R isk T e c h ni q u es

Pros and Cons of

Different CLO Models
Every firm that holds a portfolio of illiquid collateralized loan obligations
(CLOs) must figure out the best way to price these structured products.
Both stochastic and static default models can be used to value CLOs, but
what are the respective strengths and weaknesses of these approaches?
By Anthony Sepci, Dushyanth Krishnamurthy and Christian Eder

hanks to a lack of liquidity in the secondary

markets, many companies have been left holding a portfolio of structured products, including collateralized loan obligations (CLOs).
Decreased liquidity, combined with recent accounting rule changes, has prompted a number
of companies to develop models to help them determine the
fair value of these securities.
These models are highly dependent on inputs, which can
be quite difficult to determine, as many are not readily observable in the market and are prospective in nature. Additionally,
there are a number of modeling techniques available for valuing these securities, and choices must be made based on the
strengths and weaknesses of each model, familiarity with the
credit behavior of the collateral and the overall sophistication
of the valuation group.
This article will demonstrate how CLO securities can be
priced using both static and stochastic modeling approaches,
contrast the results of these two common CLO valuation
methodologies and discuss their respective limitations.
CLO Market Overview
A CLO is a special purpose vehicle (SPV) that issues debt and
equity commonly collateralized by bank loans. Unlike mortgage-backed securities (MBS), a collateral manager actively
manages the collateral pool in a CLO during what is referred

www.garp.com

to as the revolving period, typically ranging from five to seven

years, after which the principal paydown of the liability structure begins.
To value CLO securities, market participants commonly
use a discounted cash flow approach. Given the potential for
diversity in the collateral pool, varying collateral contractual
cash flow terms, complex credit and prepayment forecasting
(as well as structures that revolve), this valuation approach can
be complex. Further, the model also needs to account for dealspecific payment rules, excess interest tests, subordination tests
and credit trigger events dynamically.
The more common method for modeling these transactions
is a static approach where a single best estimate of collateral
cash flows, prepayments and defaults is made. An alternative
method is to model the cash flows for these securities using
a stochastic default model that involves Monte Carlo simulations that generate multiple default scenarios by modeling the
default behavior of the underlying collateral.
CLO Structures
As mentioned earlier, a collateral manager actively manages a
CLOs collateral pool by purchasing and selling loans as permitted in the deals indenture. Collateral loans may consist of
term loans, revolving credit agreements, acquisition or equipment lines, bridge loans, second-lien loans and covenant-lite
loans. Some CLO structures allow the collateral manager to

O C T O B E R 2 0 0 9 RISK PROFESSIONAL

43

R isk T e c h ni q u es

invest in other CLO securities, resulting in the notorious

CLO Squared securities.
CLO securities have varying capital structures, but generally feature debt that is tranched using seniority related to receiving interest and principal payments and the distribution
of losses. A typical CLO structure may have super senior,
senior, mezzanine and subordinate bonds, as shown in Figure
1 (right). Each CLO features payment rules that specify how
collateral interest and principal cash flows are distributed to
the securities.
Additionally, CLOs have a revolving period during which
principal cash flows from the collateral are reinvested rather
than paid out to the securities. During this reinvestment period, the collateral manager purchases new bank loans or
other eligible investments. The CLO structure in Figure 1
contains a pro rata split for the super senior bonds. All collateral principal is first paid to the super senior bonds, followed
by the senior, mezzanine and subordinate bonds, forming the
appropriately termed waterfall. The credit ratings and durations of these securities commonly parallel this waterfall,
respectively running from high to low ratings and from short
to long durations.
Collateral losses of interest or principal are typically first
absorbed by excess spread, then by the lowest rated bond
and then by more senior bonds, in reverse order of payment
seniority. Collateral losses can potentially reach the senior
bonds, depending on the loss levels incurred by the collateral
pool.
In addition, CLO structures typically have interest and
principal credit enhancement coverage tests to further protect senior bonds from losses. When any of these tests fall
below predetermined levels, collateral interest and principal
cash flows are commonly redirected to the most senior securities outstanding in an attempt to bring levels into compliance.
The CLO structure in Figure 1 exhibits how a coverage
test breach can impact security cash flows. In the event of
a coverage test failure, coupon payments to the super senior
and senior security holders are generally protected. However,
mezzanine and subordinate bond holders may not receive
coupon payments during any payment periods where the
coverage tests fail. Less senior security interest cash flows are
diverted to pay off the principal of the most senior security
outstanding, resulting in an early amortization.

