Using AFT and FSI Tools to Validate Components of Interest ...
Transcript of Using AFT and FSI Tools to Validate Components of Interest ...
ALCO Partners, LLC Page 1
Using AFT and FSI Tools to Validate Components of Interest Rate Risk Models
ALCO Partners, LLC Page 2
Agenda
Validation Tests of Prepayment Models • Models applicable to securities
• Models applicable to portfolio loans
Validation Tests of Stochastic Rate Generators using FSI
• Lattice
• Monte Carlo
ALCO Partners, LLC Page 3
Regressor and Back-testing
Regressor is applicable to AFT prepayment models:
• Securities
o MBS
o CMOs
MBS data at the CUSIP level require extraction routines from Bloomberg or Intex.
CMO data at the CUSIP level require license with Intex, Chasen or other cash flow model vendor.
Agency CMO data can be extracted from Bloomberg.
ALCO Partners, LLC Page 4
Batch “Back-test” in the Regressor Model
• Output created:
o Actual CPR
o AFT forecast of CPR using actual rates
• Output is written to text files with number of files equal to the number of collateral types.
o Example: 300 CUSIPs of 15 and 30 FNMA, FHMLC, and Whole loans might reside in six text files
• All CUSIPs can be run at once in Regressor with the text files created automatically.
• Text files can be opened in EXCEL and standard “goodness of fit” or other statistics applied.
Regressor and Back-testing
ALCO Partners, LLC Page 5
Once a macro is written, process takes less than an hour for the entire list of CUSIPs (We did this for 300 CUSIPs.)
Procedure can be learned in about 15 minutes via phone support, if all required extract files are available.
Regressor allows the aggregation across CUSIPs drawn from same collateral. Performance on sub-portfolios can be assessed.
Regressor and Back-testing
ALCO Partners, LLC Page 6
Why can AFT Claim Regressor Provides a Back-test”?
AFT Prepayment models are estimated using collateral data
• Securities are small samples of the collateral.
• Output in Regressor at CUSIP level is based on both overlapping historical periods and projected periods, depending on the length of history and when collateral model was estimated.
o Over the historical data model period this is an “in-sample” back-test.
o When CPRs are projected beyond historical data period used for building models, the test is an “out-of-sample” back-test.
If the connectivity in the ALM system and AFT models has been validated, then the ALM model is not needed for back-testing AFT prepayment models.
ALCO Partners, LLC Page 7
Back testing Sample Output from Regressor
0
10
20
30
40
50
60
70
80
90
Nov-0
2
Jan-0
3
Mar
-03
May
-03
Jul-
03
Sep
-03
Nov-0
3
Jan-0
4
Mar
-04
May
-04
Jul-
04
Sep
-04
CP
R (
%)
Actual Predicted
55265KTS8
R2=73% 80
One of 180 graphs created for an ALCO Partners client
ALCO Partners, LLC Page 8
Dynamic Aggregator and Back testing
Dynamic Aggregator is applicable to portfolio loans
• Similar output created as in Regressor
o AFT prepayment models provide the basis for modeling. Discrete prepayment factors can be scaled to improve model fit.
o Historical data used to fit model and model statistics can be created as part of the estimation process.
o In-sample and out-of-sample model testing can be applied with similar statistics.
o If AFT-ALM model link has been validated, then the process can substitute for back test using ALM model.
ALCO Partners, LLC Page 9
Findings from Dynamic Aggregator Projects
Dynamic Aggregator has coincidentally “solved” a classic data bottleneck in many banking institutions
• IT departments now regularly archive historical data relevant for testing prepayment models
• They are very efficient at providing raw data.
