ECONOMIC & CONSUMER CREDIT ANALY TICS

August 5, 2010

Methodology for Forecasting and Stress-Testing ABS and RMBS Deals

Prepared By
Juan Carlos Calcagno Senior Economist juancarlos.calcagno@moodys.com Anthony Hughes Senior Director tony.hughes@moodys.com Ioannis Stamatopoulos Assistant Director ioannis.stamatopoulos@moodys.com

Methodology for Forecasting and Stress-Testing ABS and RMBS Deals

he financial crisis has demonstrated the urgency of developing new tools to improve financial decision-making. The need for research and development is especially great in the structured securities market, which was at the epicenter of the crisis. Structured finance originations have now virtually dried up, to the detriment of not only investors, but also consumers. If investors can accurately value and assess the risks of their holdings, bonds can be priced correctly, and the benefits of securitization can then flow to them as well as to consumers in the form of lower interest rates on mortgages, auto loans, student loans and credit cards. In general, structured finance deals are put together by pooling groups of consumer loans and placing the principal and interest paid by borrowers into a kitty to be distributed to investors who own the rights to different tranches representing different degrees of risk. A deal is often composed of numerous pools of loans, each with distinct characteristics. Complex rules govern how the money in the kitty is distributed to investors. Some investors may be more exposed to a particular pool within a deal than others who hold bonds carrying the same rating. Conversely, the pools may all pay into a common kitty, in which case the division into pools serves a purely administrative function. The value of the bonds is thus determined by the extent to which the loan pools are able to contribute cash to the kitty. The accuracy of forecasts of these cash streams should therefore be the objective of any model used to assess bond value. In this paper, we outline the approach Moodys Analytics uses to model, forecast and stress-test collateral backing ABS deals, concentrating on U.S. RMBS. Since the Moodys Analytics Structured Finance Workstation already precisely handles the waterfall modeling problem, our task here is to develop a methodology that accurately projects the cash flow from the collateral backing the deals. Our approach considers economic conditions at loan origination, loan characteristics, past pool performance, and dynamics in the macroeconomic environment over time to explain changes in pool performance. We will describe, in detail, our methodology, and the models developed will be validated against simple and similar alternative forecasting approaches. The same methodology has been successfully applied to other types of asset-backed securities such as international RMBS, auto ABS, student loans, credit cards, and business leases. In all cases, the methodology is immediately and globally available, allowing any type of security, provided data are available, to be accurately valued and stress-tested. The output from the models described in this paper provides all necessary data to run waterfall valuation engines and thus compute fair value and expected loss under baseline as well as stressed conditions.

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

The Moodys Analytics Approach

We feel that a pool-level approach is likely to provide greater utility to end users than competing methodologies. Data at this level are available on securities from all parts of the world and on any product, meaning that the pool-level approach is universally applicable. Pool-level models are also more likely to provide accurate forecasts of pool-level aggregates than are competing approaches. The data used for this study are drawn primarily from two sources: The Moodys Analytics U.S. macroeconomic forecast databases and the Moodys Analytics Performance Data Services product. Moodys Analytics maintains one of the largest repositories of macroeconomic, demographic and financial data from a multitude of government and private sources. The data set covers the national accounts, banking and finance, demographics, personal income, prices, retail sales, labor markets, energy, financial markets and many other indicators. Moodys Analytics each month produces projections, both baseline and stressed, for economies and financial markets worldwide using large-scale structural macroeconometric models. In the broadest terms, the models are specified to reflect the interaction between aggregate demand and aggregate supply. In the short run, fluctuations in economic activity are primarily determined by shifts in aggregate demand, including personal consumption, gross private investment, net exports, and government expenditures. The level of resources and technology available for production is, in the short term, taken as givens. Prices and wages then adjust slowly to equate aggregate demand and supply and thus move the economy toward equilibrium. In the longer term, changes in aggregate supply determine the growth potential of the economy. The rate of expansion of the resource and technology base is the principal determinant of overall economic growth, which feeds, interactively, into other factors in the model. The model and subsequent forecast are overseen by a team of about 60 economists who cover the performance of the economy in real time.

The forecasts undergo continual revision and adjustment to reflect new trends and changing data. Recent advances in macroeconomic theory, new econometric techniques, and increased computing power also govern development of the models. The system has been changing to accommodate increased demand by clients who want to use the models to generate alternative scenarios and to understand the sensitivity of the macroeconomy to changing economic and financial conditions. Moodys Analytics produces one upside and several downside alternative scenarios each month for each economy it covers. A detailed description of some recent alternative scenarios is included in Appendix 2. In terms of the ABS credit data, meanwhile, PDS is a comprehensive and standardized data set that includes raw information on all active and inactive ABS and RMBS deals rated by Moodys Investors Service in the U.S. and globally. For a more comprehensive picture of the U.S. RMBS market during the model development stage, and to avoid any selection bias, we have supplemented the PDS data set with information from deals that are not rated by MIS; this supplemental database is available only internally. Given that MIS-rated deals cover about 90% of the U.S. RMBS universe, the PDS product alone is close to comprehensive, given the aims of this paper. The data set is composed of deal information at origination, augmented by continuously updated performance information from servicer and trustee monitoring reports. The data files are cleaned and validated by the MIS monitoring team as part of the ongoing ratings surveillance process. The data are of excellent quality and of a perfect structure for the forecasting models considered in this article. Included in the files are over 100 performance statistics at the deal, pool, and tranche level for over 10,000 ABS and RMBS deals in the U.S., Europe, the Middle East and Africa as well as in the AsiaPacific region. PDS coverage in each region matches the market share of MIS in rating the underlying deals. The data are not only consistent within each region or country, they are also consistent across and between regions, allowing for international analysis