44

RISK PROFESSIONAL O C T O B E R 2 0 0 9

R isk T e c h ni q u es

Figure 1: Cash Flow Structure of a Typical CLO

CF from Interest

CF from Principal

Senior Fees
Super Senior A-1

Super Senior A-1

Super Senior A-2

Super Senior A-2

Senior

Senior
Mezzanine

Mezzanine

Subordinate

Subordinate

Sub Fees & Notes

Sub Fees & Notes

Valuation Techniques
Valuation of CLOs usually involves projecting cash flows for
the underlying collateral using assumptions derived from historical collateral performance, market research, managements
judgment and deal-specific inputs such as fees and triggers.
The payment waterfall structure in the deal indenture determines how the projected collateral cash flows are allocated to
the various liabilities making up the capital structure and the
fee/expense requirements.
The resulting projection of cash flows for each security
is then discounted using a risk adjusted discount rate to determine fair value. This valuation process is summarized in
Figure 2 (below).

Figure 2: Steps Involved in CLO Pricing

Step #1

Step #2

Step #3

Step #4

Inputs &
Assumptions

Collateral
Cash Flows

Bond
Cash Flows

Bond
Analytics

CLO triggers, payment rules, assets,

and
liabilities
need to be
accounted for

Collateral cash flow

are generated by
making assumptions on prepayments, defaults,
reinvestments,
recoveries

Collateral cash
flows are paid
to the bonds
according to the
deal structure after
paying CLO fees
and expenses

The net present

value of bond cash
flows it taken. For
stochastic pricing
an average of the
simulations is
taken.

elow, we provide a more thorough explanation for each of

B
these important steps:
Step #1: Inputs & Assumptions. Initially the model
needs to be set up to account for all deal-specific inputs. The
interest rate term structure needs to be estimated to project
interest yield for the floating rate loans and securities, as well

www.garp.com

as the applicable reinvestment income. All fees and expenses

must be accounted for because they reduce available collateral
cash flows. Also, valuation dates, current security balances and
collateral amortization terms need to be updated to reflect amortizing conditions.
Step #2: Collateral Cash Flows. The model should attempt to amortize each loan according to its contractual terms,
but adjust the amortization for projected voluntary prepayments, defaults, default recoveries, interest-rate term structure
and reinvestment income. Prepayments and defaults are usually specified as constant prepayment rate (CPR) curves and
constant default rate (CDR) curves, respectively. These curves
are a vector of percentages that indicate the percentage of
current collateral expected to prepay or default in that period.
The percentages are specified as a yearly amount and must
be converted to a periodic equivalent before being applied to
the collateral balance. The recovery assumption determines
the amount of principal recovered from defaulted loans and
the timing of those recoveries. The model must also make assumptions on the type of collateral that will be purchased during the revolving period.
Step #3: Bond Cash Flows. The model needs to allocate accurately the projected collateral cash flows to the liabilities based on the specified payment rules, trigger events and
fees/expenses, as detailed in the deal indenture. Structures
vary from one CLO to the next, resulting in differences in
tranching and payment rules.
Step #4: Bond Analytics. Once security cash flows are
projected, a number of analytics can be calculated such as duration, weighed-average life, convexity, yield, discount margin
and (ultimately) fair value. To calculate the appropriate fair
value or net present value, an appropriate discount rate/rate
of return must be selected. A discount rate should account for
the risk-free rate, credit risk, model risk and liquidity premium.
For Monte Carlo simulated values, the average of all prices is
taken.
Default Modeling
Default assumptions in a CLO valuation can be modeled as
a constant default rate (CDR) or modeled stochastically using advanced mathematical tools. In the case of static default
(CDR) modeling, a certain percentage of the outstanding collateral balance is assumed to default in each period. The default assumption is specified as a vector of yearly default rates
called a CDR vector, which can be easily converted into a desired periodic rate.

