3 MAY 2011

MODELING Modeling Commercial Real Estate Loan

METHODOLOGY
FROM MOODY’S ANALYTICS Credit Risk: An Overview
Version 2.0
QUANTITATIVE RESEARCH

Authors Abstract
Jun Chen
Commercial real estate (CRE) exposures represent a large share of credit portfolios for many
Jing Zhang
banks, insurance companies, and asset managers. It is critical that these institutions properly
measure and manage the credit risk of these portfolios. In this paper, we present the Moody’s
Contact Us Analytics framework for measuring commercial real estate loan credit risk, which is the model
at the core of our Commercial Mortgage Metrics (CMM)™ product. We describe our modeling
Americas approaches for default probability, loss given default (LGD), Expected Loss (EL), and other
+1-212-553-1653
related risk measures.
[email protected]
Europe Our framework first models the CRE collateral stochastic process, as driven by both
+44.20.7772.5454 market-wide and idiosyncratic factors. We then apply a Monte Carlo technique to simulate the
[email protected] future paths of the collateral net operating income (NOI) and market value. A CRE loan credit
Asia (Excluding Japan) event is doubly triggered by the collateral financial condition at the time of default: both the
+85 2 2916 1121 sustainable NOI falls below the total debt service, and the property market value falls below
[email protected] the total outstanding loan balance.
Japan
+81 3 5408 4100 Moreover, in order to capture the actual observed borrower default behavior, we empirically
[email protected] calibrate the conditional probability of default (PD) function to large historical datasets. We
also calculate the unconditional EDF™ (Expected Default Frequency) credit measures as the
integration of conditional PD values over the many future paths of NOI and market value. We
model LGD through the same process; therefore, LGD and PD are structurally correlated and
consistently estimated within the same coherent framework. Built upon EDF measures and
LGD, we also calculate other measures such as EL, Yield Degradation (YD), Unexpected Loss
(UL), and Stressed EDF measures and loss. By establishing a strong economic causality
relationship between credit risks and real estate market- and property-specific covariates, the
model also enables large-scale scenario analysis and stress testing.
Additionally, our model facilitates many different business applications, including loan
origination, pricing and valuation, risk monitoring, surveillance, regulatory compliance, and
portfolio management.
2
 Table of Contents

1 Introduction ........................................................................................................................................................... 5

2 Model Outputs and Applications .......................................................................................................................... 6

2.1 Model Outputs ............................................................................................................................................................................................................. 6

2.2 Business Applications ................................................................................................................................................................................................... 7

3 Modeling Framework ............................................................................................................................................. 9

3.1 Understanding Commercial Mortgage Credit Events ..............................................................................................................................................9

3.2 Model Setup and Details............................................................................................................................................................................................ 11

4 Empirical Data ...................................................................................................................................................... 19

5 Model Validation ................................................................................................................................................ 20

5.1 Walk-forward Tests ................................................................................................................................................................................................... 20

5.2 k-fold Tests ................................................................................................................................................................................................................ 21

5.3 Validating Model Calibration ................................................................................................................................................................................... 22

6 Summary ..............................................................................................................................................................23

References ................................................................................................................................................................... 25

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 3

4
 1 Introduction
Commercial real estate (CRE) exposures represent a large share of credit portfolios for many financial institutions,
including banks, insurance companies, and asset managers. According to the Federal Deposit Insurance Corporation
(FDIC), CRE loans comprised about 22% ($1.6 trillion) of all outstanding loans held by U.S. commercial banks and
1
savings institutions as of September 2010. For many insurance companies, CRE loans, with a total balance of
2
approximately $300 billion, represent an important asset class in their investment portfolios. In addition, CRE
instruments comprise a large portion of the underlying collateral backing structured products. Specifically, as of the third
quarter of 2010, there was approximately $706 billion outstanding in commercial mortgage backed securities (CMBS)
3
owned by asset managers, banks, and insurance companies. For these institutions, the risk that borrowers fail to pay
either the interest and/or principal on these CRE loans poses a significant challenge. In fact, significant credit loss from
commercial mortgages can often wipe out the capital cushion and lead to the failure of a financial institution. During the
past 30 years, we have witnessed the failures of numerous financial institutions, both in the late 1980s to early1990s and
also in the recent financial crisis, caused to large degree by CRE-specific loan losses.
Given these challenges, financial institutions continue to seek better risk management of their CRE exposures. Toward
this goal, the first step is to measure the credit risk of these CRE portfolios, including the standalone credit risk
assessment of individual loans, as well as their correlation and concentration effects at the portfolio level. In this paper,
we present the Moody’s Analytics framework for measuring the credit risks of individual CRE loans. Specifically, we
describe our modeling approaches for default probability, loss given default, Expected Loss (EL), and other related risk
measures. For the Moody’s Analytics approach for measuring CRE asset correlation within a portfolio context, see Patel
and Zhang (2009).
In our framework, we begin by modeling the asset process of the underlying CRE collateral. We consider the stochastic
evolution of a commercial property’s financial performance, including income and market value, as driven by both
market-wide and idiosyncratic factors. We first estimate the local market-specific parameters that govern those processes
utilizing extensive historical datasets, then we apply a Monte Carlo technique to simulate the future paths of the
collateral’s net operating income (NOI) and market value. The Monte Carlo technique enables the model to capture the
path-dependency of the survival probability and the remaining credit risks as the future unfolds.
An important model feature is that a CRE loan credit event is doubly triggered by the collateral financial condition at the
time of default: both the sustainable NOI falls below the total debt service, and the property market value falls below the
total outstanding loan balance. Moreover, since the CRE market operates in an opaque environment that is neither
complete nor perfectly efficient, the conditional probability of the default (PD) function is empirically calibrated to large
historical datasets in order to capture the actual observed borrower default behavior. We calculate the unconditional
EDF (Expected Default Frequency) credit measure as the integration of conditional PD values over the many future
paths of NOI and market value.
Finally, we model loss given default (LGD) via the same process; hence, LGD and PD are structurally correlated and
consistently estimated in the same coherent framework. Built upon EDF credit measures and LGD, the model also
calculates other measures such as EL, Yield Degradation (YD), Unexpected Loss (UL), and Stressed EDF measures and
loss. By establishing a strong economic causality relationship between credit risks, the real estate market, and property-
specific covariates, the model also enables large-scale scenario analysis and stress testing. Another significant benefit is our
model’s ability to accurately differentiate the credit risks of senior/junior structure of the multiple loans on the same
collateral, as well as that of mezzanine loans.
In addition to the analytical modeling framework, we have sourced very extensive U.S. commercial real estate market
data and forecasts at the metropolitan and submarket levels, where available. These local market data and forecasts are
embedded within our Commercial Mortgage Metrics (CMM) system, making it a transparent, one-stop solution where
users can consistently measure credit risks across different property types and geographic locations. In the meantime, the

1
FDIC Standard Report #5 (All Commercial Banks–National) as of 9/30/2010.
2
Flow of Funds Accounts of the United States, third quarter 2010.
3
The Compendium of Statistics, update December 30, 2010, published by the CRE Finance Council.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 5

