MBS Analytics New Approaches and Techniques

1
MBS Analytics - New Approaches and Techniques

• AFT Breakfast Seminar
• April 2006

2
Agenda

• AFT Basic Modeling Approaches AFT Model
  Structure and PhilosophyAFT Model Types
  
• Prepayment ScoresShort Term ScoresLong Term
  ScoresAFT Prepayment score, its
  performanceGeographic analysisScoring of loans,
  agency pools, CMO, non-Agency CMOBurnout
  calculations, observationsEconomic value of the
  score
• Default ModelingDefault Modeling IssuesDefault
  Model StructureExamples of Fitting ProcessAFT
  non-agency database
  

3
Agenda

• Prepayment Model Performance AFT Model version
  5.42 releaseRecent Prepayments ObservationsDeal
  vs. Loan level ModelShort term vs. Long Term
  Modeling Accuracy Short term and Long term
  dynamic adjustmentsAnalyzing Model Performance
• Other TopicsPrepayment Drivers Primary to
  Secondary Spreads ModelingAFT non-agency loan
  databaseCustom Model fitting Third party
  integration levels

4
AFT Prepayment Modeling Philosophy

• In recent years, new approaches have come to
  dominate modeling of the prepayment behavior of
  mortgage loans. What makes these new approaches
  work is the focus on modeling the borrower
  behavior and the reasons for changes in that
  behavior, instead of focusing on the model's
  tenability within the existing statistical
  machinery.
• The purpose of constructing a prepayment model
  is not to project prepayments. It is impossible
  to project prepayments, since future mortgage
  rates are not known. The purpose of a model is to
  define a relationship between projected mortgage
  rates and the resulting rates of prepayment
  activity, given all available information
  regarding the mortgage, the mortgage holder, the
  current state of the economy, etc. If this
  relationship is well modeled, it will in turn
  allow one to answer all of the questions that one
  may ask as a holder of a mortgage-backed
  security. Questions like the value of the option
  to refinance, the average advantage of owning a
  mortgage vs. owning other instruments, how one
  can compare a mortgage to a collection of bonds
  and short positions in interest rate derivatives,
  etc. Therefore, a well-constructed model should
  incorporate all known factors that effect a
  mortgage holder's inclination to move or to
  refinance, as well as the overall state of the US
  housing market.
• Prediction is a very difficult art, especially
  with respect to the future.
  
• --Mark Twain

5
AFT Model Structure

• OLD
• Parsimonious
• Statistically-fit
• Poor performance
• Unstable
• Required frequent recalibration
• NEW
• Behavior-based
• Well-modeled relationship between projected
  mortgage rates and resulting prepayment behavior
  
• Proven performance in volatile markets
• Seldom recalibration

6
AFT Model Structure

• Two contributors to mortgage prepayments
• Housing Turnover
• Proper connection of housing turnover to the
  projection of the index of existing home sales
  
• Refinancing
• Incentive includes the shape of the curve
• Proper modeling of burnout
• Proper modeling of publicity effects

7
AFT Model Structure

• Step 1 Projecting existing home sales index
• Change in interest rates
• Rates increase/home sales slow down
• Rates decrease/home sales speed up
• Mean-revert over time to a normal level
• Seasonality
• Before 1998 EHS rate has been 3.5-4.3 million
  units/year. After 1998-2005 it has been
  5.5-6.8(latest)
  
• What is it long term? We assume 6.7 million
  units, Solomon assumes about 5.0 million units.
  Who is right?

8
AFT Model - Existing Home Sales
9
AFT Model - Housing Turnover Component

• Mortgage age
• A borrower with a newly purchased home is less
  likely to move.
  
• Lock in effect
• If prevailing rates are higher than current rate
  being paid, the incentive to move decreases.
  
• Self-selection
• Excess points paid discount origination
• No points/no fees - premium origination.
• Historically, AFT translation of EHS rate into
  actual HT prepayments projections has been near
  perfect.
  