if so desired. Our analysis typically concentrates on key pool performance metrics such as delinquency rates as a proportion of original balance, longer-term indicators of default, prepayment rates, and severity of losses or loss given default. For the purposes of this study, because of the lack of data on a few key variables, we restrict our attention to U.S. RMBS deals originated after December 2001. Many performance measures, however, are available back to 1995. Table 1 provides basic descriptive statistics about the pools in the sample used for modeling purposes. The pools employed are concentrated in subprime, alt-A and jumbo categories. Given that originations grew so quickly from 2004 to 2007, unsurprisingly, these cohorts are also very well represented. Data for other vintages and product types are sufficiently available for their determining factors to be accurately estimated. Once all data files are merged, the final data set used for the analysis is a multidimensional, longitudinal panel data set. Our estimation sample is composed of 12,148 unique asset pools observed over, in most cases, several years. In total, our estimation sample for most dependent variables amounts to about 660,000 unique observations with a few categories, most notably loss severities, having access to only about 140,000 separate observations. The data cover a full business cycle, including two recessions, allowing us to correctly measure the impact of cyclical factors and to weigh these factors against internal pool-specific information.

A. The econometric model

The panel nature of the data on structured securities provides a rich tapestry with which to construct high-quality forecasting models. In this paper, we consider a model of the form (1) or, decomposing components: into its constituent

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

where indicates a series of pools contained in the data set of interest, is a time series defining the cohort, or vintage, in which the pool was originated, is a standard time series indicator of the period in which the pools performance is observed and is the dependent variable of interest, appropriately transformed, be it a delinquency rate, a default rate, a prepayment rate, a volume or a rate of loss severity. The vector is an unknown random error term and defines the unobserved, time invariant, pool specific heterogeneity in the data. The vectors and contain independent variables thought to explain the behavior of . For instance, would include factors that specifically relate to the given pool in question but that do not change over time. This set of variables would thus include origination characteristics such as the average FICO score or origination LTV. Conversely, one could simply include a full set of fixed-effects dummies in this matrix, though doing so would preclude the identification of pool-specific factors and thus rule applying the model to newly originated pools. This action would also breach the principle of parsimony, which is paramount given the application of the models to forecasting. The vector , meanwhile, contains factors that pertain to a particular vintage but are common to all pools that are members of the vintage and that do not change over time; the main factors that would be represented in describe economic conditions that existed when the pools in question were being originated. The inclusion of these dynamic origination conditions variables is crucial for the problem of forecasting the way newly minted loans or loans that will soon be originated are likely to perform. Note that the inclusion of a full set of pool-level fixed effects would also preclude estimation of this component, reducing the utility derived from the model. The vector contains factors that change over time but which are common to all pools. This is therefore the component that explains the way the external macroeconomic environment affects the performance

of all legacy pools. This vector may include lags of macro variables and variables such as structural break dummies that indicate past policy changes. Macroeconomic variables, therefore, enter the model twice, describing both conditions at origination as well as ongoing conditions as the pool matures. Because origination standards tend to decline during booms and rise during recessions, the two components will generally run counter to each other, with opposite signs on variables that are common to the two components. The nonlinear, baseline lifecycle is contained in since the age of the pool is a simple function of time and vintage. This lifecycle component arises because delinquencies, defaults and prepayments tend to rise from low levels in the early months of the pools existence before peaking and then gradually declining as the pool matures. This inherently nonlinear behavior can be parsimoniously modeled using cubic spline functions with a small number of knots to capture shape changes across the pools lifecycle. We employ the Stata routine rc_spline to this end, which fits cubic splines between each knot and linear functions before the first knot and after the last.1 The last component, , includes any factors that vary across pool, vintage and time. This component could include updates of pool-specific information, such as refreshed FICOs, but, unfortunately, these factors are unavailable for the current exercise. The vector may also include interaction terms between any or all the individual components described above. For instance, it is reasonable to believe that poorer quality pools may be more sensitive to changes in the macroeconomic environment than pools of better stock. An interaction between FICO, say, or LTV and home prices should therefore be considered for inclusion. Conversely, the shape of the lifecycle may be influenced by pool quality or even by the timing of events in the broader macro environment. For example, a recession occurring close to the peak in the lifecycle may have a more pronounced
1 We also used Stata optimization routines to define the number and location of knots (Royston & Sauerbrei, 2007).