 CMM system enables total control over the inputs on the collateral’s most recent financial statistics, including NOI and
market value (either transaction- or appraisal-based) together with loan characteristics such as coupon rates.
The credit measures produced by our model have many business applications for CRE practitioners. These include risk
assessment and asset selection, risk-based pricing and valuation, risk monitoring and surveillance, regulatory compliance
and internal control, loss forecast and provisioning, scenario analysis and stress testing, and portfolio management. For
example, loan officers and underwriters can objectively and systematically assess the credit risks of a CRE loan located in
any given U.S. market; at the back end, credit risk managers and portfolio managers can quickly monitor the most recent
credit profiles of individual loans, as well as entire portfolios. By establishing a strong economic causality relationship
between credit risks and real estate market and property-specific covariates, the model also enables large-scale scenario
analysis and stress testing. For example, CMM allows users to compare results from a baseline scenario and a stressed
scenario. Users can also input their own commercial real estate market-specific views and test credit risks from those
views.
The remainder of the paper is organized as follows.
• Section 2 describes model outputs and their practical applications.
• Section 3 describes the modeling framework and inner workings of the model.
• Section 4 discusses the empirical data.
• Section 5 documents the model validation findings.
• Section 6 summarizes the paper and provides concluding remarks.

2 Model Outputs and Applications

In this section, we focus on the end results of our CMM model, and discuss how various practitioners can use these
results to make more informed business decisions.

2.1 Model Outputs

The Moody’s Analytics CMM model estimates the credit risk of commercial real estate loans, combining user-provided
portfolios with market-wide data and forward-looking scenarios.
The model provides estimates of the following risk measures, both for a single commercial real estate loan as well as for a
portfolio of loans.
• EDF (Expected Default Frequency) credit measure—measures the probability that a commercial real estate loan
experiences a default event in the future. We estimate EDF credit measures throughout the loan term, and the
model estimates an annual EDF measure for a particular point in time within the loan term. We then calculate
cumulative EDF measures to measure the cumulative holding period risks.
• Loss given default (LGD)—refers to expected loss amount, typically as a percentage of outstanding unpaid loan
balance, at the time of the default event, if the default event occurs.
• Expected Loss (EL)—measures the expected losses of a commercial real estate loan due to default events.
Mathematically, for a given point in time, EL = EDF x LGD. This relationship also holds for cumulative holding
period measures.
• Yield Degradation (YD)—measures the annualized reduction of expected yields from a commercial real estate loan
due to losses related to default events throughout the loan term. YD is similar to the measure of annualized EL, with
the main difference being that YD takes into account the timing of expected losses and discounts losses according to
the timing, whereas, annualized EL does not involve discounting and time-value of loss.
• Unexpected Loss (UL)—defined as a one standard deviation of loss from the loss distribution. We estimate the one
standard deviation of loss based on a full range of loss distribution derived from Monte Carlo simulations of all
possible combinations of systematic market risk factors and non-systematic idiosyncratic loan and property-specific
risk factors.

6
 • Stressed EDF measure and loss—measures the point estimate of EDF measures or loss from a full range of EDF
credit measure or loss distribution derived from Monte Carlo simulations. Typically, we measure the Stressed EDF
at a user-specified stressed point, such as a confidence level greater than 50%, for the tail risk at the right-hand side
of the distribution.

2.2 Business Applications

Underwriters, credit officers, risk managers, and portfolio managers can use Moody’s Analytics CMM for a variety of
different business applications. For institutions that employ internal rating systems as the foundation of many business
decisions, they can either map the EDF credit measures and LGD outputs to their internal rating scales, or combine
them with other qualitative inputs to derive an internal rating. Alternatively, they can use CMM to benchmark and
calibrate their own internal risk rating systems.
CMM applications include risk assessment and asset selection, risk-based pricing and valuation, risk monitoring and
surveillance, regulatory compliance and internal control, loss forecast and provisioning, scenario analysis and stress
testing, and portfolio management.

Using CMM in Internal Ratings Systems

Internal rating systems serve as the foundation of many business decisions within financial institutions such as credit
approval, limit setting, regulatory compliance, risk-based pricing, and active portfolio management. An effective internal
rating system has the following attributes.
• Separates default and recovery risk.
• Provides powerful differentiation of relative ranking of risk.
• Well-calibrated to provide appropriate distinctions of risk.
• Contains well-documented definitions, assumptions, and methodologies.
• Combines qualitative assessment and quantitative assessment where appropriate.
We developed CMM with the above attributes in mind. CMM measures CRE loan default risk via EDF credit measures
and recovery risk via its LGD. As shown later in this paper, the model proves to be powerful, forward-looking, and
accurately calibrated to real loss experience. For documentation, this introductory paper, together with the detailed and
comprehensive modeling documents, provides model transparency for users. Therefore, CMM is ideal as a
quantitatively-based internal ratings system for CRE exposures. If institutions use internal rating scales not generated in
absolute scales, such as PD and LGD, they can map the EDF credit measures and LGD outputs to their own internal
rating scales.
Many institutions find that market-based information, when available, is particularly relevant and powerful in internal
risk rating assessment. CMM credit measures utilize a significant amount of CRE market information, and we construct
them to reflect all the relevant property type and location-specific market information. Thus, if the user finds it
appropriate to combine qualitative assessment with quantitative components, CMM credit measures are particularly
useful as the market-based quantitative assessment component of an internal rating system. In fact, we make it feasible
and efficient to implement such an approach in the Moody’s Analytics RiskAnalyst™ system. An internal rating system
must provide sufficient differentiation of default risk. To calibrate such a system, regulators typically expect a sizeable
amount of realized default events and loss severity data spanning at least a full economic cycle. Many institutions do not
possess enough internal data and can benefit from using CMM credit measures to benchmark and calibrate internal risk
systems.

Risk Assessment and Asset Selection

CMM can be very effective during an institution’s credit underwriting process. Commercial mortgage underwriters and
credit officers can benefit significantly by using CMM to directly measure and compare credit risks at loan origination
for given loan and property characteristics. For example, CMM allows users to run multiple “what-if” analyses to
compare how credit risks, including PD, LGD, and EL, would change if either one or both the Debt-service Coverage

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 7

 Ratio (DSCR) and loan-to-value (LTV) change. Underwriters can use this information to risk-base price loans according
to a specific combination of DSCR and LTV.
Another invaluable feature of CMM is its embedded local real estate market data and forecasts, which make it possible to
compare loan risks across different property types and geographic locations. For example, a national financial institution
often conducts CRE lending in many locales throughout the country, and individual underwriter expertise and
assessment may vary significantly between local offices. CMM enables centralized credit risk management to objectively
and consistently measure credit risk without overly relying upon individual judgments from dozens of or even hundreds
of different credit underwriters.

Risk-based Pricing and Valuation

A financial asset such as a commercial mortgage must be appropriately compensated for its given risks. Because
commercial mortgages fit well in held-for-investment portfolios, the long-term credit risks become, de facto, the most
important source of risk. As such, CMM can help determine the trade-offs between loan pricing and future risks. This
type of risk-based pricing and valuation can be performed both at the primary market and the secondary market levels.

