453 by qingyunliuliu


									     The Jumbo/Conforming Loan Debate: Pitfalls from the MIRS
    Database, and a Look At Survey Rates and the Securities Market

                                   Thomas A. Lawler
                                  Senior Vice President
                                      Fannie Mae

                                       May 10, 2004

The views expressed here are the authors and do not necessarily reflect those of Fannie Mae


The Monthly Interest Rate Survey (MIRS) from the Federal Housing Finance Board
contains a lot of useful information. For example, the survey shows a high concentration of
loans originated right at the conforming loan limit, and it shows that the share of loans with
balances above the conforming loan limit that have adjustable rates is consistently and
materially higher than the ARM share of loans at or below the conforming loan limit.
These data suggest that there is a significant difference between the pricing of conforming
fixed-rate mortgages and jumbo fixed-rate mortgages in the marketplace.

Unfortunately, The MIRS database is not well suited for use in estimating loan-level
mortgage pricing models. The database has no borrower credit risk data that are widely
used to price mortgage loans, and the wide dispersion of rates charged on loans closed in
any given month cannot be explained by the loan information in this database. Evidence
from other studies, other sources, the capital markets, and surveys on rates from lenders
suggest that the MIRS data may understate materially the jumbo/conforming rate spread
available to prime mortgage borrowers.

The evidence also suggests that the MIRS data, which contain data on rates charged on
loans closed but which provide no information on when loans were priced, are not good
indicators of how mortgage loans were priced in the month the loans were closed. There is
strong evidence that rates charged in any given month are based on current and past
market interest rates, with lags that vary over time depending on the past behavior of
interest rates. As a result, studies comparing MIRS data to other contemporaneous market
data can produce misleading, and in all likelihood biased, estimates of the relationship
between mortgage rates and other market data.

There is evidence, in contrast, that surveys on 30-year fixed rate mortgage loans quoted for
prime borrowers contain useful information on rates available for jumbo loans versus
conforming loans. These jumbo/conforming survey rate spreads move quite closely with
movements in the yield spreads between Fannie Mae MBS and private-label MBS backed
by jumbo loans, based on data provided over time by Wall Street dealers.

The purpose of this paper is twofold. First, it explores the actual distribution of rates
charged on loans closed in the MIRS database, and -- using evidence from other sources –
suggests that this database is not well suited for use in developing loan-level pricing models
for mortgage loans. Second, it explores whether surveys on jumbo versus conforming loan
rates move in a fashion consistent with observed securities prices for Fannie Mae MBS
relative to those for private-label jumbo MBS securities – and concludes that in fact such
survey rate differentials are reasonably well explained by security price movements.

The study has a number of potential policy implications and suggestions for further
research. First, it suggests that the recent study by Passmore (2003) in all likelihood
produces biased or misleading results from a number of perspectives, especially in terms of
the impact of Fannie Mae and Freddie Mac on mortgage interest rates. Second, it suggests
that increased research into the role of secondary market execution on primary market
interest rates is worth pursuing. And finally, it suggests that more research is needed to
understand why the percentage of “jumbo-loan” borrowers who take out fixed-rate loans is
so low.

The Jumbo/Conforming Loan Debate: Pitfalls from the MIRS
Database, and a Look At Survey Rates and the Securities Market

Executive Summary

There have recently been a number of studies that have used data on the differential between the
rate charged on fixed-rate loans at or below the conforming loan limit and the rate charged on
fixed-rate loans above that limit in assessing the impact of Fannie Mae and Freddie Mac on
mortgage rates.1 These studies have come to considerably different conclusions, depending on
whether or not the loan-level data used include information on the creditworthiness of the
borrower. None of these studies, however, explore primary rate quotes for jumbo and
conforming loans, and compare these quotes to secondary market pricing for Fannie Mae and
Freddie Mac MBS versus private-label MBS execution.

This paper explores both the data used in previous studies, as well as how primary
jumbo/conforming rate quotes by lenders to borrowers have moved relative to secondary market
execution in the securities market. The paper comes to the following conclusions:

      •    The Monthly Interest Rate Survey (MIRS) data from the Federal Housing Finance Board,
           while containing useful information, is ill-suited for use in developing loan-level pricing
           models, in part because of the absence of borrower data (such as credit score) used in the
           industry to price mortgage loans, and in part because of the absence of data on when
           loans closed were priced.

      •    The rates charged in the MIRS data are based on current and past mortgage rates,
           depending on the extent to which homeowners “locked-in” rates before closing. As a
           result, the mortgage rates from this database in any given month are not highly correlated
           with market interest rates in that month, and the dispersion of rates charged on loans in
           any given month is not well explained by the variables in the database.

      •    These two findings suggest that studies which (1) use the MIRS database to estimate
           jumbo/conforming spreads; and (2) relate these estimated spreads to movements in other
           market interest rates are fundamentally flawed, and are likely to produce biased and
           seriously misleading results.

      •    The recent paper by Passmore (2003) uses the MIRS data to estimate jumbo/conforming
           spreads, and relates these estimated spreads to other market interest rates in attempting to
           estimate the impact of GSE activity on mortgage rates. As noted in the previous bullet,
           this methodology produces biased and potentially misleading results. As a result, the
           findings in the Passmore paper are not supported by the statistical or econometric results

      •    Another study cited by Passmore that he notes uses a database with much better measures
           of borrower credit quality than the MIRS data comes to a materially different conclusion
           than he does. Specifically, that study (Ambrose, LaCour-Little, and Sanders, 2003)
           suggests a much wider jumbo/conforming spread than Passmore finds using the ill-suited
           MIRS data. Interestingly, the jumbo/conforming spread found by that study is extremely
    E.g., Passmore (2003), Ambrose, LaCour-Little, and Sanders (2003), and McKenzie (2002).

        close to the average primary jumbo/conforming spread from the HSH Associates survey
        of lenders over the period of the study.

    •   Spreads between jumbo rates and conforming rates quoted by lenders and tracked by
        HSH Associates have moved very closely with secondary market spreads between
        private-label jumbo MBS and Fannie Mae MBS over the past six years. In contrast, the
        MIRS data on rates cited above show little correlation with observed secondary mortgage
        market rates. This finding provides further evidence that the statistical results in the
        Passmore study are fundamentally flawed.

    •   Over the last six years, movements in the primary mortgage market rates quoted to
        homeowners for fixed-rate jumbo vs. conforming loans have moved nearly to one-to-one
        with movements in secondary mortgage market prices/yields for fixed-rate jumbo MBS
        vs. Fannie Mae MBS. This result suggests that the better pricing on Fannie Mae and
        Freddie Mac securities versus private-label securities is quickly passed on to


Over the past 15 years, a number of researchers have used the Monthly Interest Rate Survey
(MIRS) data from the Federal Housing Finance Board (FHFB) to develop loan-level pricing
models, with special emphasis on estimating the impact of the conforming loan limit on loan
pricing. Recent studies include McKenzie (2002) (which also includes a review of earlier
studies) and Passmore (2003).