10
AFT Model - Housing Turnover Component
11
AFT Model - Refinancing Component

• Refinancing Incentive
• Ratio of WAC to effective mortgage rates
• Two drivers 30-year and 15-year mortgage rates
• Burnout
• Pool is not homogeneous
• There are sub-pools of fast, medium and slow
  refinancers
  
• As the ratio of sub-pools changes the response to
  refinancing incentive changes

12
AFT Model - Refinancing Component Burnout
13
AFT Model - Refinancing Component Publicity

• Publicity
• Overlooked by many models
• Historic lows cause dramatic change in refinance
  pattern
  
• Pools considered burned out start to refinance
• Loans begin to refinance for a lower incentive
• Overall sensitivity increases

14
AFT Model - Refinancing Component Publicity
15
AFT Model - Refinancing Component Other Effects

• High premium-originated loans
• Generally slower than current-coupon originated
  loans
  
• If Available Loan Size, LTV, Documentation
  Level, Geography, Single Family, Primary
  Residence, Cash-out
  
• Some of these variables can be inputted directly
  to the model, others through the scoring
  mechanism.
  

16
AFT Model - Other Model Types

• Five Basic Model Types (different types of
  algorithms)
  
• Fixed Rate Agency, WL, AA, RL, MH
• ARM - Agency, WL, AA, RL
• Hybrid - Agency, WL, AA, RL
• BC/BA
• HEL - Fixed, ARM, HELOC, Prepay Penalty

17
AFT Model - FRM

• The driver is the ratio of effective mortgage
  rate to effective WAC. The structure has been
  discussed in a variety of presentations and
  publications by AFT. Highlights
• First step- project EHS as function of mortgage
  rates
  
• Based on EHS project HT component of the model
• Using a multi-population algorithm project the
  refinancing component

18
AFT Model - ARM

• The overall structure is similar to FRM, the
  incentive structure is different.
  
• Primary driver The difference between projected
  ARM WAC and projected 30-year conforming rate.
  
• Secondary driver The difference between
  projected 30-year conforming rate and 30-year
  conforming rate at the time of ARM origination.
  
• Overall multiplier based on the difference
  between projected ARM WAC and ARM life cap.

19
AFT Model - Hybrid

• Uses an FRM like structure for the fixed period.
• Uses ARM like structure for the ARM period.
• The library accepts input of an ARM structure and
  uses the initial reset period to determine if
  its a hybrid.
  
• Currently have models for 3,5,7,10-year hybrids.

20
AFT Model - BC

• Uses an FRM like structure.
• The same structure is used for both BC fixed and
  BC ARMs.
  
• The refinancing drive is the difference between
  projected 30-year conforming rate and 30-year
  conforming rate at the time of BC origination
  that sensitivity is parameterized can be
  changed to be a function of incentive. The
  refinancing aging ramp is a 2-D function of age
  and SATO (spread at origination)

21
AFT Model - HEL, HELOC

• Uses an BC like structure for economically driven
  prepayments (defined in fr20herf.def and
  fr20heht.def).
  
• Uses a separate structure for credit driven
  refinancings (loan consolidation, term extension,
  credit improvements, etc.) defined in
  hegereri.def and hespecif.def.
• The same structure is used for both fixed and
  ARMs.
  
• Accepts as an input a vector of WAC projections.

22
AFT Model - HEL, HELOC Credit Driven Refinancing

• Measure of credit is SATO (Spread At
  Origination).
  
• Lower credits experience faster credit driven
  refinancings and slower economic driven
  refinancings.
  
• Publicity accelerates credit driven
  refinancings.
  
• Credit driven refinancings decrease as mortgage
  rates increase vs. the rates at origin.
  
• Shorter term loans experience faster credit
  driven refinancings.
  
• Generic credit refinancing function is defined in
  hegereri.def.
  
• For each HEL type FRM, ARM, Prepay Penalty,
  High LTV, HELOC there is a separate refinancing
  aging ramp and separate economic refinancing
  sensitivity (defined in hespecif.def).

23
AFT Model - HEL, HELOCCredit Driven Refinancing
Type Specific

• The model names start with HE followed by
• _GN for HEL generic
• _AR for regular HEL ARM
• _GR for HEL without prepay penalty
• _GP for HEL with prepay penalty,
• _LC for HELOC
• _LP for HELOC with prepay penalty
• _HP for high LTV HEL with prepay penalty
• _HI for high LTV HEL with no prepay penalty
• _HL for HELOC with prepay penalty
• _HA for ARM HEL.
• If only HE name is used, then the the generic
  function is applied, otherwise, for example, a
  name HE_HI would invoke a high LTV HEL model.