www.garp.com

However, defaulting a portion of a loan is not consistent

with real world default behavior, where loans either default entirely or not at all (a binary event). Stochastic default modeling
tries to incorporate this reality by modeling the approximate
time that a loan defaults.
Simulation is the process by which one can generate possible
outcomes of a random event. For example, if one wants to
simulate throwing a die, one would first need the probabilities
of obtaining the various possible outcomes one through six.
One could then simulate the outcomes using a random number
generator that produces one of the six numbers for each run,
such that the probabilities are maintained (i.e., after a number
of runs, on average one expects to get each number one-sixth
of the time). Similarly, given a probability of default curve that
specifies probabilities of a loan defaulting over time, one could
simulate the time to default using a random number generator, such that the probability distribution is maintained.
In order to simulate the time to default for each loan, the
following is needed: (1) a default curve for each loan; (2) a correlation structure; and (3) a process for generating correlated
time to default for each loan. (Most of the stochastic default
modeling techniques discussed in this article can be found in a
key paper authored by David X. Li.1)
Since the credit crisis began to unfold, there have been a
number of criticisms hurled at this stochastic model. Some
have even blamed the model as the cause of the credit crisis.2 However, though this method has serious limitations, it
has, until recently, been the most widely used model for pricing3, and it is therefore worth examining the construction and
shortcomings of this model.
Probability of Default Curves (Credit Curves)
The probability of a default curve is one of the key inputs into
the default model. It specifies the cumulative probability of
default over time for a loan. It is general industry practice to
vary the probability of a default curve based on the underlying
loan credit quality. The expectation would be that higher quality loans would have a lower probability of default over time
when compared to loans of lesser credit quality.
Default curves can be derived using a number of methods.
For example, Moodys has a table of historic probability of
default curves based on credit ratings. Alternatively, default
curves can be generated from market information such as
bond spreads or credit-default swap (CDS) spreads.
One can also adjust historic default probability curves using
market composites, such as the North American Loan Credit

O C T O B E R 2 0 0 9 RISK PROFESSIONAL

45

46

RISK PROFESSIONAL O C T O B E R 2 0 0 9

Standard Nomal Cum.

Probability Curve
100%
80

Figure 4: Impact of CDR on Bond Losses

60
40
20

16%

0
-5

N-dimensional Standard Nomal

Distribution

-4

-3

-2

-1

-1
Map each random number back to the cumulative probability using a standard Normal distribution. For example, a random number of -1 maps
back to a cumulative probability of 16%.

100%
80

lights that even at a 100 percent CDR, the super senior and
senior bond will retain some value, which is due to cash flows
from recoveries.

Figure 5: Impact of CDR on Bond Prices

100
90

Price % of Par

80
70
60
50
40
30
20
10

0
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Super Seniors

40

101.0
5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100
Super Seniors

25

Senior

Mezzanine

Subordinate

20

16%

15
10
4
0

10

12

3
Map the cumulative probability to a corresponding time to default for the asset. In this example,
a cumulative probability of 16% corresponds to a
time to default of 3 years.

www.garp.com

Figure 5 (right) shows how losses impact the pricing of the

various bonds. Prices were calculated by matching a discount
margin to each bonds accruing spread. As CDRs increase,
triggers are breached leading to cash flows being redirected
from subordinate bonds to senior bonds. This figure also high-

Bond Prices (Percentage)

30%

Subordinate

101.5

Probability of Default Curve

Mezzanine

Figure 6: Impact of Triggers on Super Senior

Bond Prices

60

20

Generate N random numbers for

the N assets from an N-dimensional
standard Normal Random distribution
with correlation structure .