There are a number of reasons often cited for using the MIRS data, as opposed to using time
series of commitment yields for jumbo vs. conforming loans compiled from various surveys of
lenders. McKenzie, for example, notes – without any evidence -- that “(m)ortgage rate quotes
(that) appear in the popular press…are unreliable estimates of the jumbo/non-jumbo rate
differential because they are offer prices and not transaction prices.” In addition, the MIRS
database is publicly available; government sponsored; and includes a fairly large sample. For
example, in 2002, 151 lenders reported a total of 350,031 individual loans. And in fact there is a
lot of useful information that can be gleaned from this database. The database is especially useful
in showing that, whatever the actual magnitude of the benefit homeowners receive from being
able to get a conforming loan, the incentive is strong enough to produce a distribution of loan size
that can only be explained by the fact that there is a material benefit. As the accompanying chart
indicates (and as other studies have shown), there are a large number of loans closed right at the
conforming loan limit. This distribution provides strong evidence that pricing in the market for a
loan that is at or below the conforming limit is viewed as highly valuable to homeowners. The
same data show that the adjustable-rate share of total loans closed is materially higher for loans
with a balance above the conforming loan limit – strongly suggesting that the pricing benefit to
homeowners for conforming loans is greater for fixed-rate loans than for adjustable-rate loans.

                                            Distribution of Conventional Loans Closed for Purchase
                                                             by Loan Amount in 2003

Percent of Number of Loans

                             2.5                                                                        Conforming Loan Limit - $322,700





                              $12,700          $92,700      $172,700      $252,700      $332,700      $412,700    $492,700     $572,700    $652,70
                                                                                      Loan Amount

                                       Source: Federal Housing Finance Board, Monthly Interest Rate Survey

                               Figure 1

                                             ARM Share of Conventional Loans Closed for Purchase
                                                           by Loan Amount in 2003
                                                                                                Conforming Loan Limit - $322,700



                 ARM Share





                               $12,700           $92,700      $172,700     $252,700      $332,700      $412,700    $492,700    $572,700    $652,700
                                                                                       Loan Amount

                                       Source: Federal Housing Finance Board, Monthly Interest Rate Survey

            Figure 2

Few studies that have used this database, however, have looked in depth at the dispersion of rates
on loans closed, or disclosed some of the shortcomings of the database for use in a loan-level
mortgage pricing model. These shortcomings include (1) the fact that the data reflect loans
closed in the last five business days of a month, but there is no information on when the loans
were priced; and (2) the fact that there are no key risk data in the MIRS database such as credit
score, debt-to-income ratio, loan documentation, subordinate financing, mortgage insurance, etc.,
that have been commonly used for years in the mortgage industry to price the credit risk of
mortgages. The only risk variable available in the database is the loan-to-value ratio of the
mortgage, and this variable alone is not a particularly good predictor of the credit risk of a loan.

The first point is fairly obvious: most lenders give borrowers the option to “lock in” the rate on
their mortgage application before the loan closes – sometimes just for a month or two, but in
other cases for longer. As a result, loans that close in the last 5 business days of, say, September
may reflect loans that were priced anywhere from March through September – with most pricing
probably concentrated between June and August. If market interest rates were volatile in the
months prior to September, then much of the dispersion of rates on loans closed in that month
will reflect previous interest-rate volatility, as opposed to the characteristics of the loans
themselves. If there were any tendency for conforming loans to have a different pattern of “rate-
locks” than jumbo loans, this volatility could impact any estimate of the jumbo/conforming
spread using the MIRS data.

This point also highlights the serious problems of trying to relate any data from this loans closed
series to contemporaneous data on market yields of other fixed-income securities. Since the loans
closed data are based on past market interest rates with lags that vary over time, relating these
data to contemporaneous yield data or any other contemporaneous data is methodologically is
unsound, and econometric models that assume a contemporaneous relationship can produce
biased and misleading results.

 The second point is also very important. Since at least the mid-1990’s, characteristics such as
credit score, debt-to-income ratio, loan documentation, subordinate financing, mortgage
insurance, prepayment penalties, etc. have been used both to underwrite and to price the risk of
mortgage loans. This is true both in the “prime” mortgage market and in the “subprime”, or
credit-impaired, market. It is clear from the dispersion of rates charged on loans in the MIRS
database that there are variables affecting pricing that are not adequately explained by the loan-
to-value (LTV) of the loan – the only risk variable in the MIRS database. Over the past several
years, the typical range of rates on fixed-rate loans closed in any given month in the MIRS
database has been around 500 basis points, with a clear “fat tail” at the higher end of the rate
distribution – with conforming loan rates generally having a fatter tail at the high end of rates
charged than jumbo loan rates. This dispersion strongly suggests that lenders have used factors
other than LTV to price loans closed in the MIRS database; that there are probably a higher
percentage of subprime loans at or below the conforming loan limit than above the conforming
loan limit in the MIRS database; and that as a result, the MIRS data are ill-suited for use in
estimating a loan-level pricing model. McKenzie (2002) notes that since the mid-1990’s, LTV
explains very little of the movement in rates charged on loans closed from this database for
conforming loans, but he only provides a partial explanation for this phenomenon.2

There are also very low charged interest rates on loans closed in the MIRS database. Past
researchers have suggested that this phenomenon was the result of lenders mistakenly reporting
  He also notes that the MIRS data suggested that LTV had a larger impact on the rate charged on jumbo
loans than it did on conforming loans.

some adjustable-rate loans as fixed-rate loans. While McKenzie (2002) argues that such
misreporting should be less of a problem since 1996, the data suggest that this misreporting still
may be an issue.3

Interestingly, many of these studies reject using quoted rates from surveys on jumbo loans versus
conforming loans, without looking at any evidence of their relevance. Moreover, none of these
studies explored how differences in the prices of Fannie and Freddie MBS versus private-label
jumbo MBS are related either to quoted commitment rate differentials on jumbo versus
conforming loans or to observed differentials on rates charged on loans closed for jumbos versus
conforming loans. This omission is surprising – from the survey side, because there exist lender-
level data on quoted yields from various private services designed to give customers and market
participants mortgage market quotes (as opposed to advertised rates); and from the securities side,
because a number of Wall Street firms regularly quote price differentials between current-coupon
Fannie Mae MBS and the AAA tranche of private-label securities backed by jumbo loans. One
would think that securities execution differences might go a long way towards explaining rates
charged on fixed-rate jumbo loans versus those charged on conforming loans. Yet none of these
previous studies explore this relationship.

The purpose of this paper is twofold. First, it explores the actual distribution of rates charged on
loans closed in the MIRS database, and -- using evidence from other sources – suggests that this
database is ill-suited for use in developing loan-level pricing models for mortgage loans.
Second, it explores whether surveys on jumbo versus conforming loan rates move in a fashion
consistent with observed securities prices for Fannie Mae MBS relative to those for private-label
jumbo MBS securities – and concludes that in fact such survey rate differentials are reasonably
well explained by security price movements.

Included in this study will be examples of how, when interest rates are volatile, the distribution of
rates on loans closed in the MIRS data is clearly based on past interest rate behavior rather than
on the characteristics of the loans themselves. These data provide strong evidence that trying to
relate the MIRS data to contemporaneous market interest rates is a misplaced exercise, and that
any studies that try to derive implications from such misplaced relationships – such as the
Passmore (2003) study -- should be viewed with skepticism.