effect on pool performance than a recession that hits a more mature pool. A parsimonious forecasting model may leave many, or all, these interaction terms on the cuttingroom floor, but a serious modeler should at least consider their inclusion in the preferred forecasting specification. The final members of the component are forecasts and actuals of other poolspecific performance metrics that enter into the determination of the dependent variable of interest. For example, if we are modeling a 60-day delinquency rate, it makes sense for the 30-day rate, observed with a one-month lag, to be a key leading indicator. This is akin to the inclusion of a roll rate or migration term in the estimated model. Models of pathologies that occur later in the default process, like REO or foreclosure rates, may conceivably include lagged 30-day, 60-day and 90-day delinquencies or a myriad of other indicators from earlier in the process. Prepayment rates may weigh on default rates, as advocates of competing risk models often suggest they should. In considering the inclusion of these terms, one must weigh the threat of promulgating forecast errors across different models. Roll rate terms should be used sparingly, though they should certainly be considered in the model specification process. We refer to these roll rate terms as pipeline connections. Taken together, it is clear that the multidimensional, vintage panel data framework provides us with a huge array of factors that may influence cash flow.

B. Estimation
In this section, we turn to estimation issues, followed by some practical matters of model specification for the problem at hand. This section will focus on functional form selection as well as independent variable and lag length determination. In modeling a large panel data set such as this one, the standard approach would be to set out using a fixedeffects estimator and merely produce the desired forecasts.2 One could conceivably test the restrictions implied by a random-effects
2 For general descriptions of panel data models, see Baltagi (2001), Hsiao (2002), or Wooldridge (2002)

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

estimator, but with more than half a million observations, this model will almost certainly be rejected. As mentioned earlier, we also reject the fixed-effects approach because its use likely affects forecasting performance3 while reducing the utility of the model in a number of important ways. Several alternatives to the fixed-effects approach are available via the econometrics literature. The most simple option would be to assume that all pool-level heterogeneity can be explained by available economic data and pool-specific features and leave out of the model altogether; this is known in the literature as the pooled OLS estimator. The ability to identify unobserved individual heterogeneity is, however, one of the primary advantages of panel data, and this simple approach would be unable to reap these rewards. The second alternative involves the standard randomeffects estimator. Using this approach, we could identify the effects of observed time invariant factors, but this estimator is built on a few strong simplifying assumptions that are unnecessary given the plethora of available data at our disposal. Most notably, the random-effects estimator assumes that unobserved individual effects are uncorrelated with other included factors.4 The third option the Hausman-Taylor (1981) estimator was built to overcome the problems inherent in both the random-effects and fixed-effects estimators. For these reasons, we prefer the Hausman-Taylor estimator. The HT estimator allows for unobserved individual heterogeneity and the estimation of time invariant factors while avoiding the assumption of no correlation between the various terms included in the model. If we rewrite (1) as (2) where and represent variables that are uncorrelated with the error term and
3 Though assertion is consistent with the principles of parsimony and has been verified by, for example, Baltagi (2008). 4 Using formal statistical validation like the Hausman test (Hausman, 1978), we rejected the null hypothesis of no correlation. As a result, we do not present results using randomeffect model here.

that are time varying and time invariant, respectively. The variables and represent equivalent sets of factors, albeit those that are correlated with the unobserved innovations. In this model, the unobserved individual heterogeneity, , is assumed to be random, independent and identically distributed with a zero mean and finite variance. The usual assumptions are made regarding . Note that the impact of the time invariant factors, , cannot be directly separated from the impact of the unobserved, time invariant random effects, . HT estimation then proceeds using an instrumental variables approach. The variables in , which are uncorrelated with and by assumption, represent the instruments for the endogenous time invariant factors contained in . In this analysis, because we seek to retain the ability of the models to project the performance of newly originated mortgage pools, the most interesting factors for consideration in are those describing poollevel characteristics such as average LTV or average FICO. As instruments for these factors, we employ various macroeconomic series, observed at origination, that we assume to be uncorrelated with the errors but that can help explain shifts in underwriting standards indicated by variables such as average FICO scores. As mentioned earlier, lending standards tend to fall during a boom, meaning that average FICOs should, for example, rise as the economy sinks into recession. We concede that there may be some residual correlation between macro conditions at origination and but weigh this against the utility derived from estimates of the pool-specific factors.

or REO), and principal repayment rates. There may be considerable interest in these other factors in their own right, but we use these factors chiefly to serve the broader issue of accurate cash flow projection. In terms of functional form, since all dependent variables of interest are bound by zero and unity, we apply a logistic transformation of the form

C. Practical considerations
Our ultimate aim in the model building process is to accurately forecast the constant default rate (CDR), constant prepayment rate (CPR), and rate of loss severity the three inputs to the waterfall engine. To support the forecasts for these factors, we also construct models for 30-day, 60-day and 90+-day delinquencies, bankruptcies, foreclosures, the rate at which defaulted properties are owned by the servicer (known as real estate owned