Senior

Figure 6 (below) highlights the impact of triggers within

the CLO structure. For this, we tested the prices of one super senior bond by using a discount margin that was higher,
lower, and equal to the accruing coupon margin. We noted
that depending on whether the bond is valued at a premium or
discount, prices decrease or increase, respectively.
The reason for this is that as coverage triggers breach, super
senior bonds receive principal back faster, as additional cash is
redirected from the subordinate bonds. Different coupon accruing and cash flow discounting rates with faster principal
payments will cause prices to vary.

100.5
100.0
99.5
99.0
98.5

CD
A

Figure 3: Process of Generating Time to Defaults

for Collateral Loans

Hypothetical Example
A hypothetical loan portfolio is used to demonstrate the differences between static and Monte Carlo valuations. The hypothetical portfolio consists of 10 B2-rated loans with varying
maturity dates, equal balances, equal coupon spreads, an assumed recovery rate of 50 percent and an assumed prepayment rate of 10 percent CPR.
The CLO structure consists of five CLO bonds with a parity ratio of 110 percent. Figure 1 (pg. 44) summarizes the CLO
structure used in this example. We further assumed that the correlation between any two loans is 0.75, and we used Moodys
historic default probability curve for B2-rated loans.
For static modeling, we examined the impact of different CDR
assumptions on the CLO securities. We assumed a constant
CDR rate for the life of the transaction and examined the effect
of increasing the CDR on security losses and resulting prices.
Figure 4 (below) illustrates the impact of different CDRs on
CLO bond losses. It is noticeable that as the CDR increases,
losses are absorbed by the lowest bond first before working
their way up the capital structure. Note that the bonds do not
experience losses immediately as a result of overcollateralization (i.e., a 110 percent parity ratio) and excess spread. The
subordinate bond experiences its first losses when the CDR
exceeds 8 percent annually.

% Bond Loss

Generating Time to Defaults

The process of generating the time to default is shown in
Figure 3 (right). One can start by generating a set of random
numbers from an N-dimensional standard normal distribution
with a correlation matrix , knowing that the generated random numbers are correlated. One can then perform an inverse
mapping of the generated numbers back to the corresponding
cumulative default probability. The determination of the time
to default for each loan is done by mapping the cumulative
default probability determined in the previous step to the corresponding time from the loans probability of default curve.
Next, one can model the loans cash flows under the assumption that the loan continues to perform until the time to default
is reached, at which point the entire loan goes into default.
Any recovery estimate would be projected as a percentage of
the aggregate defaulted amount and realized at a subsequent
date, otherwise known as the recovery lag.
The remainder of the valuation process is similar to the static approach. This process results in a one-price outcome from
a number of possible outcomes. In order to value the security,

one would have to repeat the process a number of times to derive a distribution of possible prices for the security. The steps
in the Monte Carlo simulation are summarized below:
1. Generate random numbers from a correlated standard
normal variable Z~MVN(0,) using a random number
generator.
2. Map Zs to a cumulative probability Yi = -1 (Zi) for each
of the N assets.
3. Map the probability to an appropriate time-until-default
as ti = F-1 (Yi) for each of the N assets.
4. Project cash flows for loans and securities using
assumptions and in accordance with waterfall rules.
5. Obtain a price for the security based on projected cash
flows and an appropriate discount rate.
6. Repeat above steps multiple times to generate a
distribution of prices.
7. Calculate the average of prices obtained from the
simulations.