Description of the MIRS Data

The description of the Monthly Interest Rate Survey (MIRS) conducted by the Federal Housing
Finance Board is shown below:

         The Finance Board asks a sample of mortgage lenders to report the terms and conditions of all
         single-family, fully amortized, purchase-money, nonfarm loans that they close during the last five
         business days of the month. The survey excludes FHA-insured and VA-guaranteed loans,
  He notes that “The Federal Housing Finance Board instituted loan edit procedures in May 1996 that
eliminate all loans categorized as fixed rate if their contract rate is 1 percent or more below the previous
month’s average fixed-rate mortgage rate.” However, that procedure has been changed, and most recently
FHFB economists have used an arbitrary rate level to exclude loans. However, loans below that level still
appear in the fixed-rate loans closed data – although they may be what are known as “builder buydown”
loans. In any case, the fact that FHFB economists feel they need to exclude any loans suggests that there
may be misreporting of loan terms by lenders, which should make researchers skeptical about using the
data to develop loan-level pricing models.

        multifamily loans, mobile home loans, and refinancings. In 2002, 151 lenders reported a total of
        350,031 individual loans.

        The survey provides monthly information on interest rates, loan terms, and house prices by
        property type (all, new, previously occupied), by loan type (fixed- or adjustable-rate), and by
        lender type (savings associations, mortgage companies, commercial banks, and savings banks), as
        well as information on 15- and 30-year fixed-rate loans.4

The loan terms for each loan type (30-year; 15-year; and adjustable-rate) are: contract interest
rate; initial fees and charges; term to maturity; loan amount; purchase price; and loan-to-price
ratio (generally referred to in the industry as LTV). The database also includes an “effective rate”
on loans closed, which is computed by amortizing the upfront fees and charges on a level-yield
basis over 10 years. The ARM data include the time until the first rate reset, and provide market
participants with fairly good data on trends in that market – especially the recent dramatic
increase in the “hybrid ARM” share of the total ARM market.

Not surprisingly, most studies of jumbo/conforming loan rates have focused on 30-year fixed rate
loans, the dominant fixed-rate loan originated for home purchase. ARMs originated in any one
month, in contrast, have considerable variation in margin resets, rate indexes used, etc., that make
any jumbo/conforming analysis for ARMS extremely difficult.5

The attached table shows the frequency distribution of effective rates charged on loans closed for
30-year fixed rate loans above the conforming loan limit (jumbos) and at or below the
conforming loan limit (conforming) for the first nine months of 2003 from the MIRS data. The
exhibit shows summary statistics such as mean, median, mode, maximum, minimum, standard
deviation, and frequency percentiles.

One of the first things worth noting is that the dispersion of rates is considerable: in every month
of 2003, at least one conforming loan closed had a rate over 10 percent, and the average spread
between the 95th percentile and the 5th percentile was 148 basis points. Such a distribution is well
beyond anything explainable by the loan-level data in the MIRS dataset. Moreover, in each
month there are a number of both jumbo and conforming loans with exceptionally low interest
rates, and there is still a concern that some loans that are classified as fixed-rate by reporting
lenders are, in fact, adjustable-rate loans.

A second thing worth noting is that the dispersion of rates on jumbo loans closed has generally
been lower than that for conforming loans. This is true as measured by standard deviation, as
well as by the differences in the 95th and 5th percentiles. Interestingly, for the last several years
the difference between the average rate charged on jumbo loans and conforming loans has been
less than that of the median rate, and the median rate difference has been less than that of the

  Federal Housing Finance Board Website (www.fhfb.gov).
  It is interesting to note, however, that the ARM share of jumbo loans closed has been consistently
materially higher than that for conforming loans closed. For example, in 2002 the ARM share of total
loans closed in this survey was 17 percent while for jumbo loans it was 52 percent.

                  30-year Fixed Rate Loans Closed from Mortgage Interest Rate Survey, 2003

                            Jan     Feb     Mar     Apr    May      Jun      Jul   Aug      Sep    Average

Conform ing
       Mean                 6.15    6.09    5.96    5.94    5.77    5.53    5.50    5.77    6.13      5.87
       Median               6.00    6.00    5.88    5.88    5.75    5.41    5.38    5.64    6.13      5.78
       Mode                 6.00    5.88    5.75    5.75    5.50    5.38    5.38    5.38    5.38      5.60
       St. Dev.             0.46    0.45    0.46    0.49    0.53    0.53    0.53    0.64    0.65      0.53
       Max                 10.95   10.99   10.95   10.80   11.00   11.00   10.99   11.00   11.17     10.98
       Min                  3.68    3.88    4.40    3.68    3.88    3.15    2.97    3.33    3.26      3.58
       99th percentile      8.00    7.88    7.75    7.84    7.61    7.48    7.40    7.78    8.03      7.75
       95th                 6.90    6.88    6.75    6.76    6.63    6.38    6.44    6.88    7.16      6.75
       90th                 6.61    6.55    6.43    6.42    6.29    6.05    6.07    6.56    6.88      6.43
       75th                 6.25    6.18    6.06    6.04    5.98    5.75    5.67    6.13    6.50      6.06
       50th                 6.00    6.00    5.88    5.88    5.75    5.41    5.38    5.64    6.13      5.78
       25th                 5.88    5.88    5.75    5.75    5.50    5.25    5.25    5.38    5.72      5.59
       10th                 5.79    5.75    5.56    5.50    5.38    5.05    5.00    5.16    5.25      5.38
       5th                  5.75    5.66    5.50    5.41    5.21    4.88    4.88    5.00    5.13      5.27
       1st                  5.50    5.50    5.25    4.89    4.25    4.25    4.50    4.50    4.88      4.84
       95th - Median        0.90    0.88    0.88    0.89    0.88    0.97    1.07    1.23    1.04      0.97
       95th - 5th           1.15    1.22    1.25    1.35    1.42    1.50    1.57    1.88    2.04      1.48
       5th - Median        -0.25   -0.34   -0.38   -0.47   -0.54   -0.53   -0.50   -0.64   -1.00      -0.52

Jum bo Mean                 6.21    6.12    5.97    5.94    5.79    5.62    5.67    5.84    6.13      5.92
       Median               6.14    6.13    5.96    5.88    5.76    5.68    5.64    5.75    6.13      5.90
       Mode                 6.25    6.25    5.88    5.88    5.75    5.75    5.75    5.75    5.75      5.89
       St. Dev.             0.38    0.33    0.38    0.47    0.50    0.51    0.44    0.55    0.56      0.46
       Max                  9.50    8.88   10.99    8.74    9.19    9.65    9.20    9.99    8.78      9.44
       Min                  5.25    5.13    4.88    4.63    4.25    3.89    4.25    4.25    4.25      4.53
       99th percentile      7.70    7.35    7.13    7.60    7.46    7.38    7.43    8.05    7.95      7.56
       95th                 6.75    6.50    6.39    6.67    6.50    6.25    6.38    6.75    7.00      6.58
       90th                 6.50    6.38    6.25    6.38    6.25    6.04    6.02    6.39    6.75      6.33
       75th                 6.38    6.25    6.13    6.07    5.98    5.84    5.75    6.02    6.48      6.10
       50th                 6.14    6.13    5.96    5.88    5.76    5.68    5.64    5.75    6.13      5.89
       25th                 6.00    6.00    5.80    5.75    5.68    5.38    5.40    5.55    5.75      5.70
       10th                 5.88    5.87    5.75    5.61    5.27    5.13    5.25    5.33    5.50      5.51
       5th                  5.75    5.75    5.50    5.13    4.75    4.60    5.25    5.25    5.25      5.25
       1st                  5.50    5.50    5.00    4.63    4.25    4.25    4.88    4.35    4.89      4.80
       95th - Median        0.62    0.38    0.43    0.79    0.74    0.57    0.73    1.00    0.88      0.68
       95th - 5th           1.00    0.75    0.89    1.54    1.75    1.65    1.13    1.50    1.75      1.33
       5th - Median        -0.39   -0.38   -0.46   -0.75   -1.01   -1.08   -0.39   -0.50   -0.88      -0.65