Risk Monitoring and Early Warning

The credit quality of commercial mortgages can change quickly as the market environment or property-specific
conditions change. Because the credit risk measure outputs from CMM are objective and forward-looking, risk managers
can target their risk assessment and mitigation resources toward cases where they can be the most effective. Annual
reviews and other traditional credit processes cannot maintain the same degree of speed, consistency, and objectivity.
Within CMM, accurate and timely information from the commercial real estate market can be applied consistently
across the entire portfolio, which is often difficult and expensive to duplicate using traditional credit analysis processes.

Regulatory Compliance and Internal Governence

The probability of default associated with an internal rating plays a central role in the calculation of capital requirements
within the Basel II framework. Banks may use external PD models, such as EDF measures from CMM, as part of their
internal ratings, either for regulatory capital calculations or for fulfilling internal governance and external regulatory
requirements.

Loss Forecast and Provisioning

Loan loss provisions are expenses charged to a bank’s earnings when adding to the allowance for possible bad debt. In
estimating the provisioning amount, one can use a credit risk model to estimate the potential credit losses on loans. The
model should respond to changes in the risk environment across the economy as a whole. In other words, a provisioning
calculation should be as forward-looking as possible. In fact, both the International Accounting Standard Board (IASB)
and the Basel Committee on Bank Supervision are moving toward the more forward-looking “expected loss” approach
and away from the “incurred loss” approach (e.g., Basel Committee on Banking Supervision, 2009). All CMM credit
measures are forward-looking assessments that respond to changes in the CRE market cycle and produce accurate
estimates of credit losses over a long period. Consequently, these credit measures are appropriate for expected loss-based
provisioning calculations.

Scenario Analysis and Stress Testing

The future remains inherently uncertain. No single person or entity, nor the market as a whole, possesses a crystal ball
that predicts exactly what will occur. We built the CMM system so that it contains several embedded commercial real
estate market forecast scenarios, and also allows users to input their own views regarding specific property types and/or
geographic locations. Such functionality is particularly valuable for risk managers when comparing possible outcomes
from different economic outlooks. Forecast scenarios are also becoming more and more a daily business necessity, given
increased regulator and internal risk controller requirements for periodic stress tests. The CMM on-demand scenario
analytic capabilities can significantly improve an institution’s readiness to meet such continuous demands.

8
 Portfolio Management
Portfolio management entails making numerous decisions, such as taking on additional exposures, selling or hedging
existing exposures, and calculating the prices at which to do so. The Moody’s Analytics CMM system provides a
framework that enables you to make informed decisions regarding which loans to create, under what terms, and at what
price(s). In addition, risk managers can use CMM portfolio functionality to construct strategies that exploit the relative
price differences between property types and local CRE markets. Additionally, the EDF measures and LGD outputs
from CMM can serve as inputs to calculations performed by portfolio management systems such as Moody’s Analytics
RiskFrontier™.

3 Modeling Framework
In this section, we first describe how credit events occur for commercial real estate loans in the real world. We use an
example to illustrate the importance of the collateral financials in affecting a loan’s credit risk. Next, we present the
conceptual framework as well as details of the model’s inner workings, including specifics on the asset process, the PD
model, and the LGD model. We then describe how the components work together within the CMM system. Finally, we
explain how CMM implements scenario analysis and stress testing.

3.1 Understanding Commercial Mortgage Credit Events

Since our objective is to accurately measure the probabilities of a credit event occurring and the resulting losses associated
with the credit events, first and foremost, it is important to examine why and how credit events and losses happen in the
real world. We want to make sure our model succinctly and consistently emulates real world phenomena and captures its
essence.
Why would any commercial real estate loan borrower default on their debt obligations? In principle, there are two
primary reasons under the so-called double trigger framework. The first is that cash flow from the property is inadequate
to cover the scheduled mortgage payment; the second is that the underlying commercial properties, which serve as the
secured collateral for most commercial real estate loans, are worth less than the mortgages. In other words, a commercial
mortgage borrower’s ownership value, inclusive of property resale value , plus current and future incomes, less the market
value of the mortgage (including current outstanding payments), becomes less than zero in the event of default. We
point out that the borrower’s equity value is its economic value and takes into account embedded options, so it may
differ from the book equity measure. Also, because commercial real estate is an asset class primarily focused on producing
an inflation-adjusted rental income stream while preserving capital value, it is more productive to separate and focus on
the income side of the ownership value. We illustrate the double trigger framework with the following example.

Double trigger framework example

When a commercial mortgage is originated, the mortgage lender typically requires cushions in both leverage and debt
service coverage. For example, a LTV ratio of 70% and a DSCR of 1.30 may be the threshold underwriting criteria for a
particular lender. Under this threshold, the maximum loan amount a borrower can obtain is $7,000,000, if the market
value of the property is worth $10,000,000; and the maximum annual debt service a lender would allow is $538,462, if
the property is currently generating $700,000 in annual net operating income (NOI). In fact, since both ratios need to
satisfy the threshold underwriting ratios, the actual mortgage may either carry a loan amount of less than $7,000,000 or
the annual debt service is less than the $538,462. The point here is that most commercial mortgages, if underwritten
appropriately and absent of fraud, should, in theory, carry no or very little credit risk at origination. What drives the
credit risk is the inherent future uncertainty, which can potentially be quantified.
For simplicity, assume that a commercial mortgage originates with $7,000,000 loan amount, with annual debt service of
$538,462, based on a $10,000,000 property generating $700,000 NOI a year. The realization of future NOI of the
property is unknown and can follow an infinite number of possible paths. In the particular NOI path illustrated in
Figure 1, there are periods around points A and B where the collateral’s NOI is not sufficient to cover mortgage
payments.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 9

 NOI

Trend

Mortgage payment

0 A B Maturity Time

Figure 1 Evolution of a collateral property’s NOI

Whenever a property is not generating enough NOI to cover the periodic mortgage payment, a borrower must weigh the
different options, as follows.
• Cover the payment shortfall from their own pocket, if the shortfall is deemed temporary and will be cured.
• Sell the underlying property and pay back the entire remaining mortgage balance, including outstanding interest
payments, if the market value of the underlying property is enough to cover all debt obligation plus a non-trivial
transaction cost.
• As a last resort, miss the mortgage payments and wait for the lender’s decision to either foreclose or restructure debt.

Property
Value

Trend

Mortgage balance

Time
0 A B Maturity

Figure 2 Evolution of a collateral property’s value

Obviously, what is also very important in this situation is the market value of the property, pledged as secured collateral,
which the lender can take possession of in the event of default. While the property value is usually correlated with NOI,
its evolution is also affected by the general conditions in both capital and space markets, in addition to the property-
specific NOI. In our example, with the particular NOI realization as in Figure 1, the property value does not necessarily
follow the NOI movement in lockstep.

10
 As illustrated in Figure 2, the property value drops below the mortgage balance around point A, but not point B.
Toward the end, the property value declines again around loan maturity even though the property’s NOI remains well
above the scheduled mortgage payment amount, shown in Figure 1.
It is straightforward to make the following observations with this particular example, as shown in Table 1.