CD
A

Default Swap (LCDX) index. The traded price of the index

can be used to derive an implied probability of default. A scaling factor can be applied to the historic curve to match the implied probability of default implied by the LCDX index while
maintaining the original historic curve shape.
Although a probability of a default curve can be a tool to
analyze the likelihood of a loan defaulting over time, it cannot
capture the correlation of defaults between loans. To do so,
one can start by constructing a correlation structure (or matrix) that captures the correlation between each loan in the
pool.
Correlation between any two loans can be estimated based
on historical data, such as industry-level asset correlations.
However, there is also a commonly held assumption that all
assets have the same correlation, which can be thought of as
the average correlation expected in the portfolio.
Once the correlation matrix has been determined, a copula
function can be used to incorporate the correlation between
the loans. The copula function is a multivariate distribution
that makes it convenient to link marginal distributions and introduce dependencies between them. Details on copula functions can be found in the Li paper. The most commonly used
copula function is the normal copula, which takes on the form of
an N-dimensional standard normal4 distribution function with
a correlation matrix .

R isk T e c h ni q u es

CD

R isk T e c h ni q u es

6
0 bps

www.garp.com

12
200 bps

15

18

21

400 bps

O C T O B E R 2 0 0 9 RISK PROFESSIONAL

47

R isk T e c h ni q u es

Figure 7: Gaussian Copula Model Bond Loss

Distribution
100

# Occurences

80
60
40
20
0
10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
% of Defaults
Super Senior

48

Senior

Mezzanine

RISK PROFESSIONAL O C T O B E R 2 0 0 9

Subordinate

Figure 8: Gaussian Copula Model Bond Price

Distribution
100
# Occurences

For Monte Carlo simulation, we applied Monte Carlo analysis to our hypothetical deal, using 100 default curve iterations.
The distribution of losses impacting each bond is shown in
Figure 7 (below). We can see that the super senior bond rarely
faces severe losses in the majority of the 100 scenarios. The
expected loss is the average of the 100 scenarios.
As with the static CDR approach, bonds experience losses
according to seniority. The key difference is that stochastic
modeling creates various loss scenarios while the static default
modeling creates only one outcome.
Figure 8 (right) shows the pricing distribution for the stochastic valuation approach. As with the static CDR pricing
example earlier, discount margins were matched with accruing
spreads. The final bond price is the average of the 100 pricing outcomes. Again, it is noticeable that triggers protect the
bonds higher up in the capital structure and that recoveries
also cause super senior bonds to retain value even in severe
loss scenarios.
A benefit of using a stochastic pricing approach is that it
presents a picture of possible prices at different loss severity
outcomes. A static approach, where we are using the best estimate CDR curve to forecasts future collateral losses, would
provide us with only one price point.

R isk T e c h ni q u es

80
60
40
20
0

10%

20%

30%

40%

50%

60%

70%

80%

90% 100%

Price % of Par
Super Senior

Senior

Mezzanine

Subordinate

Model Limitations
The copula-driven stochastic default model described in this
article makes certain simplifying assumptions regarding the
behavior of the underlying loans. It assumes that any two
loans in the portfolio have a fixed correlation coefficient between them. However, default correlations which in essence
measure how two companies are likely to fail at the same time
are time varying in nature. Additionally, the assumption of
a Gaussian Copula structure assigns low probabilities to extreme outcomes.
Over the past few years, market participants have moved toward a single-factor stochastic default model that assigns a single correlation number between all assets in the collateral pool,
partly because this model made it easy to back out an implied
correlation from market prices found in, for example, CDS
trades (which further translated into a simplified explanation
of risk and pricing). This fueled the acceptance of single-factor
model as a market standard, as it allowed more people to trade
these securities and it made structuring new CLO transactions
including complex instruments like CLO/CDO squares
convenient. However, as can be implied from the epic failures
of the credit crisis, people did not fully consider or understand
the limitations of this model completely.
As a consequence of an extended period of benign credit
performance and vast amounts of global liquidity, CLO credit
loss models were calibrated to an overwhelming amount of
positive economic data, resulting in low credit loss expectations and lower probabilities of extreme events. Subsequently,
capital markets rewarded CLO collateral pool diversification