24
Projected vs. Historical PrepaymentsVery Few
Surprises
25
Standard modeling vs. scoring

• There is a large number of indicatives which a
  model may not be able to accept as inputs.
  
• Different data sets may have varying contents.
  It is difficult to use the same model on
  different data sets.
  
• It is difficult to compare one loan vs. another.
• The process for interpreting/translating loan
  indicatives into value is complex and expensive
  and requires difficult to find expertise.
  
• Loan information gets lost when transferred.
• Recalculating the burnout function from the
  changing distributions of factors is impossible
  

26
AFT Prepayment Score

• AFT has created a prepayment score that can be
  attached to each loan.
  
• The score reflects all additional data that is
  available at the loan level (other than WAC and
  age).
  
• The score modifies the results of the AFT
  prepayment model and therefore the value of the
  assets.
  
• AFT has created two scores refinancing score
  and housing turnover score
  
• The score described above is a Long term Score
• Short Term Score is just a short term
  projection of prepayments for each loan.

27
These loans have 7.5 coupons and were originated
in August, 1999.
How likely are they to prepay in the next
three months? Should you offer to refinance? Whic
h one should you call first?
28
Adding more data allows the model to
differentiate each loansForecasted prepayment
behavior
So, we have developed a score for prepayment prop
ensity that modifies the prepayment forecast at
the loan level

29
How good was your intuition?
12
18
Will you make money if you solicit a refinance
from
The loans most likely to prepay?
30
Historical Loan Prepayment Speeds by Score
EXAMPLE Conforming loans originated in 1997 with
WAC of 7.5 (-25 basis points)
Bucketed into deciles by Refinancing Score.
31
Geography Contribution to the Prepayment Score

• Housing turnover related prepayments make up a
  relatively small portion of total prepayments
  over the last 5 years
  
• Refinancing related prepayments generally present
  a good statistical sample
  
• Statistics for smaller states are fairly limited
• Used Bayesian statistical analysis to come up
  with best estimates
  
• AFT tracked coefficients stability as a function
  of observation period

32
Geography Contribution to the Prepayment Score -
Refi
33
Geography Contribution to the Prepayment Score -
HT
34
Scoring of Agency Pools

• Additional data is being released by the
  agencies Loan Size, FICO, LTV, Geographic
  Distribution
  
• Using an algorithm similar the the loan scoring,
  we score all of the agency pools in the same
  manner as we score individual loans
  
• The scores for loans, pools, and CMOs are
  available from AFT
  

35
Scoring of Agency Pools
Example
36
Scoring of Agency CMOs

• Loan Size, FICO, LTV, Geographic Distribution is
  generally NOT available for Agency CMOs
  
• Knowing the pools backing CMOs and all individual
  pool scores, AFT calculates CMO scores
  
• User knows up-front which CMO will be a more
  responsive and which less
  
• The differences in economic value can be several
  
  
• Traders can immediately profit from the economic
  value, since the market does not yet recognize it

37
Scoring of Agency CMOs
Example
38
Scoring of Agency CMOs - Analysis
CMO projected vs. observed historic average SMM
using AFT standard model (not adjusted by the
scores)
39
Scoring of Agency CMOs - Analysis
Scored CMO projected vs. observed historic
average SMMs using AFT model modified by the
scores
40
Scoring of Agency CMOs - Analysis
CMO observed historic average SMMs as a ratio to
standard AFT model projections as function of CMO
score
41
Scoring of Non-Agency CMOs - Analysis

• Adequate loan-level information is not available
  for non-Agency CMOs from CMO cash-flow
  generators.
  
• AFT has put together a non-Agency CMO loan
  database or a customers database may be used in
  conjunction with AFT extraction software or AFT
  web site.
• For all non-Agency CMOs, AFT uses an extract from
  the database, where loans are bucketed by Loan
  Type, WAC, WAM, as well as OLTV and MSA if
  default calculations are required. For each
  bucket an HT score, Refi score, as well as
  Default score if needed are calculated
• AFT Software tells CMO cash-flow generator how to
  bucket the collateral.
  