• They are inefficient at “mapping” data
o Internal resources for “research and analysis” are frequently low on a priority list
o External projects where only raw data is required appear not to face the same internal hurdles because they demand very few IT resources
ALCO Partners, LLC Page 10
Findings from Dynamic Aggregator Projects
A part of prepayment model validation projects applicable to portfolio mortgages is obtaining and mapping data. But this process is no longer a large cost.
o Mapping costs have not been large. Key variable is number of systems
o Mapping rules typically remain constant over time, so annual updates become very cost effective
Once the data has been mapped into Dynamic Aggregator, the process for testing and is very cost-effective, particularly when the process is done annually as required by banking regulations
ALCO Partners, LLC Page 11
Testing Stochastic Rate Generators
In many banks stochastic rate generators are being utilized to measure value based IRR
• Base case is calibrated to market instruments
• There are known nuances and problems associated with this
• How does a bank “validate” a stochastic rate model?
ALCO Partners, LLC Page 12
Testing Stochastic Rate Generators
Approach I
• Obtain model documentation
• Build a separate model independently
• Perform double-blind tests and compare inputs and outputs
Issues with Approach I
• Vendors may be unwilling to provide proprietary code documentation
• Building a separate model can be very expensive, including its validation
• Specialized coding/math expertise required to “build” a stochastic rate generator
ALCO Partners, LLC Page 13
Testing Stochastic Rate Generators
Approach II
• Obtain a qualified stochastic rate generation system
• Calibrate to the same set of market instruments
• Compare inputs and outputs for tested samples
FSI: A qualified stochastic rate generation system
ALCO Partners, LLC Page 14
Testing Stochastic Rate Generators
FSI is a fully specified two factor lattice & Monte-Carlo system designed as a portfolio management product with validation capabilities
• Provides a basis for comfort or discomfort depending on the model results and how the model results are integrated with portfolio management decisions
• Since it contains both single factor (BK) and two-factor processes, it provides a basis for understanding the sensitivity of the risk measure to the underlying process utilized
• Helps improve calibration procedures and allows a look into the “black box” of stochastic rate generation capability
ALCO Partners, LLC Page 15
Testing Stochastic Rate Generators
Test Procedures
• Compared stochastic output from a widely used ALM model with a stochastic generator
o Compared three portfolios in two periods
o MBS & CMO
o Portfolio FRMs
o Callable Bonds
o Greeks are computed at the instrument level
o Market values and market values under rate shocks can be computed
o Market values changes compared over time at instrument level
o Results used to refine and understand calibration process and reasonableness of risk results
ALCO Partners, LLC Page 16
Callables: MV Sensitivity
-6%
-4%
-2%
0%
2%
4%
6%
8%
-150 -100 0 100 150 200
Shock (bp)
% D
M
V
FSI1F
FSI2F
ALM
Callables: MV Sensitivity Compared
-6%
-4%
-2%
0%
2%
4%
6%
8%
-150 -100 0 100 150 200
Shock (bp)
% D
M
V
FSI1F
FSI2F
ALM
Callables: MV Sensitivity Compared
12/31 3/30
Market value sensitivity measures for
callable bond portfolio where very similar
in both time periods
ALCO Partners, LLC Page 17
Callables: Duration & Convexity
Red line represents parity
0
1
2
3
4
5
6
7
8
0 1 2 3 4 5 6 7 8
Duration (FSI 2Factor)
Du
rati
on
(A
LM
)
Callables: FSI 2F vs. ALM Model
-2
0
2
4
6
8
-2 0 2 4 6 8
Convexity (FSI 2Factor) C
on
ve
xit
y (
AL
M)
Callables: FSI 2F vs. ALM Model
Duration Convexity
Results indicated some non-random differences on
individual securities. Further investigation revealed pattern
associated with “at the money” calls. Cross checks with
market values showed sensitivity to calibration mechanics
ALCO Partners, LLC Page 18
Callables: FSI 2 Factor vs. ALM Model: 12/31 to 3/31
Red line represents parity
-1%
0%
1%
2%
3%
4%
-1% 0% 1% 2% 3%
% D MV (FSI 2Factor)
% D
M
V (
AL
M)
Callables: 3/31 vs. 12/31, FSI vs. ALM Model
How does Market Value Change over Time?