Jum bo-Conform Mean         0.06    0.04    0.02    0.00    0.02    0.09    0.18    0.06    0.01      0.05
                  Median    0.14    0.13    0.09    0.00    0.01    0.28    0.27    0.11    0.00      0.11
                  Mode      0.25    0.38    0.13    0.13    0.25    0.38    0.38    0.38    0.38      0.29

Table 1

mode – something that one might expect if there were a lesser proportion of “riskier” loans above
as opposed to at or below the conforming loan limit.

A third thing worth noting is that during periods where interest rates were exceptionally volatile
and rising in the few months prior to when loans were closed, the dispersion of rates on loans
closed was not only unusually wide, but at times very “bimodal”—that is, many borrowers
“locked in” rates at the lows based on earlier market rates prior to when their loans closed, while
other borrowers were charged rates based on market rates closer to when the loan closed.

The September 2003 MIRS data show this graphically. From the middle of June to early August
of 2003, interest rates rose sharply – for example, the 30-year fixed-rate mortgage yield from
Freddie Mac’s Primary Mortgage Market Survey went from a low of 5.21% in the weeks of June
13th and 20th to a high of 6.44% in the week of September 5th (and it averaged just above 6.25% in
August). As can be seen in the distribution of effective rates for 30-year fixed-rate conforming
loans closed in September, the number of loans closed with rates around 5.25% was about the
same as the number of loans closed at 6.25%. Clearly, this rate dispersion was unrelated to the
risk characteristics of the loans closed, but instead was due to the timing of when borrowers
locked the rate on the loans they closed.

These data provide strong evidence that users of the MIRS data on the rates on loans closed need
to take into account the behavior of market interest rates prior to the date of closing.

Evidence From Other Sources

McKenzie (2002) finds that the LTV of a loan had a greater impact on the loan rate for jumbo
loans than for conforming loans in the MIRS database. He notes that this may in part be because
a greater proportion of conforming loans originated above an 80 percent LTV have mortgage
insurance than do jumbo loans with LTVs above 80 percent6.

Another possible reason for the poor performance of LTV in explaining rates charged on
conforming mortgages is that other factors play a considerable role in the pricing of mortgage
risk. This is especially evident in the subprime market. Consider some data from
LoanPerformance – a private firm that tracks data on loans backing both jumbo securities and
subprime securities. While a serious shortcoming of their data is that points charged on the loans
are not available, their data are illuminating nonetheless.

In the subprime market, rates are generally based on a combination of the rating classification of
the loan by the servicer, which is based heavily on the recent mortgage payment history of the
borrower; the credit score of the borrower; and the LTV of the loan. If the only data one had on
subprime loans were the LTV of the loan, however, one would find that there was no correlation
between the rate charged and the LTV of the loan. Why? To a large extent, it appears that many
lenders limit the LTV of loans to borrowers with very low credit scores. And, in fact, this lack of
correlation is what one finds in the subprime market. Table 2 below shows the simple
relationship between reported LTV and rate charged on 30-year fixed-rate loans closed in the
LoanPerformance subprime database in the first quarter of 2003. As the table indicates, there is
no correlation between reported LTV and rate charged.

 This tendency is in large part because by charter, loans acquired by Fannie Mae and Freddie Mac that
have LTVs above 80 percent must have private mortgage insurance or some other credit enhancement.

     Distribution of Conforming 30-Year FRM Loans in MIRS by Effective Mortgage Rate
                                      September 2003

 Number of Observations




































                                                                      Effective Mortgage Rate

 Figure 3

The data for jumbo loans closed in the LoanPerformance database show a different relationship:
as shown in the table below, rates charged do tend to rise with the LTV of the loan. That result is
not shocking, because there are fewer loans in the jumbo market with low credit scores, and fewer
jumbo loans have mortgage insurance than do conforming loans. As a result, LTV is a better
predictor of risk in the jumbo market than in the subprime market.

This evidence suggests that if there were more credit-impaired loans closed with balances below
the conforming loan limit than above the conforming loan limit – and there is a strong a priori
reason to believe that this is the case – then using the MIRS data to estimate the magnitude of the
jumbo/conforming spread will likely produce a downwardly-biased estimate. It also suggests that
any loan-level pricing model based on the MIRS dataset alone is likely to perform poorly.

A recent study of 26,179 30-year fixed-rate conventional loans closed by a national lender from
1995 to 1997 (Ambrose, LaCour-Little, and Sanders, 2003) highlights the problems with using
the MIRS data for loan-level pricing. This study used additional data on borrower credit such as
credit score, borrower age and income. The study found that, after taking into account these other
variables, conforming loans had yields that were on average about 27.7 basis points below yields
on jumbo loans. Over that same period, the average difference in rates charged on jumbo vs.
conforming loans in the MIRS data was only 10 basis points.

                        Average Interest Rate on 30-Year Fixed-Rate
                         Loans Closed in Q1 2003, by LTV Bucket

                                     Subprime                                        Jumbo
                 LTV                                Rate     LTV                                Rate

                 20-40                              7.75     20-40                              5.79
                 40-60                              7.62     40-60                              5.86
                 60-70                              7.77     60-70                              5.93
                 70-75                              7.86     70-75                              5.96
                 75-80                              7.63     75-80                              6.02
                 80-90                              8.00
                 90-95                              7.71     >80                                6.19
                 >95                                7.62

                 Source: LoanPerformance

              Table 2

Interestingly, the jumbo/conforming spread estimated from this study is very similar to the 26
basis point commitment spread between 30-year jumbo and conforming rates observed from the
HSH Associates survey over this period. The HSH Associates survey of primary mortgage rates
is widely used by industry analysts to track jumbo/conforming spreads.

Without appropriate loan-level risk attributes, it is not possible to discern the extent to which the
MIRS data might understate the jumbo/conforming rate spread. But this earlier study suggests it
could be material. Such a differential could occur, for example, if the share of conforming loans
that were “subprime” were only modestly greater than the share of jumbo loans, since subprime
loan rates are typically 200-300 basis points above prime loan rates.