• For each bucket, the software looks up the score,
  and invokes the model using the appropriate
  score.
  

42
Burnout what happens, how do we measure

• The prepayment response function of a pool of
  loans exposed to refinancing opportunity is
  different from the response function of fresh
  loans. The effect is called burn-out
• Burnout happens due to changes in population
  composition
  
• Using the scoring algorithm, we can measure it
  directly
  
• Burnout rate of a relatively homogeneous pool of
  loans will be different from the rate of a
  heterogeneous pool it is a function of the
  standard deviation of scores within pool.

43
Burnout
Example Average score over time of 8 conforming
loans originated in 1997
44
Burnout
Example Difference between historical prepayment
projections by AFT model applied to a pool with
an average score of 500 and standard score
distribution (Pool Model) and the model applied
only to loans with the score of 500
8 conforming loans originated in 1997
45
The AFT Prepayment Score improves knowledge of
the true value of mortgage assets
The score has been used to modify the prepayment
vector used in calculating the price associated
with the market OAS of 50 bps for the securities
and 200 bps for the strip.
46
The AFT Prepayment Score is available at the
point of sale

• MBS market has not priced in the economic value
  of the score. Even low loan balance pools pay-ups
  bear little relationship to their economic vale
  
• Some hedge funds and broker-dealers now bid
  aggressively for MBS with desirable scores and
  avoid the ones with less desirable
  
• Some Originators are beginning to change pricing
  to customers based on the score
  
• Several of the largest banks are using it to
  price and hedge their loans.
  
• Both Long term and Short term score are available
  to subscribers to the McDash data base.
  
• AFT can accept your loans on its FTP site and
  score them (as well as perform a complete OAS
  analysis on them).

47
Default Modeling Issues

• Issuer Related
• 1996 to 1999 data involved a lot of appraisal
  fraud (e.g.Conti)
  
• Issuers have an incentive to understate the LTV
• Issuers have a lot of latitude in dealing with
  delinquencies
  
• Many actually have departments that manage
  triggers manipulate the process of handing
  delinquencies and defaults for the purpose of
  triggering desired outcomes for their securities
  holdings
• Reported delinquencies and defaults for
  individual CMOs often have little to do with the
  underlying borrower behavior.
  
• Over the long term the data on which the models
  are based all of these games average out

48
Default Modeling Issues

• Data Related
• Due to appraisal fraud, data from 1996 to 1999 is
  not necessarily representative of newer deals
  
• Most of the data came from the last 8 years which
  experienced an unprecedented HPI. A lot of
  underwriting liberties consequences have been
  masked by the high HPI
• Losses due to defaults have often been zero
• HPI and Economy
• HPI historically has had a very low correlation
  to interest rates.
  
• Wide distribution of HPI across US. Tails are
  critical
  
• Attempts to connect unemployment simulation to
  defaults is superfluous. There is a one-to-one
  connection between unemployment and HPI

49
Default Modeling

• Drivers
• Primary driver is a function of OLTV and CLTV
• Historical CLTV is calculated based on MSA HPI
  indexes
  
• There has been a wide enough distribution of HPI
  for different MSAs that we can observe the
  effects of negative HPI, but the data set is
  somewhat limited
• The model uses all loan and borrower level
  information available
  
• Structure
• There are a large number of variables that affect
  default propensity which makes it impossible to
  perform aggregations. AFT generates a default
  score based on time-independent characteristics
  thus cutting the dimensionality of the problem
• Transition matrix based

50
Default Modeling Input/Output

• Inputs To Scoring Algorithm
• Originator, MSA, FICO, ?FICO, IO, DocLevel,
  ResidencyType, SingleFamily, LoanPurpose,
  LoanSize
  
• Inputs to The Model
• CLTV, OLTV, Default Score, Prepayment Score, Age,
  Current WAC, Projected WAC, Initial Payment,
  Projected Payment, Projected HPI, Projected
  Mortgage rates (30, 15, 5, 7)
• The prepay/default model may, as an option,
  include the scoring algorithms internally.
  
• Output
• Del30(360), Del60(360), Del90(360),
  Foreclosure(360), Liquidation(360),
  Severity(360), Prepay (360)

51
Default and Prepayment Model deployment for
Non-Agency CMOs - Analysis

• Adequate loan-level information is not available
  for non-Agency CMOs from CMO cash-flow
  generators.
  
