QUANTITATIVE PERSPECTIVES.

AN IMPLIED PREPAYMENT MODEL FOR MBS

Eknath Belbase, PhD.

January 2001
 AN IMPLIED PREPAYMENT MODEL FOR MBS
Eknath Belbase, PhD.

JANUARY 2001

QUANTITATIVE PERSPECTIVES ..
Introduction
In this article we describe a market-implied prepayment model for
residential mortgage-backed securities. The methodology takes, as
given, the market prices of securities together with a yield curve
environment and produces a small set of parameters which describe
a prepayment model.

This prepayment model can, then, be used on a broader portfolio to

obtain market-implied valuation and risk measures. In addition to
having potential hedging applications, this technology is a useful
gauge of market sentiment over time and can be used to guide the
evaluation of a prepayment model based on historical prepayment
data.
A Two-Bond Example

We begin by motivating the development of an implied prepayment model

with a two-bond example. Given two FNMA 30 year bonds with net
coupons of 6.5 and 8.0 and the current yield curve environment, an option-
adjusted spread (OAS) is usually calculated by generating a set of interest
rate paths consistent with the yield curve, generating prepayments using
the longer end of the simulated yield curves, and using the simulated short
rates plus a spread to discount these cash flows. The OAS is that single
number which, when added as a spread to all the rate paths, matches the
average net present value of cash flows from the paths to the market price.
In this analysis the prepayment model and market prices are given and we
solve for OAS.
 Instead of solving for OAS, we could try to speed up and slow down the
prepayment model being used by scaling its output SMMs by a constant
factor and try to match the OASs of the two bonds; since the OAS is the
excess spread paid for the bond by the market after taking prepayment risk
into account, if other characteristics of the bonds are similar, it would be
reasonable to expect their OASs to be equal. Table 1 applies this idea to a
FN6.0 and FN8.0 using the Andrew Davidson & Co., Inc. prepayment
model (ADP) with an October 27th, 2000 treasury yield curve.

Table 1
Treasury
FN6.0 and OAS at Curve in
Price 100% ADP Months Yield
FN8.0 OASs FN6 94.34 118 3 6.335
FN8 101.56 132 6 6.372
using three ADP
OAS at
speeds and 80% ADP 12 6.159
FN6 110 24 5.930
10/27/00 data FN8 140 60 5.785

OAS at 120 5.713

120% ADP 360 5.741
FN6 126
FN8 125 Volatility % 17
Reversion % 2
No. of Paths 50

The first analysis shows the OAS of the two bonds using an unadjusted
ADP model. When we decrease the prepayment speed by applying 80% of
ADP, the timing of principal is extended for both bonds. For the discount
this means a lower spread is required to match the same price, and for the
premium, a higher spread. In the final run, we apply 120% of ADP; this
shortens principal timing, increasing net present value for the discount and
decreasing it for the premium. Hence, this increases the OAS of the
discount and decreases the OAS of the premium. The values of the OAS
match almost exactly at 120% of ADP.

2 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 A General Portfolio of Bonds

In Figure 1 we plot the OASs of FN6-FN9 bonds under the three scalings
used earlier with the same yield curve and their market prices from that
date.
Figure 1
FNMA Portfolio OASs
A FNMA
200
Portfolio with 180

10/27/00 Market 160

OAS
140
Prices 120
100
FN6 FN6.5 FN7 FN7.5 FN8 FN8.5 FN9
Bond

1 ADP 0.8 ADP 1.2 ADP

With more than two bonds, it is not always possible to match all the OASs
using a simple scaling. While it is clear that the FN6.5 through FN8 bonds
have almost the same OAS using 120% of ADP, some more flexibility is
required to bring the OASs of the other bonds closer to some common
OAS.

The Andrew Davidson & Co., Inc. prepayment model has additional tuning
features which give us this flexibility. The steepness tuning parameter
changes the shape of the refinancing incentive curve.1 This leads to faster
prepayments on premiums and slower speeds on discounts in one direction
and the reverse in the other direction. Figure 2 displays the same OAS
profile curves under 20% changes in this parameter.