In many cases, we could have effectively applied a simple log transform to the data, which asymptotes only at zero but not one. In some cases, however, particularly in short-term delinquency rates on subprime mortgages, the data start to reach into the higher part of the available number space, where the upper bound starts to come into play. Further, for loss severities in home equity lines and some mortgages, we find most of the action close to the uppermost bound. In all cases, the logistic transform is found to perform well in terms of forecast accuracy, allowing a simple, theoretically tenable and consistent transformation to be applied. A key aspect of model development is variable selection identifying which credit and economic variables best explain the dynamic behavior of the dependent variable in question. For panel data structures such as this, one must also decide on the number and nature of the lags for each variable in the equation. Aligned with principles of modern econometrics, we prefer to choose the variables based on a combination of economic theory or intuition, together with a consideration of the statistical properties of the estimated model. We feel that models built using pure data-mining techniques or principles such as machine learning, though they may fit the existing data well, are more likely to fail in a changing external environment because they lack theoretical underpinnings. A strong signal can be derived from a consideration not only of how mortgage markets are seen to work but of how they should work given our understanding of markets and human folly. The best prediction models employ a combination of statistical rigor with a
4

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

TABLE 1

Number of pools by collateral type and closing year

Deal Closing Year Pools Collateral Type HELOC High LTV Home Equity/Closed End 2nds Subprime Alt-A FHA-VA Jumbo Prime Conforming Scratch & Dent Subprime 2nds Option Arms Total Pools Average Time-Series Available 2002 14 14 8 287 155 9 207 10 2 706 92 2003 23 10 8 526 463 17 583 25 13 5 1,673 82 2004 61 6 6 749 1,080 17 534 5 63 27 71 2,619 70 2005 50 7 6 795 1,335 25 452 43 51 221 2,985 58 2006 46 6 19 777 904 20 315 49 83 284 2,503 46 2007 20 1 14 386 676 2 252 41 41 173 1,606 36 2008 2 4 12 30 3 51 23 2009 2 1 3 10 2010 Total 1 1 2 2 216 44 61 3,524 4,626 90 2,376 6 231 217 757 12,148

healthy dose of economic principle. Models built this way enjoy the additional benefit of ease of interpretation. The selection of the explanatory macroeconomic variables was also based on our ability to forecast them easily and sufficiently accurately. In econometric forecasting, the future realizations of the dependent variable rely on accurate forecasts for the independent variables. Forecast errors can easily be multiplied if input variable forecasts are not closely monitored or if the data backing the forecasts are noisy and thus inherently difficult to project. Identifying macro variables that are useful in this context is straightforward: They are the factors that often have the power to change the direction of financial markets and that are thus routinely debated in the public square. Variables such as unemployment rates, home price indices, key interest rates, and retail sales numbers are generally better candidates than industry-level mortgage foreclosure rates, for instance, which are of interest only during certain brief periods. We find that a combination of such core variables generally act as a reasonable proxy for these more peripheral factors. Since core variables are generally easier to predict, we find them more useful in building optimal forecasts of ABS pool performance.

Model Results
A summary of the variables included in each regression for the vectors we modeled is presented in Table 2. The first group of variables, the pool origination factors ( in (2)), include average LTV and FICO, a set of dummies for vintage origination year, dummies defining loan or collateral type (i.e. whether the pool represents subprime or jumbo mortgages, for instance) and a set of dummies indicating whether the originator of the pool in question was in the top 10 in terms of total number of pools originated during the in-sample period. The vintage year, collateral type and originator dummies are included in most regressions in order to control for group heterogeneity in the data. Weighted average LTV and FICO at origination are key pieces of information not only because they determine pool quality and hence performance, but also because we seek to retain the ability of the models to project the performance of newly originated mortgage pools using information that is readily available at deal origination. The second group of variables includes the economic condition factors at deal origination, which is our set of instrumental variables, . In this category, we found that gen-

eral measures of economic activity, like GDP and the unemployment rate, are correlated with changes in pool performance; we also find good support for factors specific to the mortgage market such as home prices and interest rates. Broadly speaking, house price changes and an average benchmark interest rate at origination tend to have a countercyclical impact on measures of performance observed early in the default process. Labor market conditions at deal origination, however, become more important when modeling later stages in the default process, including foreclosure, REO, and losses. The third set of variables is the macroeconomic series or those factors related to the pools exposure to the business cycle. We found that the types of macroeconomic factors that affect delinquency are standard fare labor and housing market variables, income, and overall economic activity with close attention paid to lag structures and transformations applied to the economic data being used. House price changes and refinancings become much more important drivers at later stages in the foreclosure process. We use both current home price changes and price changes since origination as key housing market factors. This second variable allows for a differential negative equity ef5

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

fect to be explicitly modeled. The models of foreclosure, REO, charge-offs, and severity are tied largely to home prices and various other factors affecting housing and legal cost structures. We also found prepayment rates to be more sensitive to interest rates, refinancing activity, and existing-home sales. As the structure of the model is recursive in nature, the next set of covariates constitutes the so-called pipeline connections. We model delinquencies largely independently of one another but model subsequent events prepayments, foreclosures, charge-offs, and severity conditionally on the delinquency performance the pools attain. We also take advantage of our panel data estimation techniques to measure the unobserved pool heterogeneity embedded within several early performance measures and then use these estimates as an exogenous quality metric in the subsequent net charge-off, prepayment, and principal payment models. Finally, the models include a cubic spline baseline lifecycle with four knots5 and interaction terms between the spline variables, current economic conditions, and collateral type dummies. By introducing these interactions, we measure the differential impact of economic drivers on different types of assets. For example, we allow the sensitivity to unemployment rate changes to vary between pools classified as subprime and those defined as jumbo deals. Similarly, by interacting the spline with collateral type, we allow the shape
5 Three in the case of the LGD model.