Given both the timing of pricing versus closing issue and the lack of loan-level risk attribute
issue, one would think that loan-level pricing models using the MIRS database would perform
poorly -- especially over the most recent period, as risk-based pricing has become more prevalent
in the mortgage market. And in fact that appears to be the case. For example, Passmore,
Sherlund, and Burgess (2003), using MIRS data on 30-year fixed-rate loans from April 1997
through February 2003, assess the extent to which LTV, loan size, jumbo status, and other
variables available in the MIRS database explain the rates charged on loans closed -- with
particular attention focused on the impact of the conforming loan limit on rates. Their results
suggest that these variables do a poor job in explaining the rates charged on loans closed over this
period – e.g., the adjusted R2 of their equations averaged only 11 percent over their sample, and
no variable had a coefficient significantly different from zero at the 95% level of confidence.
Their results also produced an average coefficient on a jumbo “dummy variable” of 0.16, but with
a high standard error of 0.10. (Although, as Kogut (2004) notes, this coefficient is highly
dependent on the specification of the variables measuring the effect of loan size on mortgage

rates7.) These results provide additional evidence that the MIRS database is probably not
appropriate for use in estimating the impact of the conforming loan limit on mortgage rates.

A Look at Commitment Yield Surveys and the Capital Markets

Numerous analysts – as well as the Federal Reserve -- have over the years used survey rates on
mortgage commitments as good indicators of overall mortgage rate trends8. Similarly, surveys on
commitment rates for jumbo loans versus conforming loans are widely tracked by mortgage
market participants as indicators of the relative cost of jumbo loans and conforming loans.
Casual empiricism comparing rate quotes available on various lenders’ websites with such
surveys suggests that these surveys are reasonably accurate indications of what the same high
credit-quality borrower would be quoted for a loan at the conforming loan limit versus a loan just
above the conforming loan limit. While it would certainly be preferable to have actual
transaction data on loans priced for borrowers with otherwise identical credit profiles, the analysis
above demonstrates conclusively that the MIRS database on loans closed does not contain such
data. In addition, the results of one study that had better borrower credit data for loans closed
from 1995-1997 suggested that the commitment rate series provided a good indication of actual
jumbo/conforming rate differentials during this period.

In terms of the primary market, one of surveys most widely used by mortgage market participants
to gauge jumbo/conforming spread is that conducted by HSH Associates. This survey asks
lenders for their quoted rates on both jumbo loans and conforming loans for “prime” mortgage
borrowers, which, while not defined, is meant to mean rates that would be quoted for highly
credit worthy borrowers. These rate quotes are not “advertised” rates but instead are
representative rates the lenders would quote prime borrowers. HSH Associates notes that:

        Our survey largely focuses on prices available to ‘A’ credit quality borrowers for the purchase of
        a single-family home. Where required, HSH Associates defines ‘A’ credit as loosely adhering to
        FICO scores of 680-720.

They also note that their survey rates are “…derived from our objectively surveyed data – not
advertising rates.” (their italics).

  Kogut finds that if one uses the specification for measuring the effect of loan size on mortgage rates
recommended by McKenzie (2002), then the coefficient on the jumbo dummy variable over the sample
used by Passmore, Burgess, and Sanders is 0.212. Kogut also details the high degree of sensitivity of the
coefficient on the jumbo dummy variable to alternative specifications, and notes that whatever the
specification, the fit of regressions over this period of time using the MIRS data is poor.
  E.g., the Freddie Mac survey rate is widely used by mortgage prepayment analysts, and it is the only
mortgage rate shown on the Federal Reserve’s “H.15” weekly interest rate release.

                               Jumbo to Conforming Primary Spreads
                                               (Source: HSH Associates)



     Spread (bps)


                                                       Average spread
                    20                                 = 28 bps


                    6/12/1998 3/12/1999 12/12/1999 9/12/2000 6/12/2001 3/12/2002 12/12/2002 9/12/2003

Figure 4

 The attached chart shows how the spread between 30-year fixed-rate loan quotes for jumbo and
 conforming loans has moved over the last six years, using weekly survey data. As the chart
 shows, this spread has been variable over time, but has averaged about 28 basis points since 1998.

 Secondary mortgage market participants track the spread between Fannie Mae MBS and the AAA
 tranche of jumbo MBS as a gauge of how these two markets are performing relative to one
 another. A number of Wall Street firms have tracked the price difference between Fannie Mae
 MBS and the price of a AAA-rated tranche of a private-label security backed by jumbo loans for
 a number of years. Lehman Brothers and UBS Warburg, for example, have made this data
 available since at least 1998. One would think that this relative execution differential would be
 highly correlated to the pricing in the “primary” market for jumbo loans versus conforming loans.
 Yet it appears that no previous academic study of the jumbo-conforming mortgage spread has
 explored this relationship, even though mortgage market participants regularly do so.

 Unfortunately, the data tracked are not readily translated to a secondary market yield differential
 relative to that in the primary market. The Lehman and UBS data, for example, only track the
 “price drop” of a AAA-rated jumbo tranche versus a Fannie Mae MBS, and the “effective rate
 equivalent” of such a price drop is dependent on the market’s “option-adjusted duration” of the

MBS – a number that varies widely both across time and across market participants at any given
point in time.

In addition, an originator who securitizes jumbo loans and sells the created securities must take
into account not just the price of the AAA tranche but also the prices of the subordinated tranches
that provide the credit support for the AAA tranche. The cost of such “subordination” depends
both on the “sizing” of the structure – that is, the size of the subordinated tranches relative to the
senior, AAA tranche – as well as the pricing of the subordinated tranches. Data on such sizing
and pricing are less readily available, and may not be totally reliable. However, both Lehman and
UBS have kept such estimated spreads over time, and it is possible to approximate an “effective
guaranty fee equivalent” using the subordinated tranche sizing and pricing estimates for the
jumbo market. However, one challenge is that there is very little historical data on the pricing of
the very low-rated or unrated subordinated tranches created from private-label deals. These
tranches are not actively traded; are sometimes retained by the issuer of the security; and often are
priced at extremely low prices. While these tranches are generally very small relative to the total
“deal” in a private-label security, their spreads are extremely wide, and can effect overall
execution. The appendix shows a “typical” deal structure.

One approach to comparing primary and secondary market yields using the AAA jumbo MBS
price drop would be to convert this price drop to yield, using the option-adjusted duration of a
current-coupon, 30-year, fixed-rate MBS. That is,

Current-coupon jumbo yield = Current-coupon Fannie yield + (price drop)/duration.

However, there are widespread differences among market participants over the past few years on
estimates of that duration at any point in time. The assumption used in this paper was that, based
on a number of Wall Street firm’s estimate over the last several years, the duration of a newly
originated 30-year fixed-rate mortgage was five years. Using this duration would mean that if a
AAA Jumbo MBS were one percentage point in price lower than a comparable coupon Fannie
Mae MBS, then the yield spread between the two would be 20 basis points.

The second stage in comparing primary and secondary market yield spreads is to convert
subordinate-tranche pricing into an effective yield equivalent, and compare that “implied
guaranty fee” to the guaranty fee charged by Fannie Mae. This exercise is not straightforward.
Dealers have typically kept track of subordinated tranche spreads to Treasuries, so one needs to
convert these to spreads to the AAA tranche. Next, one needs to take into account the fact that
deals are typically structured so that prepayments all go to support the AAA tranche in the first
several years, so that the duration of the subordinated structures is generally considerably longer
than that of the AAA tranche. An additional complication is that the lowest-rated or unrated
tranche spreads have typically not been tracked over time.