Table 1 Decision analysis on default probabilities

Decision Point DSCR LTV Decision Analysis Credit Risk Measurement

A < 1.0 > 100% High probability of default EDFA

B < 1.0 < 100% Not a clear-cut default choice EDFB

Maturity > 1.0 ≈ 100% High refinancing risk EDFMaturity

While the above discussions illustrate the financial aspects of CRE borrower default drivers, we should note that decades
of actual experience with commercial mortgage defaults also clearly teach us that a borrower’s decision to default is not
purely a financial matter. For a CRE asset that is illiquid and difficult to value and to sell easily, the borrower’s decision
to default is influenced by both financial facts and subjective assessment of the situation, leading to the so-called “sub-
4
optimal” (non-ruthless) default behaviors observed at the aggregate level. We emphasize here that a vast majority of
borrowers do make very rational and near-optimal decisions regarding defaults; it is the inability to observe and record
many loan, property, and borrower-specific decision factors that lead to empirically-observed, “sub-optimal” default rates
in aggregate. Furthermore, even perfectly explainable and rational behavior on the individual level can still appear to be
“sub-optimal” using aggregate data alone.
To create a credit risk model that is relevant for business users, we must anchor the analysis on empirical data. While
theoretical thinking is very useful as a starting point in disentangling the causes and consequences from a rational
economic-reasoning perspective, a model will not be accurate and useful for business applications if it is simply an
intellectual exercise without the benefit of actually-observed historical data. Thus, our approach to modeling CRE credit
events is to combine rational economic reasoning perspectives with insights gleaned from careful analysis of empirical
data.

3.2 Model Setup and Details

Because commercial mortgage credit events fundamentally depend on the realized financial conditions of the underlying
collateral and the market environment it operates within, our modeling process starts by understanding and quantifying
the dynamic processes and uncertainties surrounding CRE assets. In addition, at the time of financial distress, because
commercial mortgage borrowers make default decisions based beyond purely financial considerations, we employ
empirical data and statistical analysis to measure the “sub-optimal” or “behavioral” aspect of mortgage defaults. In other
words, our modeling framework can be thought of as consisting of two major parts: the underlying commercial
properties’ stochastic dynamics following a structural approach; and empirically calibrated default probability
measurements, conditional upon realized property financials.

4
Note here that “sub-optimal” is viewed from the borrower’s perspective. In fact, since the lender takes the opposite position, the
“sub-optimal” borrower’s behavior actually adds value to the lenders’ CRE portfolio. Because, in aggregate, CRE borrowers do not
exercise their default options ruthlessly, most lenders’ CRE operation has been able to live through the cyclical troughs of the CRE
market downturns without being completely destroyed.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 11

 Figure 3 The CMM modeling framework consists of two major parts

Mathematically, a commercial real estate loan’s EDF credit measure at a particular point in time t is:

𝐸𝐷𝐹𝑡 = � 𝑃𝑟𝑜𝑏(𝑋𝑡 ) ∙ 𝑃𝑟𝑜𝑏(𝐷𝑒𝑓𝑎𝑢𝑙𝑡|𝑋𝑡 ) (1)

where Xt denotes all the relevant financial variables at the loan, property, and market levels measured at time t,
Prob(Xt) measures the probability of a particular realization of Xt,

and Prob(Default|Xt) is a conditional default probability function given the realized Xt. EDFt, (i.e. the unconditional
EDF), is simply an integration of the probabilities of all possible realizations of Xt multiplied by corresponding
conditional default rates for those realizations.
We find that this composite modeling approach leads to the most effective credit risk model by combining the best of
both worlds. On the one hand, the dynamics of variable, Xt, follows a structural stochastic process that is fully
parameterized based on extensive historical observations of the commercial property financials. On the other hand, the
calibration of the conditional default rate function Prob(Default|Xt), derived from rigorous statistical analysis, captures
the “sub-optimal” exercise of the default options, and therefore, produces accurate and realistic EDF measures. This
composite approach is similar in spirit to the Moody’s Analytics EDF credit measure model approach for publicly traded
firms, as that model is also a structural approach with a robust implementation grounded in empirical data.
Another benefit of our modeling approach is that it naturally leads to LGD measures that are economically and
structurally correlated to PD, because LGDt can be estimated as another conditional function that draws Xt as the
dependent variables. This method offers significant improvement over an ad hoc approach to approximate the
relationship between PD and LGD.
In essence, our CMM model consists of three key elements.
• Parameterizations of Asset Dynamics and Volatility (the Asset dynamics model). In this step, the CRE collateral’s
NOI and value processes in the future are parameterized, based on its property type and geographic location, in
conjunction with the known financial and leasing information at the starting point.
• Calculation of EDF Measure (the EDF model). At a future time t, given realized NOI and values generated by
Monte Carlo simulations, the model first calculates the default drivers, including DSCR and LTV, and then
estimates conditional PD through the conditional default rate function Prob(Default|Xt). Provided by the known
distribution characteristics of NOI and value from the first step, the model then estimates unconditional EDF
measures as well as Stressed EDF measures as point estimates from the full range of conditional PD distributions.
• Calculation of LGD (the LGD model). Given a simulated market value of the collateral at a future time t, the
model estimates conditional LGD through an empirically-determined loss function. An unconditional LGD is the
weighted sum of conditional LGD values, with weights being the corresponding conditional PD values.
We can also easily calculate other risk measures such as EL, UL, and yield degradation since the model performs a full
Monte Carlo simulation from which EDF measures and LGD have been calculated. The remainder of this section offers
a brief introduction to these key model elements.

12
 3.2.1 Asset Dynamics and Volatility
Asset dynamics refers to the inner workings and quantifiable causal relationships of the commercial properties’ financial
performance, and asset volatility refers to the uncertainty around the financial performance. While conceptually one can
measure all sorts of financial-related variables including rents, occupancy rates, revenue, expense, NOI, capital
expenditure, and market value etc., the direct drivers behind commercial mortgage defaults are mainly DSCR and LTV,
which are, in turn, driven by NOI and the property’s market value.
Modern financial theory views any asset, including commercial real estate, as having two independent sources of risk
drivers:
• One related to the overall market movement, also known as market or systematic risk
• Another related to the specifics of individual assets, also known as non-systematic or idiosyncratic risk
Intuitively, this separation is no different than the commonly practiced attribution analysis, where both performance
(returns) and risks (volatility) can be traced back to either market-wide trends or asset-specific conditions. It has been
established that the process of a commercial property’s income or value can be approximated as follows:

𝑃𝑖,𝑡 = 𝑃𝑚,𝑡 + 𝜀𝑖,𝑡 (2)

Systematic Non-systematic
component component

where Pi,t represents the realization of NOI or value in log form for the ith property at time t, and Pm,t represents the
market-wide index at the same time t, and εi,t is the idiosyncratic component of the Pi,t movement that remains after
stripping out the market-driven component.
The specification also means that, in terms of risk composition,

Variance [𝑃𝑖,𝑡 ] = Variance [𝑃𝑚,𝑡 ] + Variance[𝜀𝑖,𝑡 ] (3)

Systematic Non-systematic
Total risk risk risk

The importance of a risk model when considering both systematic and idiosyncratic risks is illustrated in Figure 4, which
shows a typical CRE asset that displays substantially higher total risk than market risk alone.