of the lifecycle to depend upon the underlying quality of the pools under consideration. Table 2 summarizes the specifications used to model each of the 11 vectors studied in this article. All the variables (or groups of variables) are significant at the 5% level, and all variables have appropriate signs consistent with economic theory. The models have been tested rigorously to ensure that they are optimal, among the set considered, for forecasting purposes. In Table 3 we present the results for one regression, that for the foreclosure rate, as an example of the estimation results achieved for all models. Other specifications are, naturally, available from the authors on request. Given that our general model specification included factors from all the subcomponents described above, we can reasonably conclude that factors specific and internal to the pool, together with external macroeconomic factors, are given appropriate loadings in the calculation of performance forecasts. Once we feed in forecasts of macroeconomic variables derived from our structural macroeconomic model, generating pool-level forecasts is straightforward. Extending this principle to stress tests which really constitute little more than pessimistic economic forecasts is equally simple. This process delivers a set of vectors under each alternative macro scenario, including a baseline projection. The models can be applied to custom scenarios provided all necessary macroeconomic variables are covered by the scenario. The particular shape of these fore-

casts and scenarios depends on the estimated elasticity of the risk vector to the included macroeconomic series. As mentioned earlier, the elasticities have been found, in many cases, to be heterogeneous across different collateral groups. Moreover, we find that the slopes vary depending on the underlying quality of the pool in question and the time at which the pool was originated. To illustrate these features, we present forecasts and scenarios for a small number of pools drawn from the subprime, alt-A, jumbo, and option ARM collateral types under two alternative macroeconomic scenarios the current baseline forecast (a recovery) and an alternative, depression-like event. These projections are depicted in Charts 1 and 2. Under baseline assumptions, we find that foreclosure rates will stay flat for another quarter as the real estate market stabilizes and then steadily decrease as bad accounts move out of the foreclosure state and are flushed from the system; the vector then converges to the steadystate baseline lifecycle. A depression scenario, however, generates a sharp double dip, with figures across collateral types reaching historically high levels in 2011, similar to the peaks observed in mid-2009.

A. Out-of-sample validation
One of the most important findings in the forecasting literature is that the model that best fits the data is not necessarily the one that will provide the most accurate out-ofsample forecast (Fildes and Makridakis (1995)). Typically, to assess accuracy, the data will be

Chart 1: Projections for Alt-A and Subprime Pools

Foreclosure rates, %
35 30 25 20 15 10 5 0 07 08 09 10 11 12 13 14 15
Source: Moodys Analytics
FROM MOODYS ECONOMY.COM 1

Chart 2: Projections for Jumbo and Option Arm pools

Foreclosure rates, %
25

Baseline Subprime Alt. Scenario

20 Option ARM 15 10

Baseline Alt. Scenario

Alt-A

5 0 07 08 09 10 11 12 13

Jumbo 14 15

Source: Moodys Analytics

FROM MOODYS ECONOMY.COM 2

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

split into two data sets: The first set, a development sample, is used to specify the model and estimate its coefficients; the second set is used to evaluate forecast accuracy and is known as the hold-out sample. To validate our preferred models, we hold out the last six available data points at the end of the time series for each pool. The models were fitted to the remaining historical data. The forecasts generated using the models were then compared with actual values observed during the hold-out sample, and aggregate root mean squared forecast errors were computed. Underlying our forecast accuracy evaluation was the need to test our preferred model against reasonable alternatives (Baltagi (2008)). For this purpose, the out-of-sample performance of the HT estimator was compared against that of three benchmark models. Our first benchmark was a nave forecast that assumed no change from the most recently observed value in the development sample. The other two benchmarks were specified similarly to our preferred specification but were estimated using pooled OLS and the fixed-effects estimator, respectively. These approaches are described in previous sections; both should provide far stiffer competition to the preferred specification than the corresponding nave forecast.

Errors for each approach were aggregated and results presented in Table 4. We present the performance of the alternative models relative to the preferred HT specification. Numbers greater than unity therefore indicate cases in which the alternative approach is able to beat the preferred method in a head-to-head competition. The results of the validation exercise are clear-cut. The preferred HT estimated panel data model suffers less squared forecast error than each of the three considered alternative specifications for all 11 estimated vectors. The one exception is for charge-offs, in which the fixed-effects estimator is able to eke out a small victory by just a couple of percentage points. We feel that any test applied enough times will yield contrary results and place this finding in that category. Folding in the fact that the fixed-effects estimator reduces model utility, we can easily conclude that a 2% improvement is not enough to warrant a change in technique. In terms of the three primary inputs to the Structured Finance Workstation CDR, CPR and severity the HT estimated model easily dominates its competitors and can thus be recommended. Vectors derived from this approach are likely to be optimal in terms of deriving accurate bond valuations.