The approach taken in this analysis was as follows: first, based on discussions with dealers, an
approximation of the pricing of the lowest-rated and unrated tranches was made based on
movements in the spreads on the higher-rated subordinated tranches. Second, the sizing used was
based on an example shown in Lehman (2002). Next, the resulting “spread effect” was converted
to a price drop using a duration of 7.75 years, which is close to the duration of the subordinated
tranches shown in Lehman (2002). And finally, this price drop was converted to yield on the
whole private-label deal using a duration of 5 years. Details are in the appendix.
Obviously, this methodology is far from perfect: deals vary based on the credit characteristics of

actual loans; sizing from rating agencies has changed over time; the spreads are only indications;
and the duration of the subordinate pieces almost certainly varies over time. Still, this seemed
like a reasonably starting point.

The next step was to see how these secondary market indicators were related to primary market
rates, and this was done by regressing primary jumbo/conforming spreads against these secondary
market spreads. The dependent variable in the regressions is the effective jumbo/conforming
commitment rate spread from the HSH Associates weekly survey (primary j/c/spread). The
independent variables are: (1) the spread AAA jumbo/Fannie Mae yield spread defined above
(secondary j/c spread); and (2) the spread between the “effective” guaranty fee from subordinate
tranches of jumbo MBS deals and the Fannie Mae guaranty fee (gfee j/c spread), the latter of
which was assumed to be 18 basis points over this period.

It should be noted that one should expect some noise in the regression, as the timing of the survey
and the timing of the price “marks” from dealers in any given week was not the same.

With that caveat in mind, the results were as follows, using weekly data from June 1998 to
December 2003:

Regression 1

primary j/c spread = 11.7 bp   + 0.85 * secondary j/c spread + 1.11 * gfee j/c spread
  t-statistics       (9.20)      (12.77)                       (10.29)
R2 = 0.45                        D.W. 0.18

Given the timing differences; the somewhat crude adjustment of price to yield; and the absence of
hard data on the riskiest subordinated tranches, the results are actually quite good. If
contemporaneous and one-week lagged secondary spreads are used instead, a slightly better fit is

Regression 2

primary j/c spread = 10.7 bp   + 0.91 * secondary j/c spread + 1.06 * gfee j/c spread
  t-statistics       (8.50)      (13.22)                      (9.92)
R2 = 0.47                        D.W. 0.13

<coefficients reflect sum of current and one-week lagged coefficients)

The same regression, using monthly average data, yields the following results:

Regression 3

primary j/c spread = 10.9 bp   + 0.90 * secondary j/c spread + 1.14 * gfee j/c spread
  t-statistics       (4.12)      (6.42)                        (5.18)
R2 = 0.47                        D.W. 0.27

The results suggest that the primary rates implied by the HSH Associates’ survey are reasonably
well explained by movements in the secondary market. An interesting finding is that the
coefficients on the secondary market variables for all of these regressions are fairly close to one.

To be sure, in looking at the residuals from this simple regression, there is clearly positive
autocorrelation in the error terms. Based on what was happening in markets over this time, this
pattern is highly likely to be caused by “missing variables” – and, as a result, a simple
autocorrelation correction is not appropriate. For example, residuals from this simple regression
suggest that primary market jumbo/conforming spreads were consistently lower than this simple
regression would have suggested in 1999; higher during most of 2001; and lower during most of
2002 and the first half of 2003.

As the next chart shows, these residuals seem to be related to the ebbing and waning of bank
demand for whole mortgage loans during this period. This chart graphs the residuals from
regression 3 and the annualized growth rate in residential whole-loan mortgage holdings by banks
over this period, from the Federal Reserve Flow of Funds – which, unfortunately, are available
only quarterly. That bank demand for whole loans might affect primary vs. secondary market
spreads is intuitive: if bank demand for whole loans rises, many banks become aggressive in
pricing relative to secondary execution. Since banks can purchase the higher-yielding jumbo
loans, this activity can put downward pressure on the primary jumbo/conforming spread relative
to that seen in the secondary market. Weak bank demand for whole loans would be expected to
have the opposite result.

Other factors appear to have been affecting jumbo/conforming spreads during part of this period,
however, especially during 2001. During that year the primary jumbo/conforming spread moved
up considerably higher than the secondary market spread, and stayed there for much of the year.
In looking back at that period, there were a number of forces at work that may have explained this

    (1) A sharp rate drop at the end of 2000 caused overall mortgage applications to jump
        sharply, while at the same time the mortgage origination industry was still downsizing in
        terms of total employment. Those conditions may have caused a temporary rise in
        originator profit margins, with lenders increasing their margins more in the jumbo market
        than in the conforming market because of less-intense competition.

    (2) Volatility in the overall credit markets, and especially the lower-rated credit markets,
        increased dramatically near the end of 2000 and into 2001, with the spreads on B-rated
        corporate bonds rising considerably and becoming more volatile relative to A-rated
        bonds. While there are no solid data available, these conditions may have caused the
        spreads on the very low-rated or unrated subordinate tranches from private-label jumbo
        deals to widen by more than past relationships might suggest, causing mortgage
        originators to increase their jumbo mortgage rates relative to their conforming mortgage

                                Annualized Growth in Banks Whole Loan
                               Holdings and Residuals from Regression 3

                      30%                                                                                                         30
                                Annualized Growth
                                in WL Holdings
                                at Banks

                                                                                                                                        Regression Residual
                      20%                                                                                                         20
       WL Holdings

                      10%                                                                                                         10

                      0%                                                                                                          0

                     -10%                                                                                                         -10






































      Figure 5

   (3) Another development was the implementation for GAAP accounting of Financial
       Accounting Standards No. 133, which required that derivative instruments be treated as a
       financial asset and reported on the balance sheet at fair value, at the beginning of 2001.
       This change had a profound potential impact on earnings resulting from hedging
       Mortgage Servicing Rights (or MSRs) with derivatives, producing considerably angst in
       the mortgage industry. It also produced a flurry of Wall Street research pieces on MSR
       hedging and FAS 133. (Lehman (2000) and Lehman (2001) are excellent examples of
       such pieces). Since the liquidity of agency-eligible MSRs tends to be greater than that
       for non-agency MSRs, this may also have caused originators to increase their primary
       jumbo loan rates relative to secondary market rates during this period.

   (4) Finally, during the first three quarters of 2001, most mortgage market modelers observed
       an increase in the option-adjusted spreads on mortgages – attributed in part to some of
       the developments above. That widening in spreads, in turn, led to considerable increases
       in Fannie Mae and Freddie Mac purchases of mortgages for their on-balance sheet
       portfolios. Their “bid in the market” during this volatile time may have led mortgage
       lenders to build less of a “risk premium” into their primary conforming loan rates than
       was the case for their primary jumbo loan rates.