Market
Total volatility
volatility only

95% of
68% of realizations
realizations

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Year Year
Figure 4 Distributions of NOI amounts or values

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 13

 Because real estate is a very location- and property type-specific business, the market here is defined by property types
and metropolitan areas. In other words, the San Francisco office market is considered a distinct market from the San
Francisco apartment market or the New York office market, and so on. Drawing from a large historical time-series
database, which includes both market-wide performance statistics and property-specific operating financials covering a
substantial portion of the U.S., we can estimate and parameterize the asset dynamics and volatilities as specified in
Equations (2) and (3) for most of the active CRE markets in the U.S.
To help understand how CMM implements the Monte Carlo simulations of the collateral NOI, we provide an example
of the actually observed property-level NOI values of office properties in the San Francisco office market.

300

250

200

150

100

50

-50
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009

Figure 5 Property-level observations of normalized NOI for San Francisco office properties.

Since DSCR is determined to be the most important empirical variable, in CMM Monte Carlo simulations, we focus on
the NOI simulations based on the observed historical patterns by property type and metropolitan areas, such as the one
shown in Figure 5. In particular, we simultaneously simulate the random realization of independent factors: the market-
wide and the property idiosyncratic factors. For example, as shown in Figure 6, the market factor could rise, stay flat, or
lower in the future.
Meanwhile, a particular realization of the collateral’s NOI can also deviate from the market factor by idiosyncratic
variations. Given a particular realization of the market path, the final realized NOI could be better than, the same as, or
worse than the market-wide growth rates. We have observed in the actual model implementation that the end results of
CMM’s NOI simulations for the San Francisco office market are indeed quite similar to that shown in Figure 5, which
confirms the validity and robustness of our collateral models.

Figure 6 The asset risk of a commercial property comes from both market and idiosyncratic sources

14
 3.2.2 Modeling Conditional Default Probability
Once we fully specify the asset processes, the next problem is to solve the probability of borrowers choosing to default,
based on a particular realization of NOI and market value of the underlying collateral. Fundamentally, this is a question
about the conditional default probability. Note, that in an abstract structural default modeling approach, such as the
Merton model, a loan would automatically default once the market value of the asset falls below that of the mortgage,
since the amount of the debt serves as the absorbing boundary. While still incorporating this powerful notion, our
modeling framework expands to include the following very important practical considerations, making the model truly
relevant in the real business world.
• The market value of a specific asset is actually unknown and is a somewhat subjective measure in the case of
commercial real estate. As such, in an empirical model, we must resort to other directly-observable measures to
complement imprecise measures of the asset value. It is also important to recognize that a periodic income stream
should be explicitly factored into the valuation equation of a CRE asset.
• Net operating income (NOI), which is directly observable and ubiquitously measured and recorded, is of
predominant financial and decision-making importance to commercial mortgage borrowers. Notice that a particular
CRE asset is fundamentally an income-producing asset, and, unlike a corporation, it has no potential to grow its
business base, since its physical size and location are fixed once built. As we find throughout our empirical work, the
debt-service-coverage ratio (DSCR), which measures the level of NOI relative to the periodic mortgage payments,
explains a large portion of the historical default incidents.
• Defaulting (or not) is a borrower’s choice, rather than a strict rule that must be followed. While financial factors
such as DSCR and LTV are of critical importance, choices are behavioral in nature and subjective to both
quantifiable and non-quantifiable factors. Note here, we consider that the conditional default rate is not only a
default probability conditional upon known financial and operating ratios, but it is also a probability of a borrower
choosing to default. To date, it appears to us that the best way to model this kind of behavioral issue is through
extensive statistical analysis using large panel datasets, which is exactly what we incorporate.
Effectively, our conditional default probability model is a multi-factor empirical function that links empirical default
rates to key explanatory variables at both the property/loan and the market levels. Our empirical research identifies the
following key statistically-significant variables in explaining historical defaults.
• Asset level financial ratios: DSCR, LTV, and collateral size
• Market cycle factors: vacancy rate, market-wide price changes, and market condition at origination
• Other: core vs. non-core collateral and loan seasoning
The functional model form looks like:

𝑃𝑟𝑜𝑏(𝐷𝑒𝑓𝑎𝑢𝑙𝑡|𝑋𝑡 ) = 𝑓[𝑇1 (𝐷𝑆𝐶𝑅𝑡 ), 𝑇2 (𝐿𝑇𝑉𝑡 ), … ] (4)

where Xt denotes all the explanatory variables listed above,

T1 and T2 are semi-parametric transformations of DSCR and LTV,
and f is a logistic function to capture the non-linear relationship between PD and explanatory variables.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 15

 Table 2 Explanatory variables in the conditional default probability model
No Category Measurement Relationship
1 Debt Coverage Monotonously transformed DSCR (debt- Lower DSCR leads to higher PD.
service-coverage ratio).
2 Financial Leverage LTV (loan-to-value ratio) Higher LTV leads to higher PD.
3 Systematic Factor Standardized market-wide vacancy rate Weak space market (higher vacancy rate) leads
(Space Market) (i.e. z-score for space market). to higher PD.
4 Systematic Factor Standardized market-wide price change Weak capital market (slower and negative price
(Capital Market) (i.e. z-score for year-to-year price change). appreciation) leads to higher PD.
5 Origination Quality Transformed market-wide vacancy rate at Loans originated in stronger CRE market
loan origination. environment tend to have higher PD due to
loosening of the underwriting criteria.
6 Collateral Size Percentile of property size for Larger collateral leads to higher PD.
corresponding types.
7 Seasoning Transformed loan age PD peaks around year 3-7.
8 Property Type Dummy for non-core property types, such Non-core property types have higher PD due to
as hotels. higher operating business characteristics.

As shown in Table 2, in addition to loan and property level factors, market-wide factors play a significant role. We use
the market vacancy rate as a proxy for the contemporaneous space market condition, the market-wide price change as a
proxy for the commercial real estate capital market condition, and the market condition at origination to approximate
the average underwriting quality. That is, when the commercial property market is tight and experiences low vacancy
rates, myopic lenders tend to loosen underwriting criteria and admit more lower-quality loans than when the property
market experiences high vacancy rates.
The other explanation is that these market factors serve as proxies for the option value of borrower equity positions.
Given the same DSCR and LTV, when the market is good, the option values of borrower equities tend to be higher (at
least from a regular borrower’s viewpoint) than when the market is bad. In other words, if the DSCR is 0.8 and LTV is
100%, the troubled borrower is more likely to hold onto the property without defaulting then when the general market
condition is favorable, and the same borrower is more likely to default given the same DSCR and LTV then when the
prevailing market condition is deteriorating.
In general, for the conditional default probability model, we find that the variables we identify lead to accuracy ratios
(AR) of 50% to 60% or more (equivalent to ROC ratios of 75% to 80% or more), both in- and out-of-sample.