Figure 2
FNMA Portfolio OASs
The Effect of the
200
Steep parameter
180

160
OAS

140

120

100
FN6 FN6.5 FN7 FN7.5 FN8 FN8.5 FN9
Bond

1 Steep 0.8 Steep 1.2 Steep

1 With the ratio between the gross rate and market rates on the x-axis and the refinancing incentive,
which can be thought of as the CPR due to refinancing, on the y-axis.
3 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 In contrast to Figure 1, where increasing the scale parameter seems to
rotate the OAS profile clockwise while keeping the center relatively
constant, the OAS profile moves up and down; in addition, the slopes of
different parts of the curve can be changed by different amounts. This
flexibility is useful when we try to equate the OASs of more than two
bonds.

The slide parameter controls the ratio at which the refinancing incentive
curve is "at the money." Decreasing this parameter changes a slight
discount to a deeper discount while increasing it changes a discount to a
premium. Figure 3 shows the effect on the OAS profile curve.

Figure 3
FNMA Portfolio OASs
The Effect of the
230
Slide parameter
180
OAS

130

80
FN6 FN6.5 FN7 FN7.5 FN8 FN8.5 FN9
Bond

1 Slide 0.8 Slide 1.2 Slide

The burnout tuning dial tends to have the least impact on valuation: on
discounts it tends to have no impact at all, and on premiums, an impact that
is small relative to the first three tuning dials. This can be seen in Figure 4.

Figure 4
FNMA Portfolio OASs
The Effect of the
180
Burnout parameter
170
160
OAS

150
140
130
120
FN6 FN6.5 FN7 FN7.5 FN8 FN8.5 FN9
Bond

1 Burn 0.8 Burn 1.2 Burn

4 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 Minimizing OAS Dispersion

Because the bonds in the FNMA portfolio have the same credit quality, we
would expect their OASs to be the same under the market-implied
prepayment model.2 This is because the OAS is the total spread minus the
option cost; differences in prepayment risk have already been accounted
for in the option cost.

We can use the standard deviation of the OASs of the portfolio of bonds as
a measure of the dispersion of the OASs. Using this definition, the implied
prepayment model is, then, the prepayment model (defined implicitly
through the tuning parameters) which minimizes this measure of
dispersion.3 Figure 5 shows a 3-dimensional representation of a surface on
which we would search for the minimum if we were using only two tuning
parameters.
Figure 5

A Surface

Using Two

Parameters

2 Assuming there is no liquidity premium for some of the coupons. We will revisit this
assumption later.
3 It is also possible to use a target vector of OAS.

5 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 While the general optimization would be in 5 dimensions (with 4 tuning
parameters), this simpler case makes several things clear. First, it is
possible that the boundary of the acceptable-tuning parameter space will
be hit. There is no guarantee that the value of OASs for the portfolio at
such a point would make any sense. In addition, there appear to be several
regions where the standard deviation is fairly low; to choose between such
regions, it would be useful to have some sense of a preferred region and
some ability to distinguish between nonsensical OAS values and
reasonable OASs built into the model. Finally, to obtain a minimum to
within 5-10 bp, it is clear that a simple grid search would have to search
points no further apart than 0.1 in the space of tuning parameters.

Methodology and Software Implementation

A complete grid search of tuning parameters with a 0.1 spacing and each
tuning parameter ranging between 0.5 and 1.5 would force us to perform
over 14,000 portfolio OASs. This amounts to over an hour for each
instrument (assuming that we are dealing with TBAs and not CMOs) in
our portfolio. In addition, this method does not include any information
about preferred regions or output OAS values that are acceptable.

For these reasons, we chose to implement the implied prepayment model

using penalty functions and a state-of-the-art non-linear least squares
optimization technique.4 The penalty function essentially says that the
market view should not severely disagree with a view based on historical
prepayments. Thus, any movement away from the default tuning
parameters of 1.0 must be earned by lowering the standard deviation of
OAS, and the further away from the default values, the larger the burden.
One way to specify this linearly would be that each 1% change in tuning
must be earned with a 1 bp reduction in the standard deviation of OAS per
security in the portfolio.