• AFT has put together a non-Agency CMO loan
  database or a customers database may be used in
  conjunction with AFT extraction software or AFT
  web site.
• For all non-Agency CMOs, AFT uses an extract from
  the database, where loans are bucketed by Loan
  Type, WAC, WAM, OLTV, MSA. For each bucket an HT
  score, Refi score, a Default score are
  calculated
• AFT Software tells CMO cash-flow generator how to
  bucket the collateral (only WAM-WAC buckets).
  
• For each WAM-WAC bucket, There may be 100s of
  OLTV, MSA buckets. For each of them the software
  looks up the scores, calculates CLTV based on MSA
  specific historic HPI and invokes the model using
  the appropriate score and an input HPI projection
  for each MSA.

52
Default Modeling Structure

• Model Structure
• Calculate principal transition function (from any
  state into the subsequent state) D(t).
  
• Calculate modifier transition function M(T) (that
  modifies other allowed transitions,) keeping D(t)
  unchanged
  
• D(t), M(t) are functions of default Score, CLTV,
  OLTV, Accumulated Prepayments, Change in Pay
  Level, age, seasonality
  
• Follow the transition matrix over time to
  calculate all delinquency transitions.
  
• Loss severity is calculated based on Coupon,
  Average Time to Liquidate, Loan balance, Fraction
  of value recovered, Fixed costs per loan

53
(No Transcript)
54
(No Transcript)
55
Default Modeling Structure
56
Default Modeling Structure
57
Default Modeling Structure
58
Default Modeling Structure
59
Default Modeling Structure

• Total access to parameters. All functions are
  defined as piecewise linear and saved in
  parameter files. Easy to edit, easy to create new
  models and model types.
• F(age)
• 0 0
• .3
• .8
• 1.1
• 0.7
• Fully integrated with AFT libraries (including
  full integration with INTEX)

60
HPI Simulatoin

• There are four ways to simulate HPI.
• A) User supplies a single HPI scenario for all
  MSA
  
• B) User supplies an HPI scenario for every MSA
• C) User does an OAS simulation using an MSA level
  HPI simulation in conjunction with interest rates
  based on MSA level HPI variance-covariance matrix
  provided by AFT.

61
AFT Model New Release Version 5.42New
Structures I

• Changed historical rates file from monthly
  FNCR3010 to weekly MBA survey rates
  
• Changed the connection from between the last
  mortgage rate in Mort30.dat file and the last
  existing home sales projection in htecnmy.dat
  file to a file that contains both of the numbers
  in the same place avoids consistency issues
• Allows to keep a history of the mort30 and
  htecnmy pairs for back-testing
  
• Allowed dynamic creation of the 5 model types
  discussed above
  
• Allows weighting of the intra-month lows to be
  heavier than other rates
  
• Added the ability to have origination year
  dependent RF multiplier
  
• Created the capability to either treat the
  incentive as a ratio of WAC to effective mortgage
  rates or their difference
  

62
AFT Model New Release Version 5.42 New
Structures II

• Lags can now be a function of age for the housing
  turnover component (in addition to the being a
  function of age for the refinancing component)
  
• Elbow can shift as a function of publicity
• Refinancing aging ramp can be a function of SATO
  (the spread at origination)
  
• Accumulator-like function to deal with capacity
  constraint situations
  
• Hybrid-specific ARM parameter files capability
  added which can have sensitivity to when first
  reset took place
  
• Additional spreads as a function of program e.g
  WL, AA, HE
  

63
AFT Model New Release Version 5.42 New
Structures III

• Prepay penalty function can now be used as a
  multiplier to overall refinancing response or a
  modifier to the effective mortgage rate as before
  
• The DLL can accept prepayment scores
• The DLL now reads the normal and actual
  standard deviations of the scores and modifies
  the burnout rate based on that
  
• Ability to handle real loan sizes and
  normalized
  
• Ability to have an elbow be a function of
  normalized loan sizes
  
• FICO sensitivity
• Three ways to interpret age-dependent refinancing
  lags
  

64
AFT Model New Release Version 5.42 Major
Parameters Modifications

• Changed weights for calculating effective
  mortgage rate from month 2 at 25 and month 3 at
  75 to month 2 at 100
  