3.2.3 Modeling Loss Given Default

Loss given default (LGD) measures loss severity if the loan is already in default. When a commercial mortgage is in
default, if the lender decides to first take possession of the collateral property and then dispose of it to recover mortgage
principal, then the LGD is simply a function of the disposition value of the collateral in relation to the unpaid loan
amount plus transaction and administrative cost. The lender may decide to work out and restructure the loan if the
expected LGD is too high by taking the foreclosure route, but the decision to restructure largely depends upon the
perceived collateral liquidation value. Our LGD model is a sum of two components.
• Loss from principal. This is the loss due to the difference between a collateral property’s liquidation value and the
face unpaid principal balance of the commercial mortgage.
• Loss due to costs and expenses, including, but not limited to, lost interest, transaction cost, legal and administrative
expenses, and property maintenance and renovation cost, etc.
In other words, a commercial mortgage’s LGD can be expressed as:

16
 𝐿𝐺𝐷𝑡 = 𝑔[𝑇3 (𝐿𝑇𝑉𝑡 ), 𝑌] (5)

where T3 is a transformation function of LTV to make it linear to LGDt ,

Y denotes the empirical variables proxy for disposition costs,
and g is a linear specification.
From a theoretical standpoint, the principal loss part of a commercial mortgage’s LGD is simply the ratio of principal
shortfall from collateral liquidation and remaining loan balance, i.e.,

𝐶𝑜𝑙𝑙𝑎𝑡𝑒𝑟𝑎𝑙 𝑉𝑎𝑙𝑢𝑒 − 𝐿𝑜𝑎𝑛 𝐴𝑚𝑜𝑢𝑛𝑡 1

𝐿𝐺𝐷𝑃𝑟𝑖𝑛𝑐𝑖𝑝𝑎𝑙 = − = −1 (6)
𝐿𝑜𝑎𝑛 𝐴𝑚𝑜𝑢𝑛𝑡 𝐿𝑇𝑉

LGDPrincipal should be an inverse function of LTV if the CRE market is efficient, with negligible transaction costs. Because
the CRE market is not as efficient and carries significant transaction costs, the liquidation price of distressed commercial
properties tends to trade below the market price of comparable non-distressed assets. We find that the degree of distress
in the actual loss severities data could be partially proxied by another observable variable: time to liquidation.
Drawing upon a large and updated historical LGD dataset, we can confirm the model’s empirical validity. Using actual
observable variables, empirical evidence clearly supports the statistical and economic significance of LTV, time to
liquidation, and collateral size. Beyond what is explained above, time to liquidation is directly correlated to lost interest,
transaction, and property maintenance cost. Collateral size captures the percentage cost of some relatively fixed expenses
on the legal and administrative sides. For example, $1 million legal and administrative cost would add 2% LGD to a $50
million loan, while the same cost would add 10% LGD to a $10 million loan. Altogether, our empirical model explains
about 50% of the historical loan-level LGD variation.

Table 3 Explanatory variables in the conditional LGD model

No Category Measurement Relationship
1 Loss on Principal 1/LTV Higher LTV leads to higher LGD.
2 Loss on Interest and Time to resolution (time lag between first Longer time to resolution leads to higher
property maintenance cost. missed payment and eventual resolution). LGD.
3 Administration and legal Percentile of the collateral size by property Larger collateral corresponds to slightly
cost. types. lower LGD due to some fixed costs on
asset liquidation.

Since the implementation of CMM model uses Monte Carlo simulation techniques, we can simulate numerous future
asset income and value paths. These simulated asset values provide a direct measure of loss from principal, thus enabling
an LGD measure that is also structurally and causally correlated to the EDF measure.

3.2.4 Putting It All Together

In summary, our CMM model applies an asset-based composite approach that combines a structural asset evolution
process and empirically calibrated functions for both EDF measures and LGD. The steps to calculate EDF and LGD
measures for a given commercial real estate loan include the following.
1. Simulate a large number of asset income paths and values by considering both systematic and idiosyncratic
uncertainties.
2. Calculate a series of DSCR, LTV, and other explanatory variables along each simulation path.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 17

 3. On each simulation path i, estimate its conditional EDF and LGD measures at time t, EDFi,t and LGDi,t, given the
simulated realizations of DSCRi,t, LTVi,t and other contemporaneous explanatory variables using Equation 4 and
Equation 5.
4. Calculate unconditional EDF measure at time t, EDFt, by using Equation (1).
5. Calculate unconditional EL at time t, ELt, using the formula:

∑𝑁
𝑖=1(𝐸𝐷𝐹𝑖,𝑡 × 𝐿𝐺𝐷𝑖,𝑡 )
𝐸𝐿𝑡 = (7)
𝑁

where i refers to individual simulation path that has equal probability of asset realization, and N is the total number of
simulation trials. Unconditional LGD at time t, LGDt, follows by LGDt=ELt/EDFt.
One key benefit of implementing the Monte Carlo simulation technology is the resulting full-range distribution of
conditional PD and conditional loss rates through a large number of random draws. Very naturally, from there we obtain
all the point estimates regarding UL and various stressed EDF measures and stressed loss rates at any user-specified
confidence levels.

3.2.5 “What-if” Questions and Scenario Analysis

As our model adopts a strong economically sensible causative specification during each step, it provides users with a
powerful and reliable framework in which they can ask “what-if” questions regarding the model’s various inputs and then
examine the effects on credit risks of any input changes. For example, a user may be interested in comparing the EDF
measures between the baseline economic forecast scenario (S0) and a stressed economic forecast scenario (S1).
Furthermore, the user may also want to test the differences of EDF measures if the future CRE market volatility differs
from the past, which is a legitimate exercise, given the vast amount of literature pointing to the existence of time-varying
volatilities. It would be of tremendous business value to complete the report shown in Table 4 by leveraging CMM as
one of the main tools.

Table 4 Scenario analysis over the holding period: an example of model flows
Macroeconomic CRE Market CRE Asset CRE Portfolio CRE Portfolio
Scenario Assumptions Assumptions Volatility EDF EL

S0: GDP0 NOI growth0 1.0 times EDF0 EL0

historical
Base case Unemployment0 Value growth0 market vol
…… ……

S1: GDP1 NOI growth1 x times EDF1 EL1

historical
Stressed case Unemployment1 Value growth1 market vol
…… ……

Macroeconomic CRE Market Forward-looking CMM Output:

Outlook Outlook Volatility Credit Risk Measures

Users can change one or all three categories of forward-looking assumptions to conduct scenario analysis. In addition,
both macroeconomic and CRE market outlook scenarios can be either direct user inputs or sourced to third-party
vendors within the CMM application. In fact, there are a number of third-party vendors who provide forward-looking
macroeconomic and CRE market scenarios as fee-based forecasting services. Users may find it cost-effective and
informative to take advantage of the available location-specific forecasts from those forecasters. For example, CMM users
can use the CRE markets forecasts provided by CBRE Econometric Advisors (EA) and/or macroeconomic forecasts
provided by Moody’s Economy.com.

18
 Moody’s Analytics does not endorse any third-party forecast services nor validates their accuracy. Rather, in absence of a
large, publicly traded market for the CRE asset class, we find that the forward-looking views given by industry experts
provide useful perspectives, although the eventual market movement may occasionally deviate from their actual forecasts.