In addition, our analysis showed that seasoned, high-coupon loans had

consistently higher OASs, and these could not be decreased within our
framework. Generally, these loans tended to be illiquid and the accuracy
of the price information on them was in question. Rather than completely

4 The algorithm is a subspace trust region method based on the inferior reflective Newton method.

6 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 discarding these higher coupons, we used weights on the different coupons.
This allows us to place different amounts of emphasis on the information
contained in the prices and OASs of different securities.5

The non-linear optimization was performed using the lsqnonlin function in

the optimization toolbox in the MATLAB environment. This function is a
general non-linear search method which can be used when information
about the slope of the surface on which the optimization is being performed
is not available in closed-form. The software implementation involves this
function calling our MBSOAS and CMOOAS libraries via the
mexFunction feature of MATLAB.

This methodology is currently available to anyone who has access to the

MATLAB environment. In the future we plan to design a stand-alone
routine which has these capabilities.

Some Results

Figure 6 displays October 27, 2000 results from the implied-prepayment

model for GN30 and FN30 portfolios with coupons from 6 through 9.5.6

Figure 6
GN30 OAS Results FN 30 OAS Results
GN30 and
250 250
FN30 OAS

runs for
200 200
10/27/00
basis points

basis points

150 Untuned 150 Untuned

OAS OAS
Tuned Tuned
100 OAS 100 OAS

50 50

0 0
6 7 8 9 6 7 8 9

Net Coupon Net Coupon

5 We can use weights that represent the amount of trading on those coupons which occurred
over a certain period, or the proportion of outstanding loans the coupons represent, etc.
6 The full set of WAC/WAM and volatility assumptions are available on our website at
www.ad-co.com under Quantitative Analysis.

7 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 The actual OAS and tuning values are displayed in Table 2.

Table 2
GN 30-Year FN 30-Year
GN30 and Untuned Tuned Untuned Tuned
Coupon OAS OAS Coupon OAS OAS
FN30 Before
6 104 117 6 118 136
and After 6.5 111 119 6.5 120 135
7 114 120 7 121 137
Tuning
7.5 120 120 7.5 124 136
8 131 119 8 132 136
8.5 142 116 8.5 142 136
9 177 168 9 180 180
9.5 223 211 9.5 218 211
Scale 1 1.1358 Scale 1 1.1072
Steep 1 0.9270 Steep 1 0.9002
Slide 1 1.0798 Slide 1 1.0294
Burn 1 1.0385 Burn 1 1.0052

First, if we look at the standard deviation before and after optimization for
the whole portfolio, for the GN30s it decreases from 41 to 35bp and for
FN30s from 36 to 29bp. However, we weighted the 9 and 9.5 coupons
significantly less. If we look exclusively at the 6 through 8.5 coupons, for
GN30s the standard deviation decreased from 14 to 2bp and for FN30s
from 9 to less than 1 bp.

When the method weighed the higher coupons equally, it tended to lead to
absurd results. There are several explanations for this discrepancy. First,
the price data on the highest coupons is not as accurate as for the more
liquid ones. Secondly, there may be a liquidity premium for the higher
coupons. Finally, the market may be assigning a spread based on faster
prepayments than the base prepayment model for those coupons under
falling rates in a way that is not consistent with the structure of the model
for the lower coupons.

For both collateral types, the market seems to favor a slightly less steep refi
incentive curve than our base model combined with a speeding up overall.
In addition, the values of slide indicate that securities are becoming "in the
money" in the market model at slightly lower ratios than in our base

8 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 model. This is shown in Table 3, where we compare the base case (static
interest rate environment) WAL equivalent PSAs for the two collateral
types under the untuned prepayment model and under the market-implied
model.

Table 3
Untuned Tuned Untuned Tuned
Equivalent Security Base PSA Base PSA Security Base PSA Base PSA
PSAs Before GN6 123 164 FN6 135 174
GN6.5 130 173 FN6.5 139 181
and After GN7 132 178 FN7 144 189
GN7.5 137 298 FN7.5 173 277
GN8 219 460 FN8 310 435
GN8.5 355 584 FN8.5 471 642
GN9 316 392 FN9 375 419
GN9.5 301 363 FN9.5 380 418

Finally, we look at the series of solutions over time. Figure 7 shows the
tuning parameters for the GN30 portfolio based on market close data from
the last Friday of each month between January and December 2000.