• Lowered burnout by about 20
• Lowered sensitivity to publicity
• Hybrids are driven by rates difference rather
  than their ratio as an incentive
  
• Prepay penalty function now multiplies the
  refinancing component for HEL/HELOC
  
• Modified lags sensitivity as a function of age,
  rates direction
  
• Made elbow be a function of publicity
• Slightly modified the refinancing curves
• Slightly lowered refinancing elbow
• Variety of modifications for credit collateral

65
Prepayments Observations

• Unprecedented increases in existing home sales
  rate, have been stubbornly high for the last 7
  years.
  
• Every existing home sale is a housing turnover
  related prepayment
  
• What is the long term expected existing home
  sales rate?
  
• The drop off in refinancings of the marginally
  refinancible collateral was greater than implied
  by last 5 years of data. Is the original (version
  5.4) burnout function a better way to model? Is
  amelioration of burnout rate a transient
  phenomenon?
• HT aging ramp has stayed short
• Refinancing aging ramp is now present in all
  models
  

66
Changes in jumbo hybrid responses, near-zero
incentive refinancings and WAC dispersion

• Jumbo hybrids have exhibited significant changes
  in refinancing response
  
• Refinancing response increased for near-zero
  incentive
  
• Refinancing responses for larger incentives are
  inconsistent between different deals
  
• Substantial differences by originator
• Some changes may be explained by the borrowers
  expectations that interest rates would rise
  term extension behaviors
  
• Jumbo deals tend to have a very wide WAC
  distribution
  
• Near zero incentive refinancings may actually be
  refinancings for a substantial incentive
  

67
Deal vs. Loan Level Models

• CMO deals tend to have a wide distribution of
  WACs, origination dates, prepay penalty
  structures, and even of collateral types. It is
  especially true for non-agency deals
• AFT models have been generally fit against
  deal-level data since they usually have to
  perform against CMO deals, or other relatively
  wide aggregates
• Users may need to analyze individual loans or
  loans bucketed tightly by WAC or other
  characteristics.
  
• AFT has come up with a closed form solution to
  translate deal-level model into a loan-level
  model based on the expected and realized standard
  deviation of WACs and other factors
• Will demonstrate the corrections based on WAC
  dispersion, corrections based on other factors
  will be discussed in the scoring part of the
  presentation

68
Deal vs. Loan Level ModelsWAC Distribution
69
Deal vs. Loan Level ModelsRefinancing Curve
Correction
70
AFT Additional Naming Conventions

• Third party systems allow for differing sets of
  inputs
  
• AFT attempts to allow users bypass these systems
  constraints by using the agency name field to
  pass information to the model that these systems
  wont allow. Result complex naming conventions
• Example FNMA_7BLN152.30.711_at_25.B1
• Decoding
• FNMA model type
• _7BLN means use 7-year balloon model. Used for
  systems that would not indicate to our model
  that a balloon collateral is being run
  
• 151.30.71 means 151.3 loan size, 0.71 LTV. Used
  for systems that would not send to our model the
  loan size and LTV
  
• 1 means use the scoring algorithm if available
  and 0 means dont use scoring algorithm
  
• _at_25 means that the WAC distribution of the
  collateral that you are running is 25 basis
  points.
  
• .B1 means use fnma.b1 file for the definition of
  the prepayment penalty structure.
  

71
Short term vs. long term modeling accuracy

• Valuations of MBSs are largely functions of the
  models long-term projections, e.g. over more
  than a year. These valuations will be accurate if
  given an interest rate scenario, the models
  projections are quite close to the actual on that
  scenario
• Retention efforts, dollar roll values, and
  quarterly income projections depend on a models
  short-term projections.
  
• Model evaluations should be clear on whether one
  evaluates the short term accuracy or the long
  term one
  
• There is certain information that is available
  for making short-term projections that is not
  available for making long-term projections.
  
• Along with monthly data releases, AFT also
  releases a set of short term adjustment
  coefficients for Agency collateral. These files
  just need to be placed in the parameters folder
  to be utilized.