4 Empirical Data
A model is only as good as the empirical data allows. The shortage of long-term, detailed reliable data poses a major
challenge for developing a credit risk model for commercial real estate loans. In the process of developing our model, we
assembled a collection of datasets covering different aspects of the CRE asset and loan markets. Collectively, these
datasets serve as the foundation for both model development and model validation.
The datasets used in CMM model development and validation provide both aggregate market statistics and property-
and loan-level information. Data sources include the following.
• National Council of Real Estate Investment Fiduciaries (NCREIF)—provides aggregate market statistics, mainly
capital appreciation, income, and total returns, for five major property types and more than 50 metropolitan areas.
Some series are offered at the property subtype levels. Despite some shortcomings, the NCREIF dataset is widely
regarded as the industry benchmark measurement in terms of aggregate market movement. The earliest data points
go as far back as 1978 in the NCREIF database.
• CB Richard Ellis Econometric Advisors (CBRE EA)—provides aggregate market statistics, mainly market rents and
vacancies, at the MSA and submarket (is applicable) levels. CBRE EA market rent data are probably the most well-
constructed rent indices, which carefully control for many idiosyncrasies in the actual observed lease rates. The
earliest data from CBRE EA goes back to 1980 for office and industrial properties. CBRE EA data also covers
multifamily, retail, and hotel property types.
• Real Capital Analytics (RCA)—provides disaggregated property-level transaction data on price and cap rates (if
available) for any commercial property transactions valued over $5 million in the U.S. One can examine the
aggregate statistics at any level as long as sufficient transactions exist. This data powers the Moody’s/REAL
Commercial Property Price Index (CPPI), and the earliest observations begin in 2000.
• Trepp’s CMBS Deal Library—one of the largest commercially available databases of the U.S. CMBS universe. The
Deal Library contains comprehensive information and history on both individual loans and properties that serve as
collateral within the CMBS transactions. Since CMBS is a relatively young segment of the commercial mortgage
market, the useful history for empirical research goes back approximately ten years.
• American Council of Life Insurers (ACLI)—publishes periodic reports on the commitment profiles and credit
performance of commercial mortgages originated and held by life insurance companies. ACLI data offers the longest
aggregate time-series of loan performance spanning several CRE market cycles, beginning in 1965. Hence, this
database serves as an invaluable data source for the most credible through-the-cycle analysis for held-for-investment
commercial mortgages.
• Federal Deposit Insurance Corporation (FDIC)—publishes quarterly bank-specific reports that detail the holdings
and credit performance of banks’ loans and leases, including longer-term commercial mortgages and short-term
construction loans. Data goes back to 1992, and provides a unique perspective as it is the only public source for the
credit performance of the CRE loan portfolios held by banks and savings institutions, the largest players in the CRE
lending marketplace.
• Proprietary data contributed by financial institutions—through its business relationships, Moody’s obtains
proprietary access to historic loan-level data from lenders in the commercial mortgage business. The earliest
observations of such datasets go back as far as the 1970s.
• Various published studies and reports, including the series of studies conducted by Snyderman and Esaki et al. The
unique importance of the Snyderman-Esaki studies lies in its tabular mortality tables that track individual loan
cohorts from cradle-to-grave in terms of default rates. These loan cohorts were originated from 1972 through 1997
by major life insurance companies and experienced 2,700 defaults (15.3%) out of the 17,978 total loans originated.
The data is widely used by market participants in benchmarking and in conducting scenario analyses throughout
cycles.
While each of the above datasets is indispensible for our model development, none is completely sufficient on its own.
We take a mosaic approach that pieces all the information and empirical analysis together within the coherent

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 19

 overarching framework explained in Section 3. The final CMM model is the result of exhaustive empirical analyses and
careful triangulation from complementing insights garnered from multiple data sources.

5 Model Validation
Validating a quantitative risk model is both a theoretical and an empirical problem. By theoretical, we refer to the
general guidelines of model development, which involve not only academic rigor, but also sound and time-proven
business experience. Validation refers to whether or not the model makes econometric sense as well as common business
sense, and also whether or not the model specifies an economically-sensible and long-lasting causative relationship that
links credit risks with key risk drivers.
Model theoretical validation also involves imposing a parsimonious structure and designing the model so that it is
intuitive and simple to understand. A parsimonious structure allows the model to focus on the most important factors
while leaving out “accidental” variables that are either redundant or have a one-time transient effect only. Meanwhile, an
intuitive model is more traceable and more likely to be used appropriately by regular business users with limited
statistical backgrounds. Moreover, a model should be theoretically validated by examining the directional response of
PD/LGD to explanatory variables. For example, one would expect PD to increase with the decline of DSCR, and this
type of directional relationship can be first validated through a theoretical angle. Finally, common sense and business
experience also apply when checking the theoretical reasonableness of model results. For example, since we will not be
able to predict ex ante which loans will default and which will not default, a model is probably poorly constructed if it
consistently outputs PDs in the 95% to 100% range for an input DSCR of 1.0.
Beyond theoretical considerations, a model must be empirically validated, most importantly, through the out-of-sample
data. Moody’s Analytics has pioneered and refined the use of empirical validation in commercial credit models, and we
validate the CMM model using those proven testing processes.
It is useful to consider the model’s empirical validation from two separate yet related metrics.
• Model power—the ability of the model to rank-order individual loans from more to less risk. Power describes how
well a model discriminates between defaulting (“bad”) and non-defaulting (“good”) loans.
• Model calibration—the consistency of the aggregate-level EDF and LGD measures when compared to the actual
realization for a portfolio of loans. It also measures relative risk magnitudes between subsets of a portfolio. For
example, whether or not a model can predict correctly if group A has twice the default rates as group B.
These two dimensions are indispensible for a good model, as model power ensures its ability to rank score individual
risks, and model calibration ensures the EDF level closely matches the actual default rates throughout portfolios and
through time.
For model power, we apply walk-forward, k-fold, and cross-validation techniques. These tests include out-of-sample
testing (using defaults and non-defaults that were not used in the model development, such as “hold-out” sample) and
comparisons to alternative model specifications.

5.1 Walk-forward Tests

The walk-forward testing approach allows users to test models and modeling methodologies while controlling for sample
5
and time dependence . It proceeds as follows. We estimate the model up to a certain year and score the observations in
the next year. These model scores are out-of-time and out-of-sample. We then re-estimate the model including one more
year of data and repeat the analysis for the next year and continue until the end of the sample. For each version of the
model, we calculate the accuracy ratio of next year’s scores by comparing them to actual realized default rates. Under this
approach, we select the development dataset prior to a certain cutoff date, and the validation dataset begins after that
date. Multiple tests can be conducted as one moves the cutoff dates backward and forward. This technique reduces the
chances of “over-fitted” models, since the testing process never uses data used to fit model parameters. At the same time,
the approach allows modelers to take greater advantage of the data by using as much of the information as possible to fit

5
See Dwyer and Kocagil (2005).

20
 and to test the models. The results from our out-of-sample tests show a high degree of discriminatory power of CMM
model, shown in Table 5.

Table 5 Out-of-sample model power from the walk-forward method

Development Data Out-of-sample Year ROC Ratio 6 Accuracy Ratio

Up to 2006 2007 77.6% 55.2%
Up to 2007 2008 80.8% 61.6%
Up to 2008 2009 78.8% 57.6%

A model is considered to be better than a random guess if the ROC ratio is above 50% and the accuracy ratio is above
0%. A perfect model achieves the maximum possible 100% for both ROC and accuracy ratios.

