q for Portfolio Y. Because percentage expected loss is invariant to restructuring it
follows that the mean value of q for Portfolio X is the same as that for Portfolio Y. As is well
known, when two distributions have the same mean there cannot be strong first order stochastic
dominance between them. This proves that expected loss satisfies the no-arbitrage condition.

It can be shown that a credit quality measure that is any monotonic function of EL also satisfies
the no-arbitrage condition, but some other plausible credit quality measures do not. An example
of a credit quality measure that does not satisfy the no-arbitrage condition is μ + λσ where μ is
the expected loss, σ is the standard deviation of the loss distribution, and λ is a positive constant.
This has the reasonable property that increases in both expected loss and loss dispersion reduce
credit quality. To see that it does not satisfy the no-arbitrage condition, consider a large portfolio
                                                            
3
 See for example Gorton (2008). 

5 
 
of bonds with zero default correlation where each bond has the same μ and the same σ. The
credit quality measure for each bond in the portfolio is μ+λσ. The credit quality distribution for
the portfolio considered as a single asset is close to μ. A credit quality distribution where all the
probability mass is at μ+λσ can therefore be restructured into a credit quality distribution where
all the probability mass is at μ. This trivially violates the no-arbitrage condition.

4. Conclusions

It is likely that one of the incentives for the tranching of mortgage portfolios was the probability
of loss criterion used by S&P and Fitch. Investors, who did not carefully study the criteria used
by rating agencies, were liable to implicitly assume that tranching somehow increased the quality
of a portfolio of assets.

References

Gorton, Gary, 2008, “The Subprime Panic,” European Financial Management, forthcoming.

6 
 
 Table 1: Example illustrating credit quality dominance

Portfolio A dominates Portfolio C. Portfolio B dominates Portfolio C.

There is no credit quality dominance between A and B

Portfolio A Portfolio B Portfolio C

Asset 1 (q=1) 0% 80% 0%
Asset 2 (q=2) 100% 10% 90%
Asset 3 (q=3) 0% 10% 10%

7