72
Projections Without ST adjustment
73
Projections With ST adjustment
74
Deal-Specific Long Term Adjustments

• Individual deals or cohorts may behave
  differently from the expected for a variety of
  not easily identifiable reasons.
  
• AFT has created a solver for HT multiplier, Refi
  multiplier, and elbow shift to minimize the
  r-squared errors. The first few months of data a
  ignored (since they tend to be least reliable and
  most volatile)
• The solver is integrated within several
  analytical systems and within Regressor
  
• One can use the solver against CMO deals or
  against any cohort
  
• One needs to be careful that the modifications
  are predictive
  

75
Deal-Specific Long Term Adjustments
Historical fit without long term adjustments
76
Deal-Specific Long Term Adjustments
Historical fit with long term adjustments
77
Analyzing a models performance

• The primary consideration in evaluating a models
  performance is its stability. Frequent releases
  indicate an attempt to predict the last few
  months of prepayment speeds, which are well
  known, leaving no room for prolonged
  out-of-sample observations.
• When comparing historical model performance,
  ALWAYS compare models of the same vintage, e.g.
  if you are looking at 2004 performance, you
  cannot compare a 2001 vintage model to 2004
  vintage.

78
Analyzing a models performance

• Given two models of similar vintage, one can
  compare their OUT-OF-SAMPLE performances using a
  variety of statistical tools.
  
• Comparisons of in sample model performances are
  of less value. They can serve as indicators (and
  indicators only) of the extent of the model
  flexibility in reflecting accurately and
  completely the underlying phenomenon.
• Fit to history in sample, if analyzed blindly,
  may give no information as to the model potential
  for future performance.

79
Analyzing a models performance

• The analysis still has to involve a deep
  understanding of model structure in order to
  attain a degree of confidence that it reflects
  the phenomenon. The in sample performance has
  to be a result of that structure, rather than of
  a blind collection of parameters that happen to
  fit well but have no predictive power.

80
What rates should drive prepayments?

• Mortgage prepayments can only be driven by rates
  that a borrower sees what happens in the
  secondary markets is of no consequence to the
  borrower.
• A prepayment model may only connect primary, not
  secondary, rates to prepayments - otherwise you
  need a new model each time primary/secondary
  spreads change.

81
Primary/secondary spreads modeling

• Most analysis is based on rates available in the
  secondary market. These rates change minute by
  minute
  
• Primary rates are available on a weekly basis at
  best
  
• One needs to make an assumption about the effect
  on the primary rates from the secondary market
  
• The common assumption is to assume a constant
  spread
  
• The assumption breaks down when the production
  volume hits capacity constraints
  

82
Primary/secondary spreads modeling

• AFT prepayment model can accept projected primary
  mortgage rates, or current coupon rates, or
  treasury rates, or LIBOR rates as inputs. Flags
  in espparamconfig.def tell the model which it is.
  
• Normally, unless a primary mortgage rate is sent,
  the model adds a constant spread to the rate. The
  spreads are specified in either
  mrprimccsprds.dat, or mrprimcomsprds.dat, or
  mrtrsprds.dat depending on the control flags
• Another method of calculating primary mortgage
  rates is to use a spread model based on these
  rates.
  
• The spread model may be applied externally by the
  system vendor, or internally through our
  prepayment library directly by using the above
  flags.

83
Primary/secondary spreads modeling

• There are three spread models. A separate
  algorithm and a separate parameter set is used to
  calculate primary mortgage rates based on each of
  the following Current Coupon rates, LIBOR, FNMA
  10-day commitment rates.
• In addition to the above, one can define manual
  spread adjustments. The files that control that
  are described in AFT Model Technical Structure
  III, Spreads Control Files

84
New AFT Non-Agency Database

• AFT has been collecting data for loans backing
  all non-agency CMOs from a variety of sources
  
• Currently have collected loan-level data for
  about 80 of non-agency universe CMOs.
  
• The data is processed into Dynamic Aggregator, or
  will be soon available through McDash or in
  raw format
  
• AFT uses McDash agency database to complete its
  data. AFT has data for about 70 of the loans in
  existence
  

85
Model fitted to customer data setuncorrected
86
Model fitted to customer data setcorrected
87
AFT Vendor Integration