None of the above factors would necessarily be expected to have a permanent impact on

jumbo/conforming primary spreads vs. secondary spreads. Taken all together, however, they may
have had a temporary but significant effect on spreads in 2001. It is beyond the scope of this
paper to estimate the effect of each factor on overall spreads. However, there was sufficient noise
during that period to suggest running a regression excluding the 2001 data. The results are shown

Regression 4

primary j/c spread = 5.9 bp   + 1.09 * secondary j/c spread + 0.80 * gfee j/c spread
  t-statistics       (4.16)      (6.47)                      (5.25)
R2 = 0.69                         D.W.    0.52

<excludes 2001>

This result produce materially better results in terms of fit than does the regression including the
2001 data. Moreover, the coefficient on the secondary market spread is not materially different
from one.

An alternative specification would be to assume that a “shock” occurred in the mortgage market
in early 2001, but that shock gradually eroded in the latter part of the year. To proxy this, a
dummy variable (D) was created that is equal to zero up to 2001; is equal to one at the beginning
of 2001 through mid-2001; and then decays linearly down to zero through the rest of the year,
was used. The results of the regression including this dummy variable are as follows:

Regression 5

primary j/c spread = 6.3 bp + 1.09 * sec. j/c spread + 0.89 * gfee j/c spread + 11.9 bp *D
  t-statistics       (3.08)    (10.24)                  (5.36)                   (7.36)
R2 = 0.71                        D.W. 0.59

Finally, this last regression was run using a correction for first-order autocorrelation. The results

Regression 6

primary j/c spread = 10.1 bp + 0.94 * sec. j/c spread + 0.83 * gfee j/c spread + 4.8 bp *D
  t-statistics       (3.73)    ( 7.85)                 (3.46)                   (3.73)
R2 = 0.87                        D.W. 1.69        AR (1) = 0.82 (9.95)

Actual vs. fitted values of the primary jumbo/conforming spread from this regression are shown
in the attached figure.

While the relative coefficients from these various specifications are slightly different, there

implications are similar: first, over most of the last six years, primary fixed-rate mortgage rates
quoted by lenders to homeowners for jumbo vs. conforming loans are highly correlated to
secondary market yields on private-label MBS backed by jumbos vs. market yields on Fannie
Mae MBS; and second, that these primary rates move about one-to-one with secondary market

It is worth talking a bit about the constant term found in these regressions– which implies that the
spread between jumbo and conforming rate quotes in the primary market tends to be higher than
that observed in the secondary market. This finding is not too unexpected. Mortgage lenders who
originate jumbo loans with the intent to securitize and sell feel that they face greater hedging
costs, higher structuring costs, and more aggregation risks than lenders who originate conforming
loans with the intent to securitize and sell. That greater risk is related to considerable uncertainty
about subordinate tranche execution; the hedging of the AAA jumbo tranche arising from swings
in the Fannie Mae/AAA jumbo spread (most of the originators of jumbo loans planning to sell in
securitzed form hedge with Fannie MBS); and the need to hedge for longer periods of time,
because private-label deals have much larger minimum sizes than do Fannie Mae MBS. Lenders
who originate conforming loans with the intent to securitize and sell, in contrast, know that if
they hedge with Fannie Mae MBS and have a contractual guaranty fee, that their hedge risk is
reasonably well known. In addition, the structuring costs and bid/offer spreads in the jumbo
market exceed those in the Fannie Mae MBS market (the secondary market executions used here
are based on “offer” prices). While summing up the impact of all of these factors in basis points
is not simple, an overall difference in the 5-10 basis point range is well within the scope of

                             Actual And Estimated Jumbo to
                         Conforming Spreads From Regression 6

         40                                                                       Actual












 Figure 6

The MIRS data and Securities Prices

In order to see if relating the MIRS data to contemporaneous market data might produce biased
results, a regression relating the average jumbo/conforming spreads from the MIRS data to these
jumbo/conforming secondary market spreads might be illuminating. Such a regression, using
monthly average data, is shown below:

Regression 7

MIRS j/c spread = -4.5 bp + 0.60 * secondary j/c spread + 0.50 * gfee j/c spread
                 (-0.86)     (2.16)                       (1.14)
R2 = 0.05                    D.W. 0.94

Obviously, this regression suggests that MIRS data regarding effective rates on loans closed are
not highly correlated with or well explained by contemporaneous market interest rates. This
result is not surprising, since the MIRS data on loans closed reflect recent and past interest rates.

The next regression relates the MIRS data to market data lagged one month.

Regression 8

MIRS j/c spread = - 8.6 bp + 0.84 * secondary j/c spread(-1) + 0.30 * gfee j/c spread(-1)
                   (-1.68)    (3.10)                            (0.69)
R2 = 0.11                       D.W. 0.93

This regression suggests a stronger relationship between the MIRS data and lagged market rates,
although not nearly as strong as the relationship between the survey rate spread and the secondary
market rate spread shown earlier.

These data suggest that the HSH survey rate information, which moves much more closely with
market interest rates, almost certainly provides analysts with a more accurate picture of
jumbo/conforming loan rate differentials than do the MIRS data.

Note that the regression relating the current month’s MIRS data to current market interest rates
produces a dramatically lower coefficient on the jumbo/conforming yield spread coefficient than
is seen either in the lagged regression, or in the regression relating survey data to MBS data. This
result provides evidence that regressions relating the MIRS data to contemporaneous market data
are likely to produce downward-biased estimates of the impact of market forces on mortgage
rates and jumbo/conforming spreads.


The discussion and results presented here suggest that recent studies that have used the MIRS
data to estimate the difference between mortgage rates on jumbo loans and rates on conforming
loans have almost certainly understated the differential between fixed-rate jumbo mortgages and

fixed-rate conforming mortgages. Moreover, studies that have related the MIRS data to
contemporaneous movements in other market data almost certainly have produced biased and
potentially misleading results of the effect of those market data on jumbo/conforming mortgage
rate differentials.

A particular example is the recent paper by Passmore, Sherlund, and Burgess (PSB, 2003). This
paper first formulates a simple loan-level pricing model and applies it using the MIRS data for
30-year fixed-rate loans. They find that this model, using this database, does a poor job in
explaining the rates charged on loans closed. The paper then uses the resulting statistically
insignificant results on the estimated jumbo/conforming spread over time to contemporaneous
market data. As this paper strongly suggests, this methodology, using the MIRS database, is not
well suited either to estimate the jumbo/conforming spread, or to estimate the impact of other
market variables on this spread. It also strongly suggests that the conclusions drawn from the
Passmore (2003) paper on the impact of GSE activity on mortgage rates, which are based heavily
on the PSB results, should be viewed with a high degree of skepticism.

In marked contrast, the data from surveys of lenders on jumbo/conforming spreads are highly
correlated with secondary market movements in Fannie Mae and Jumbo MBS prices, and appear
to provide a much better indication of jumbo/conforming spreads to homeowners. Recent loan-
level studies incorporating borrower credit data provide additional support for this view9.

However, much more research needs to be done. This study was just a first, quick look at the
data. More data and analysis are needed on secondary market prices and yields to understand
more fully how they impact mortgage lender behavior and interest rates for homeowners.

    See Ambrose, LaCour-Little, and Sanders (2003).