ALCO Partners, LLC Page 19
FRMs: MV Sensitivity
-10%
-8%
-6%
-4%
-2%
0%
2%
4%
-150 -100 0 100 150 200
Shock (bp)
%D
M
V
FSI1F
FSI2F
ALM
FRMs: MV Sensitivity
-12%
-10%
-8%
-6%
-4%
-2%
0%
2%
4%
6%
-150 -100 0 100 150 200
Shock (bp)
%D
M
V
FSI1F
FSI2F
ALM
FRMs: MV Sensitivity
12/31 3/31
Prepayment models were identical
ALCO Partners, LLC Page 20
FRMs: Duration & Convexity
Red line represents parity
0
1
2
3
4
5
6
0 1 2 3 4 5 6
Duration (FSI 2Factor)
Du
rati
on
(A
LM
)
FRMs
-5
-4
-3
-2
-1
0
-5 -4 -3 -2 -1 0
Convexity (FSI 2Factor) C
on
vexit
y (
AL
M)
FRMs
Duration Convexity
ALCO Partners, LLC Page 21
FRMs: FSI (2 Factor) vs. ALM: 12/31 to 3/31
Red line represents parity
-3%
-2%
-1%
0%
-3% -2% -1% 0%
% D MV (FSI 2Factor)
% D
MV
(A
LM
) FRMs
FRMs where there are good market value indicators
combined with the OAS calculations lead to similar results
ALCO Partners, LLC Page 22
CMOs & MBS: MV Sensitivity
12/31 3/31
-8%
-6%
-4%
-2%
0%
2%
4%
-150 -100 0 100 150 200
Shock (bp)
M
V
FSI1F
FSI2F
ALM
CMOs &MBS: MV Sensitivity
Compared (12/31/2004)
-8%
-6%
-4%
-2%
0%
2%
4%
-150 -100 0 100 150 200
Shock (bp)
M
V
FSI1F
FSI2F
ALM
CMOs &MBS: MV Sensitivity
Compared (12/31/2004)
ALCO Partners, LLC Page 23
CMOs & MBS: Duration & Convexity
Red line represents parity
Duration Convexity
Circled outliers were later identified as input errors in the
ALM model
0
2
4
6
8
0 2 4 6 8
Duration (FSI 2Factor)
Du
rati
on
(A
LM
)
CMOs & MBS: Duration
-6
-5
-4
-3
-2
-1
0
1
-6 -5 -4 -3 -2 -1 0 1
Convexity (FSI 2Factor)
Co
nvexi
ty (
AL
M)
CMOs & MBS: Convexity
ALCO Partners, LLC Page 24
CMOs & MBS: FSI (2 Factor) vs. ALM: 12/31 to 3/31
Red line represents parity
Four outliers are the same CUSIPs as in the prior graph
-5%
-4%
-3%
-2%
-1%
0%
1%
-5% -4% -3% -2% -1% 0% 1%
%D MV (FSI 2Factor)
%D
MV
(A
LM
)
CMOs& MBS: 3/31 vs. 12/31
ALCO Partners, LLC Page 25
Yield Curve Distributions: Single Factor vs. Two Factor Stochastic Processes
Single Factor
Process:
Yield Curve
Distribution
Mean
Min
2.510th
25th
Median
75th
90th
97.5
Max
0%
5%
10%
15%
20%
25%
30%
0 24 48 72 96 120 144 168 192 216 240
Month
1M
LIB
OR
Rate
(%
)
MC Short Rate
Percentile
Single Factor 3-31
Mean
Min
2.5
10th
25th
Median75th90th97.5 Max
-6%
-5%
-4%
-3%
-2%
-1%
0%
1%
2%
3%
0 24 48 72 96 120 144 168 192 216 240
Month
Yie
ld C
urv
e (
%)
MC Yield Curve
Percentil
e
Single Factor 3-31
Mean
Min
2.5
10th
25th
Median
75th
90th
97.5
Max
0%
5%
10%
15%
20%
25%
0 24 48 72 96 120 144 168 192 216 240
Month
5Y
r L
IBO
R R
ate
(%
)
MC Long Rate
Percentile
Single Factor 3-31
ALCO Partners, LLC Page 26
Yield Curve Distributions: Single Factor vs. Two Factor Stochastic Processes
Two Factor
Process:
Yield Curve
Distribution
Mean
Min
2.510th
25th
Median
75th
90th
97.5
Max
0%
5%
10%
15%