Final Remarks
These results demonstrate that pool-level performance aggregates are well forecast using aggregate panel data specifications that accurately capture the impact of macroeconomic dynamics on the behavior of loans. The models described are simple, parsimonious structures in which every coefficient estimate can be readily understood using simple economic intuition. The subsequent models forecast the aggregate performance characteristics accurately and easily lend themselves to the construction of relevant stress-testing scenarios. Structured finance is a complex business. After the recent recession exposed the failure of the models used in the industry, many sought to solve the problem with even more complex solutions and models. Though it may seem counterintuitive, the correct response to the failure of complex models in 2007-2009 was a flight to simplicity and functionality. The models that form the backbone of the Moodys Analytics system are simple yet comprehensive and, importantly, are designed specifically to tackle the problem of forecasting and stress-testing the collateral underlying ABS deals.

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

TABLE 2

Summary of Models inputs

Vector Group Variable LTV Origination Conditions FICO Top Originators Vintage Year Loan Type Economic Conditions at Origination Fed Funds Home Prices GDP Unemployment Rate Disposable Income Home Prices (HP) Change in HP since 0 Current Economic Conditions Negative Equity Dummy Unemployment Rate Avg. Hourly Earnings GDP Personal Bankruptcies REFI Volume Fed Funds Debt Service Burden Existing Home Sales Mort. REFI Originations 30 day delinquency Pipeline Connections 90+ day delinquency Bankruptcy Foreclosure REO Unobserved effect (CDR) Other Variables Lifecycle * Loan Type Lifecycle * Economic Loan Type*Economic Dec 2005 Bankruptcy Law X X X X X X X X X X t-3 t-3 t-3 t-3 t-1 t X X X X X X X t-12 t-6 t-3 t-6 t-3 t-6 t-3 t-6 t-1 t-6 t-6 t t-6 t-12 t t t t t-12 t t t t t t t, t-3 t, t-6 t, t-3 t, t-6 t-3 t-6 t-3 t, t-6 t t t t X X X X X X X X X X X X X X X X X X X X X X X X 30 day 60 day 90+ day delinquency delinquency delinquency CDR X X X X X X X Bankruptcy X X X Foreclosure X X X REO X X X X X X X X LGD X Net Chargeoff CPR X X X X X X X X X (relative to WAC) X X X X X Principal

Lifecycle

Number of knots

Notes: All models include seasonality factors (month) and dummies for zeros

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

TABLE 3

Partial Regression Results for Foreclosure Rate

Coefficient LTV FICO (non-linear spline 1) FICO (non-linear spline 2) Unemployment Rate, % at origination House Prices, % change year ago lagged 12 months GDP, % change year ago Dummy indicator for Negative Equity 90+ day delinquency lagged 3 months Bankruptcy lagged 3 months Observations Number of unique pools 0.053 -0.006 -0.001 -0.443 -0.028 -0.015 0.108 0.110 0.081 576651 11727 Standard Error 0.004 0.001 0.001 0.014 0.000 0.001 0.004 0.001 0.001 t-statistic 11.809 -8.402 -1.889 -30.842 -119.093 -27.163 29.156 173.511 151.339

Note: model includes lifecycle, seasonality, and dummies for zeros, originators, negative equity and loan type and a constant term

TABLE 4

Out of Sample Forecasting Results

Ratio between Hausman-Taylor and benchmark models RMSE Vector 30 day delinquency 60 day delinquency 90+ day delinquency CDR Bankruptcy Foreclosure REO CPR Principal Severity Chargeoff Nave Forecast 0.935 0.871 0.688 0.830 0.948 0.572 0.973 0.884 0.668 0.911 0.762 Pooled OLS 0.742 0.853 0.479 0.802 0.720 0.636 0.645 0.957 0.586 0.928 0.905 Fixed Effects 0.941 0.904 0.531 0.875 0.711 0.777 0.758 0.993 0.667 0.850 1.020

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

ECONOMIC & CONSUMER CREDIT ANALY TICS

References
1. Armstrong, J. S. (1985). Long-Range Forecasting: From Crystal Ball to Computer. New York, John Wiley & Sons 2. Allen, P. G. and Fildes, R. (2001). Econometric Forecasting. Principles of Forecasting. J. S. Armstrong. Norwell, MA, Kluwer Academic Publishers. 3. Baltagi Badi H. (2001). Econometric Analysis of Panel Data. Wiley and Sons, Chichester, UK. 4. Baltagi, Badi H. (2008). Forecasting with panel data, Journal of Forecasting, 27(2), pp 153-173. 5. Hendry, David F. and Hubrich, Kirstin, (2009). Combining Disaggregate Forecasts or Combining Disaggregate Information to Forecast an Aggregate. ECB Working Paper No. 1155. 6. Fabbozzi, Frank and Kothari, Vinod, (2008). Introduction to Securitization, Wiley. 7. Fildes, R and Makridakis, S (1995), The impact of empirical accuracy studies on time series analysis and forecasting, International Statistical Review, 63, pp. 289-308. 8. Green, Richard and Wachter, Susan M. (2005), The American Mortgage in Historical and International Context. Journal of Economic Perspectives, 19(4), pp. 93-114. 9. Hausman, J.A. (1978). Specification Tests in Econometrics, Econometrica, 46 (6), pp. 12511271. 10. Hausman Jerry A. and Taylor William E. (1981) Panel Data and Unobservable Individual Effects. Econometrica 49, pp.137798. 11. Hsiao (2002). Analysis of Panel Data. Cambridge University Press 12. Kendall, Leon, and Fishman, Michael (2000). A Primer on Securitization. The MIT Press 13. Lutkepohl, H. (2006). Forecasting with VARMA processes, in G. Elliott, C.W.J. Granger & A. Timmermann (eds), Handbook of Economic Forecasting, vol. 1, pp 287-325, Elsevier. 14. Royston, Patrick, and Sauerbrei, Willi (2007). Multivariable Modeling with Cubic Regression Splines: A Principled Approach. The Stata Journal 7(1), pp. 4570 15. Wooldridge Jeffrey M. (2002) Econometric Analysis of Cross Section and Panel Data. MIT Press, Cambridge, MA.