CMM Model

The more distance, the

better the model’s
Predictive Power

Random Guess

Figure 7 A typical CMM model validation power curve

5.2 k-fold Tests

Similar results on the model’s power can be obtained using alternative validation techniques such as k-fold and cross-
validation analysis. The k-fold analysis tests model stability vis-à-vis different data segments. In this analysis, we divide
commercial mortgages into k sub-sample (we typically set k = 5). We then estimate the model on the sample while
excluding the observations in the set {k=1}. This model is used to score the observations in the set {k=1}. Such scores
represent true out-of-sample estimates. We repeat this process for each of the k sub-samples. Afterward, we combine the
out-of-sample scores into one data set and calculate the accuracy ratios and the power curve. We then compare these
results with the corresponding in-sample accuracy ratios and power curve. In addition, we check to see whether the
parameter estimates for each explanatory variable are stable across the different samples. We find that model performance
is well maintained both in- and out-of-sample in the k-fold analysis.
As shown in Table 6, our out-of-sample analysis also confirms that the model’s discriminatory powers remain consistent
across different property types.

6
The ROC ratio is also commonly called the area under the ROC curve (AUROC), where ROC refers to receiver operating
characteristics.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 21

 Table 6 Out-of-sample power curves by property types
Property Type ROC Ratio Accuracy Ratio
Office 80.8% 61.6%
Retail 81.7% 63.4%
Industrial 79.7% 59.4%
Multifamily 81.8% 63.6%
Hotel 86.8% 73.6%

5.3 Validating Model Calibration

In terms of validating model calibration, we conduct a thorough out-of-sample exercise. That is, we first develop a
version of the model using the latest default data from the 2000s era and then apply the model to a pseudo-historical
commercial mortgage portfolio of the life insurers to compare the modeled PD/LGD to actual realized default rates and
loss severities. Since calibration is about the correct overall aggregate levels of the risk measures, we focus on the 10-year
cumulative PD values for vintages 1985 through 1995, which we were able to track their credit performance through
7
2002 . The out-of-sample model results match closely with realized defaults, as shown in Table 7 and Figure 8.

Table 7 Out-of-sample comparison of PD levels with actual

Statistical Measure CMM Model PD Actual Default Rate
Average 13.0% 13.1%
Standard Deviation 10.4% 9.7%
Minimum 1.9% 0.7%
Maximum 31.0% 27.3%

Cumulative
Default Rate
Model PD Actual Default Rate
35%

30%

25%

20%

15%

10%

5%

0%
85 86 87 88 89 90 91 92 93 94 95
Origination Cohort

Figure 8 Comparison of the cumulative 10-year model PD and actual, realized default rates

Additional calibration exercises simulate a representative commercial mortgage industry portfolio throughout its history
and compare model PD and EL to actual, realized default rates and charge-off rates (proxy for credit losses) using the

7
See Snyderman (1991) and Esaki and Goldman (2005).

22
 aggregate time-series statistics of the CRE loan portfolios held by FDIC banks. The result also confirms that calibration
of model PD, LGD, and EL is largely consistent with banks’ historical experience.
Because model calibration also involves relative risk measurements between subsets of a portfolio, we conduct another
validation exercise to compare out-of-sample PD level accuracy by buckets. In this exercise, we bucket out-of-sample
observations into 30 buckets based on our estimated PD and then compare the average PDs with actual realized default
rates for each group, shown in Figure 9. The overall PD levels appear to fit quite well across rating groups based on
estimated PD.

Figure 9 Fitting between PD Buckets and actual realized default rates

Using a variety of validation techniques, we find that the model estimates credit risks reasonably well, conditional upon
accurate inputs. We believe that the model produces satisfactory risk measurements over the past several commercial real
estate cycles given the empirical evidence we have access to; we also believe that the model is applicable to all the major
segments of the market: insurance companies, CMBS, and commercial banks.
The CMM application is flexible enough to allow the user to conduct historical analyses. As a result, users can perform
back-testing using an institution’s historical portfolios to compare model results to see if they match realized default and
loss rates when using the validation metrics from the power, as well as the calibration perspectives. Before users
implement the CMM application within their organizations, we encourage them to thoroughly understand and evaluate
the model to ensure appropriate use.

6 Summary
The commercial real estate and commercial mortgage markets continue to evolve. We have seen numerous changes on
many important fronts: demand for commercial real estate space driven by macroeconomic forces, local supplies driven
by each MSA’s physical layout and zonings, and capital sources increasingly fluid and global in nature. Furthermore, the
recent credit crisis has heightened regulatory scrutiny of commercial mortgage lending.
Despite all the changing factors, we find that a commercial property’s financials, transpired through DSCR and LTV,
continue to play a dominant role in affecting mortgage defaults and losses. Other market-wide factors, such as the
prevailing commercial real estate market condition and underwriting pressures, continue to influence the magnitude of
point-in-time default rates through their impact on default option values and borrower behaviors. This little-changed
relationship between a loan’s financial performance and its credit risk makes a quantitative model not only possible, but
extremely useful for any serious risk management practice.
After extensive research on a very rich collection of both public and private data sets, we have developed unprecedented
insights into the drivers of credit risks of commercial mortgages. The result is our CMM modeling framework, which
balances the need for a highly predictive model with a robust, intuitive, transparent, and highly flexible model.
Based on all the evidence seen to date, we firmly believe that our CMM framework will prove indispensible for financial
institutions when implementing quantitative credit risk tools in loan origination, portfolio management, scenario
analysis, stress testing, and meeting regulatory requirements.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 23

24
 References

Acknowledgements References
The authors thank Kevin (Qibin) Cai for
substantive research assistance. Black, Fischer, and Myron Scholes, 1973, “The Pricing of Options and
Corporate Liabilities.” Journal of Political Economy, Vol 81, No 3, 637–
Copyright © 2011 Moody's Analytics, Inc.
and/or its licensors and affiliates. All rights 659.
reserved.
Chen, Jun and Yongheng Deng, 2004. “Commercial Mortgage Workout
Strategy and Conditional Default Probability: Evidence from Special
Serviced CMBS Loans.” Real Estate Research Institute funded paper.
Crosbie, Peter and Jeff Bohn, 2003. “Modeling Default Risk.” Moody’s
KMV White Paper.
Dwyer, Douglas W. and Ahmet E. Kocagil, 2005. “Moody’s KMVTM
RiskCalcTM V3.1, United States.” Moody’s KMV White Paper.
Esaki, Howard and Masumi Goldman, 2005. “Commercial Mortgage
Defaults: 30 Years of History.” CMBS World, 21-29, Winter 2005.
Merton, Robert C., 1973, “Theory of Rational Optional Pricing.” Bell
Journal of Economics, Vol 4, No 1, 141–183.
Patel, Nihil and Jing Zhang, 2009, “Modeling CRE Correlations in Credit
Portfolios,” Moody’s KMV white paper, 2009.
Snyderman, Mark P., 1991. “Commercial Mortgages: Default Occurrence
and Estimated Yield Impact.” Journal of Portfolio Management, Fall
1991.
Vandell, Kerry D., 1992. “Predicting Commercial Mortgage Foreclosure
Experience.” AREUEA Journal, Vol 20, No 1, 55-88.
Wheaton, William, Ray Torto, Jon Southard and Robert Hopkins, 2001.
“Evaluating Real Estate Risk: Debt Applications.” Real Estate Finance,
Vol 18 No 3, 29-41.

MODELING COMMERCIAL REAL ESTATE LOAN CREDIT RISK: AN OVERVIEW 25