20%
25%
30%
0 24 48 72 96 120 144 168 192 216 240
Month
1M
LIB
OR
Rate
(%
)
MC Short Rate
Percentil
e
Two Factor 3-31
Mean
Min
2.510th
25th
Median
75th
90th
97.5
Max
0%
5%
10%
15%
20%
25%
0 24 48 72 96 120 144 168 192 216 240
Month
5Y
r L
IBO
R R
ate
(%
)
MC Long Rate
Percentil
e
Two Factor 3-31-2005
Mean
Min
2.5
10th
25th
Median75th90th97.5 Max
-6%
-5%
-4%
-3%
-2%
-1%
0%
1%
2%
3%
0 24 48 72 96 120 144 168 192 216 240
Month
Yie
ld C
urv
e (
%)
MC Yield Curve
Percentil
e
Two Factor 3-31
ALCO Partners, LLC Page 27
Yield Curve Distributions: Single Factor vs. Two Factor Stochastic Processes
Two Factor
Process:
Yield Curve
Distribution
Mean
Min
2.5
10th
25th
Median75th90th97.5 Max
-6%
-5%
-4%
-3%
-2%
-1%
0%
1%
2%
3%
0 24 48 72 96 120 144 168 192 216 240
Month
Yie
ld C
urv
e (
%)
MC Yield Curve
Percentil
e
Two Factor 3-31
Mean
Min
2.5
10th
25th
Median75th90th97.5 Max
-6%
-5%
-4%
-3%
-2%
-1%
0%
1%
2%
3%
0 24 48 72 96 120 144 168 192 216 240
Month
Yie
ld C
urv
e (
%)
MC Yield Curve
Percentil
e
Single Factor 3-31
Single Factor
Process:
Yield Curve
Distribution
ALCO Partners, LLC Page 28
FRMs: MV Sensitivity: Single vs. 2 Factor
-10%
-8%
-6%
-4%
-2%
0%
2%
4%
-150 -100 0 100 150 200
Shock (bp)
%D
M
V
FSI1F
FSI2F
ALM
FRMs: MV Sensitivity
-12%
-10%
-8%
-6%
-4%
-2%
0%
2%
4%
6%
-150 -100 0 100 150 200
Shock (bp)
%D
M
V
FSI1F
FSI2F
ALM
FRMs: MV Sensitivity
12/31 3/31
Impact of process is limited in this
case even though distribution of yield
curves is noticeable
ALCO Partners, LLC Page 29
Observations from Test
Stochastic models will frequently produce different measures of risk at the instrument level for instruments with embedded options
• We believe – but it is difficult to confirm without access to source code – that this is largely due to differences in calibration procedures as well as differences in modeling methods.
• With larger portfolios that include instruments with a wide variety of option structures, the differences in results will tend to cancel each other out if the overall levels of implied volatility used in the calibrations are similar.
ALCO Partners, LLC Page 30
Observations from Test
Validation tests of stochastic models using a qualified stochastic model, such as FSI, can:
• Identify modeling errors, because large differences in results invite further investigation
• Enhance understanding of the “black box” of ALM model calibration procedures, when calibrations aren’t “all that close”
• Provide guidance for the product characteristics that are stochastic modeling technique sensitive
Validation tests of stochastic models lead to the following warning:
Beware of relying on a single stochastic model when measuring risk at the instrument level for instruments with embedded complex options