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

10

ECONOMIC & CONSUMER CREDIT ANALY TICS

Appendix A: Definition of Variables

Field Pool Performance 30 day delinquency 60 day delinquency 90+ day delinquency CDR Bankruptcy Foreclosure REO Loss-Given Default (LGD) Net Chargeoff CPR Principal Pool Origination LTV FICO Top Originators Vintage Year Loan Type Macroeconomic series Disposable Income Home Prices Change in HP since 0 Negative Equity Dummy Unemployment Rate Avg. Hourly Earnings GDP Personal Bankruptcies REFI Volume Fed Funds Debt Service Burden Existing Home Sales Mort. REFI Originations Income: Per capital disposable income, (Constant$, SAAR, annual growth rate) Median Sales Price Existing Single-Family Homes, (Ths. $, SA, annual growth rate) Difference in Median Sales Price since Deal Origination (%) 1 if change in HP since 0 is negative; 0 otherwise Household survey: Unemployment rate, (%, SA) Avg. Hrly. Earnings: Total Private, ($ Per Hrs., SA, annual growth rate) NIPA: Gross domestic product, (Bil. 2000 $, SAAR, annual growth rate) Bankruptcies: Personal, Total, (# 3-Month Ending, SAAR, annual growth rate) MBA: Percent REFI Volume, (%, annual growth rate) Interest Rates: Federal Funds Rate, (%,P.A.) Debt Service Burden: Total, (% of Disposable Personal Income) Existing Home Sales: Single-Family, (Mil., SAAR, annual growth rate) Mortgage Originations: Refinances, (Mil. $, SAAR, annual growth rate) Weighted Average Loan To Value at deal origination (%). Weighted FICO score at deal origination (%). Indicator variables for Top Mortgage originators during the sample time period including Countrywide Bank, Wells Fargo, IndyMac, BofA, Citibank, Greenpoint Bank (CapOne), Option One Mortgage, and Residential Funding (GMAC) Vintage origination year (2001 to 2010) Loan collateral type: Heloc, High LTV, Home Equity/Closed End 2nds, Subprime, Alt-A, FHA-VA, Jumbo, Prime Conforming, Scratch and Dent, Subprime 2nds, Option Arm Amount of receivables that are 30-59 days past due divided by the original collateral balance (%). Amount of receivables that are 60-89 days past due divided by the original collateral balance (%). Amount of receivables that are 90 or more days past due divided by the original collateral balance (%). Constant default rate based on amount of receivables in default in the current month, annualized (%) Amount of receivables in which the obligor has declared bankruptcy divided by the current collateral balance (%) Amount of receivables in foreclosure divided by the current collateral balance (%). Amount of receivables that are real estate owned divided by the current collateral balance (%). Monthly net losses during the related monthly period divided by gross losses during the related monthly period (%). Net losses during the related monthly period divided by the prior months ending collateral balance (%). Constant prepayment rate based on unscheduled principal paid by obligors from the current month, annualized (%). Total principal collected during reporting period divided by the current collateral balance (%). Description

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

11

ECONOMIC & CONSUMER CREDIT ANALY TICS

Appendix B: Summary of U.S. Economics Scenarios (July 2010)

Scenario Real GDP Median Home Price Fed Funds Target Unemployment

S1

Stronger Recovery in 2010 Baseline, Current Mild Second Recession Deeper Second Recession Complete Collapse, Depression Aborted Recovery, Below-Trend Long-Term Growth Fiscal Crisis, Dollar Crashes, Inflation

Real growth of 3.9% in 2010, 4.6% in 2011 Real growth of 2.9% in 2010, 3.6% in 2011 Real growth of 2.0% in 2010, 1.3% in 2011 Real growth of 1.6% in 2010, -0.7% in 2011 Real growth of 1.3% in 2010, -1.9% in 2011 Real growth of 2.1% in 2010, 1.8% in 2011 Real growth of 2.2% in 2010, 2.6% in 2011

Expected to rise 2.9% in 2010 and 2.0% in 2011 Peak-to-trough decline of 26%, turnaround in mid 2011 Peak-to-trough decline of 36%, turnaround after mid 2011 Peak-to-trough decline of 40%, turnaround at beginning of 2012 Peak-to-trough decline of 45%, turnaround in late 2012 Peak-to-trough decline of 29%, turnaround in mid 2012 Peak-to-trough decline of 37%, turnaround in mid 2013