Senior-Subordinated Structures for Private-Label MBS

Private-label – that is, “non-agency” – mortgage-backed securities are typically structured in what
is known as “senior-subordinated” structures, in which credit risk is redistributed through “credit-
tranching”. This “tranching” generally involves the creation of a AAA-rated, senior tranche, as
well as lower-rated subordinated tranches that absorb all credit losses up to a certain amount.
This structure reflects the fact that many investors are willing to take the prepayment risk
associated with uncertain mortgage cash flows, but are less willing to take the credit risk
associated with borrower default. The analogue in the agency MBS market is that Fannie Mae
and Freddie Mac guarantee principal and interest payments to MBS investors, making their MBS
AAA-rated, and in return they charge mortgage servicers a guaranty fee.

The size of the subordinated tranches relative to the senior tranche is dependent on the credit
quality of the underlying loans, and is based on rating agency models of loan performance.

An example of a deal representative of a $250 million new-issue jumbo transaction, taken from
Lehman (2002), is shown below:

Hypothetical Non-Agency CMO

Bond              Credit Rating            Amount        % of Deal

Senior           AAA                     $240.000 mm            96.00%
Sub. 1             AA                       3.125                 1.25
Sub. 2              A                       2.500                1.00
Sub. 3           BBB                        1.875                0.75
Sub. 4            BB                        1.250                0.50
Sub. 5              B                       0.625                0.25
Sub. 6         Not Rated                    0.625                 0.25

In this case, credit losses up to $10 million, or 4 percent of original balance, are borne solely by
investors in the subordinate tranches.

Lehman(2002) notes the following:

         Typically, a generic non-agency senior-subordinate structure incorporates shifting interest,
         through which both credit and prepayment risk are re-allocated by redirecting principal payments
         to the senior tranche according to a specified schedule. The objective of a shifting interest
         structure is to increase subordination to the senior tranche. Specifically, to build credit
         enhancement to the senior tranche, the pro rata share of prepayments attributed to the subordinate
         class is allocated to the shifting interest schedule. Scheduled principal payments are distributed
         between the two classes pro rata.

Lehman gives the following example, in which during the first five years, the AAA tranche
receives 100% of all prepayments. After that, the AAA bonds receive a decreasing proportion of
the subordinates’ pro rate share of prepayments, while still receiving their pro rata prepayment
amount. By effectively “locking out” the subordinates from receiving prepayments during the
first five-years, the subordinates have a dramatically longer average life, but more average life

stability, than the AAA tranche.

Class AAA Prepayment Allocation

Distribution Period (months) Class AAA Total Prepayment Percentage

1-59                                       100% of Total Prepayments
60-71                                      70% of Sub Prepayments+Class Pro Rata Share
72-83                                      60% of Sub Prepayments+Class Pro Rata Share
84-95                                      40% of Sub Prepayments+Class Pro Rata Share
96-107                                     20% of Sub Prepayments+Class Pro Rata Share
108- Forward                               Class Pro Rata Share

Typically, the subordinate tranches are quoted on a yield-spread basis – that is, as a spread to the
Treasury curve, or often the 10-year Treasury rate. They often have the same coupon as the AAA
tranche, but because of their lower credit quality and longer durations, these securities typically
trade at a discount – with the lower-rated tranches typically trading at very steep discounts. There
are seldom quotes available for the unrated tranche, and at times the B-rated tranche is not

Lehman has kept historical spreads to the Treasury curve for representative subordinate tranches
from jumbo deals rated AA, A, BBB, and BB. Averages from June 1998 through December 2003
are shown in the chart below.

                   Subordinate Spreads to Treasuries: Averages from
                                   6/1998 - 12/2003

   basis points

                          AA           A           BBB              BB

  Source: Lehman Brothers, AA – BB rated.

While spreads on the B-rated and unrated tranches are unavailable over this entire period, UBS
did for a time keep estimated spreads on B-rated tranches. The average spreads for the six months
ending in December 2000 are shown in the next chart.

                     Subordinate Spreads to Treasuries: Averages from

   basis points

                          AA       A       BBB       BB          B        Unrated

  Source: UBS Warburg; Unrated spread estimated

These spreads are well explained by a simple function

Spread = a + b * exp (rating)

 where rating is converted to a number, with AA = 0, A = 1, BBB = 2, etc. This function was
used to estimate the spread on the unrated tranche.

Using these spreads and the weights from the representative structure above, one can get a
“weighted average” spread on the subordinated structures. These spreads, unfortunately, are
relative to the 10-year Treasury. To estimate the spread relative to the senior tranche, one needs
to compare that spread to the AAA tranche/10-year Treasury spread.

This spread, however, does not measure the effective basis-point cost of the subordination, since
the durations, or average lives, of the subordinated tranches are materially longer than the
underlying mortgages, and as a result each basis point in yield on the subordinated tranche
translates into a larger price discount relative to its proportionate share of the principal balance of
the total structure. A better proxy would be to convert the incremental spread to price, and then
translate that price drop to yield using the effective duration of the underlying collateral.

While each structure is different, and mortgage durations change over time, an approximate
conversion would be to use a duration of about 7.75 years on the subordinate tranches, and a
duration of about 5 years for the underlying mortgages. Such a simple conversion would mean
that for each 10 basis points of incremental yield spread on the subordinate tranches, the
“effective” cost on the whole pool of mortgages to the issuer converts to about 15.5 basis points
(10*7.75/5) times the percentage of the issue that is subordinated.

To be sure, this methodology uses only representative spreads, and employs a relatively simplistic
conversion of these spreads to yields to “effective credit cost”. However, it seems to produce
reasonable results. For example, using this methodology, the average “guaranty-fee equivalent”
of a private-label jumbo MBS deal was about 19 basis points, very similar to the average
guaranty fee charged by Fannie Mae for 30-year product over this period. Moreover, the
statistical results suggest that movements in this guaranty-fee equivalent are correlated with
primary jumbo/conforming spreads.

Ambrose, B.W., La-Cour-Little, M., and Sander, A. (2003). “The Effect of Conforming Loan
Status on Mortgage Yield Spreads: A Loan Level Analysis,” mimeo from authors, November 5,

Kogut, D. (2004). “Alternative Specifications for Estimating the Jumbo/Conforming Spread
Using the MIRS Data: A Look at the Passmore Results,” internal Fannie Mae memo.

Lehman Brothers (2000). “Mortgage Servicing Rights: Risk Exposures and Hedging Strategies.”
Fixed-Income Research.

Lehman Brothers (2001). “SFAS 133/138 – Observations on Mortgage, CP and FX Hedging.”
Bank Strategies Group.

Lehman Brothers (2002). “An Introduction to the Non-Agency CMO Market.” Fixed-Income

McKenzie, J. (2002). “A Reconsideration of the Jumbo/Non-Jumbo Mortgage Rate Differential,”
Journal of Real Estate Finance and Economics, 25, 197-213.

Passmore, W. (2003). “The GSE Implicit Subsidy and Value of Government Ambiguity,”
Federal Reserve Working Paper 2003-64.

Passmore, W., Sherland, S., and Burgess, G. (2003). “The Effect of Housing Government-
Sponsored Enterprises on Mortgage Rates,” mimeo from authors, December 29, 2003.


To top