Figure 7

Tuning
GN30 History Jan-Dec 2000
Parameters
1.25
Parameter Value

for GN30s for 1.15 Scale

Jan-Dec 2000 1.05 Steep
0.95 Slide
0.85
Burn
0.75
00
0
00

00
-0

-0

l-0

v-
n-

p-
ar

ay

Ju

No
Ja

Se
M

Date

The scale and burn parameters seem to be the most stable over time for
GN30s. Steep does seem to fluctuate over time, both above and below the
steepness of the base model. However, scale and slide are consistently
above 1, within a fairly small (5%) band.

9 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
 It is important to note that the results of the implied prepayment model do
not necessarily provide a forecasting model. Rather, the parameters specify
a prepayment model which reflects the market's assessment of risk in much
the same way that forward rates reflect the market's assessment of yield
curve movements, or implied volatilities from caps and swaptions reflect
the market's view of interest rate volatility. In all these cases, the implied
curve or set of parameters is a composite view based on the often divergent
views of different market participants.

Further Studies & Enhancements

We will continue to examine the relationship of the tuning parameters

output by the implied prepayment model over time. In addition, we will
begin to examine the stability of the implied prepayment model using the
swap curve to forecast mortgage current coupon rather than the treasury
curve.

While we used the dispersion of OASs within a portfolio to infer a

prepayment model, this technology can be used with any other measures
which can be fit to a reasonable target vector, or for which it makes sense
to minimize some measure of distance. For example, we could solve for a
prepayment model that produces a set of effective durations for our
portfolio that matches a set of target durations.

A useful empirical test is to compare hedging performance using the base

prepayment model against the implied prepayment model over some
period of time. Finally, this technology allows us to compare across
different prepayment models by stating one in terms of the other's tuning
parameters. All of these topics will the be subject of further research at
Andrew Davidson & Co., Inc.

10 Q U A N T I TAT I V E P E R S P E C T I V E S J A N U A RY 2 0 0 1
Conclusion

The market-implied prepayment model estimation methodology takes, as

given, a yield curve environment, volatility assumptions and the prices of
MBS securities and outputs a set of tuning parameters which specify a
prepayment model. In contrast to a prepayment model based on historical
data, this implied prepayment model is one which minimizes the
dispersion of OASs among the portfolio securities. The methodology can
be used both to track market sentiment as the basis of an alternate hedging
strategy and as a comparison between different views of prepayments and
prepayment models. The general strategy can be used with measures other
than OAS.

Eknath Belbase is responsible for the research, development and testing

of valuation/risk management tools and quantitative methodology for
Andrew Davidson & Co., Inc. Eknath applies his mathematical and
statistical experience to the development of cutting edge quantitative
financial tools that are efficient, accurate and well-tested. He has worked
on the MBS fixed-rate and ARM prepayment models, designed a lattice
implementation of the Black-Karasinki interest rate process tailored to be
extremely efficient on mortgage instruments, and researched LIBOR-
Treasury and mortgage current coupon dynamics. He has worked on
consulting projects involving structured products, credit risk and
regulatory testing.

Eknath holds a PhD in mathematics with specialization in probability, and

an MS in Statistics, both from Cornell University. He received his BA in
mathematics and computer science from Ohio Wesleyan University.

Eknath Belbase has also written:

A Lattice Implementation of the Black-Karasinski Rate Process June ‘00
Fixed Rate Mortgage Prepayment Model March ‘00
Quantitative Perspectives

ANDREW DAVIDSON & CO., INC.

520 BROADWAY, EIGHTH FLOOR
NEW YORK, NY 10012
212•274•9075
FAX 212•274•0545
EMAIL [email protected]

QUANTITATIVE PERSPECTIVES is available via Bloomberg and the internet.

BLOOMBERG ADQP<GO>
WEBSITE www.ad-co.com

We welcome your comments and suggestions.

Contents set forth from sources deemed reliable, but Andrew Davidson & Co., Inc. does not guarantee its
accuracy. Our conclusions are general in nature and are not intended for use as specific trade recommendations.

Copyright 2001
Andrew Davidson & Co., Inc.