www.garp.com

accomplished through geography and industries (and even

through varying asset-backed issuers), resulting in thinly capitalized CLO structures and overvalued securities.
The stochastic and static default models both fail to address
complex contractual loan features, such as loan conversion options, refinance options, facility extension options, potential
interest rate increases in the event of delinquencies and other
drivers of collateral behavior.5 These features have always
required a more traditional fundamental approach to credit
analysis or more robust default modeling coefficients developed through rigorous regression analysis utilizing full cycle
data sets.
The stochastic model has also proven ineffective in hedging, because it has been unable to capture large movements in
price consistently. The rating downgrade of General Motors
debt in May 2005 is one example on the ineffectiveness of
delta hedging using this model.6
Closing Thoughts
In spite of the limitations associated with the models described
in this article, they are still commonly used for CLO valuations. Using a static CDR curve for CLO valuations is more
commonly used by market participants for fair value accounting purposes, due to its more simplistic approach and easier
interpretation of results. Research is usually available to calibrate market inputs such as losses and prepayment rates.
The stochastic modeling approach requires additional effort
in setting up the model and determining the appropriate probability of default curves, correlation matrices and discounting
spreads. Stochastic models are more effective when one can
regularly calibrate calculated prices to market prices.
Static modeling does result in reasonable valuations when
assumptions are constructed appropriately based on contemporary market consensus on credit, careful monitoring of
both the collateral pool and CLO managers performances,
corroboration (using quality industry research) and price comparisons (using relevant trade proxies from both the secondary
and primary markets).
One advantage of stochastic modeling is the consideration
for distribution of outcomes. Stochastic modeling makes it possible to analyze the distribution of possible prices, as apposed
to a static valuation that produces a single outcome. Unlike
simpler instruments such as interest-rate swaps, CLOs have
complicated structures that include dynamic rules for allocating cash and losses to the bonds, which can have a significant
impact on both duration and pricing.

www.garp.com

For example, a slight increase in losses could trip a trigger,

thereby cutting off cash flows to a particular bond or reducing
the life of another bond. A static valuation with a single outcome would not capture these nuances, but a stochastic model
would be able to capture a distribution of possible loss outcomes and the implication that loan loss scenarios would have
on the respective bonds in the capital structure.
There is a popular saying in statistics, Essentially, all models are wrong, but some are useful.7 This should be kept in
mind when interpreting results from any model. However,
when faced with the task of choosing a model, one has to
weigh complexity and cost, and make a determination on the
most appropriate model for the application at hand while being aware of the limitations a model presents to the user. That
said, investors in structured collateralized products like CLOs
can and should supplement credit loss modeling with sound,
fundamental credit analysis and careful input from a collateral
manager.
FOOTNOTES
1. D. X. Li. On Default Correlation: A Copula Function Approach, The RiskMetrics Group, Working Paper (April 2000). Lis
method has become well known in the CLO industry.
2. F. Salmon. Recipe for Disaster: The Formula That Killed Wall
Street, Wired Magazine (February 23, 2009).
3. D. Duffie. Innovations in Credit Risk Transfer: Implications for
Financial Stability, Stanford University paper (July 2007).
4. A standard normal distribution is a mean zero and variance of
one.
5. A Guide to the Loan Market, Standard & Poors (October
2007).
6. D. Duffie. Innovations in Credit Risk Transfer: Implications for
Financial Stability, Stanford University paper (July 2007).
7. George E.P. Box, a famous British statistician, is thought to be
the original source of this saying.

Anthony Sepci is a partner in KPMG LLPs financial risk management practice, based
in Los Angeles. With more than 15 years of experience, he co-leads a team of more
than 30 structured finance professionals. He can be reached at asepci@kpmg.com.
Dushyanth Krishnamurthy is a manager in KPMG LLPs financial risk management
practice. He can be reached at dkrishnamurthy@kpmg.com.
Christian Eder is a senior associate in KPMG LLPs financial risk management practice. He can be reached at ceder@kpmg.com.
The opinions expressed in this document do not necessarily represent the views of
KPMG LLP.

O C T O B E R 2 0 0 9 RISK PROFESSIONAL

49