The funds rate is expected to end 2010 at 1.6% and 2011 at 2.5% The funds rate is expected to end 2010 at 0.2% and 2011 at 1.8% The funds rate is expected to end 2010 at 0.1% and 2011 at 0.5% The funds rate is expected to remain below 1% until Q3 2012 The funds rate is expected to end 2010 at 0.1% and 2011 at 0.2% The funds rate is expected to end 2010 at 0.1% and 2011 at 0.4% The funds rate is expected to end 2010 at 0.1% and 2011 at 1.9%

Peaks at 10.1% in Q4 2009 and ends 2010 at 8.3% Peaks at 10.0% in Q1 2011 Peaks at 11.8% in Q2 2011 Peaks at 14.0% in Q4 2011 Peaks at 15.1% in mid 2012 Peaks at 11.4% in mid 2011 Peaks at 13.0% in early 2013, ends 2010 at 10.2%

BL

S2

S3

S4

S5

S6

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

12

About Moodys Economy.com

Moodys Economy.com, a division of Moody's Analytics Inc., is a leading independent provider of economic research, analysis and data. As a well-recognized source of proprietary information on national and regional economies, industries, financial markets, and credit risk, we support strategic planning, product and sales forecasting, risk and sensitivity management, and investment research. Our clients include multinational corporations, governments at all levels, central banks, retailers, mutual funds, financial institutions, utilities, residential and commercial real estate firms, insurance companies, and professional investors.
With one of the largest assembled financial, economic and demographic databases, Moody's Economy.com helps companies assess what trends in consumer credit and behavior, mortgage markets, population, income, and property prices will mean for their business. Our web and print periodicals and special publications cover every U.S. state and metropolitan area; countries throughout Europe, Asia and the Americas; and the world's major cities, plus the U.S. housing market and 58 other industries. We also provide up-to-the-minute reporting and analysis on the world's major economies on our real-time Dismal Scientist web site from our offices in the U.S., the United Kingdom, and Australia. Our staff of more than 50 economists, a third of whom hold PhD's, offers wide expertise in regional economics, public finance, credit risk and sensitivity analysis, pricing, and macro and financial forecasting. Moody's Economy.com became part of the Moodys Corporation in 2005. Moody's Economy.com is headquartered in West Chester PA, a suburb of Philadelphia, with offices in London and Sydney. More information is available at www.economy.com.

2010, Moodys Analytics, Inc. and/or its licensors and affiliates (together, Moodys). All rights reserved. ALL INFORMATION CONTAINED HEREIN IS PROTECTED BY COPYRIGHT LAW AND NONE OF SUCH INFORMATION MAY BE COPIED OR OTHERWISE REPRODUCED, REPACKAGED, FURTHER TRANSMITTED, TRANSFERRED, DISSEMINATED, REDISTRIBUTED OR RESOLD, OR STORED FOR SUBSEQUENT USE FOR ANY PURPOSE, IN WHOLE OR IN PART, IN ANY FORM OR MANNER OR BY ANY MEANS WHATSOEVER, BY ANY PERSON WITHOUT MOODYS PRIOR WRITTEN CONSENT. All information contained herein is obtained by Moodys from sources believed by it to be accurate and reliable. Because of the possibility of human and mechanical error as well as other factors, however, all information contained herein is provided AS IS without warranty of any kind. Under no circumstances shall Moodys have any liability to any person or entity for (a) any loss or damage in whole or in part caused by, resulting from, or relating to, any error (negligent or otherwise) or other circumstance or contingency within or outside the control of Moodys or any of its directors, officers, employees or agents in connection with the procurement, collection, compilation, analysis, interpretation, communication, publication or delivery of any such information, or (b) any direct, indirect, special, consequential, compensatory or incidental damages whatsoever (including without limitation, lost profits), even if Moodys is advised in advance of the possibility of such damages, resulting from the use of or inability to use, any such information. The financial reporting, analysis, projections, observations, and other information contained herein are, and must be construed solely as, statements of opinion and not statements of fact or recommendations to purchase, sell, or hold any securities. NO WARRANTY, EXPRESS OR IMPLIED, AS TO THE ACCURACY, TIMELINESS, COMPLETENESS, MERCHANTABILITY OR FITNESS FOR ANY PARTICULAR PURPOSE OF ANY SUCH OPINION OR INFORMATION IS GIVEN OR MADE BY MOODYS IN ANY FORM OR MANNER WHATSOEVER. Each opinion must be weighed solely as one factor in any investment decision made by or on behalf of any user of the information contained herein, and each such user must accordingly make its own study and evaluation prior to investing.

MOODYS ANALYTICS / METHODOLOGY FOR FORECASTING AND STRESS-TESTING ABS AND RMBS DEALS

13

CONTACT US U.S./CANADA +1 866.275.3266 Email us: help@economy.com Or visit us: www.economy.com

Copyright 2010, Moodys Analytics, Inc. All Rights Reserved.

EMEA +44 (0) 20.7772.1646

ASIA/PACIFIC +61 2 9270 8111

OTHER LOCATIONS +1 610.235.5299