risk metrics hedging cdt
TRANSCRIPT
Risk Management
Hedging credit index tranchesInvestigating versions of the standard model
Christopher C. Finger
2www.riskmetrics.com Risk Management
Subtle company introduction
2
2www.riskmetrics.com Risk Management
Motivation
A standard model for credit index tranches exists.
It is commonly acknowledged that the common model isflawed.
Most of the focus is on the static flaw: the failure to calibrateto all tranches on a single day with a single model parameter.
But these are liquid derivatives. Models are not used forabsolute pricing, but for relative value and hedging.
We will focus on the dynamic flaws of the model.
3
2www.riskmetrics.com Risk Management
Outline
1 Standard credit derivative products
2 Standard models, conventions and abuses
3 Data and calibration
4 Testing hedging strategies
5 Conclusions
4
2www.riskmetrics.com Risk Management
Standard products
Single-name credit default swaps
Contract written on a set of reference obligations issued by onefirmProtection seller compensates for losses (par less recovery) inthe event of a default.Protection buyer pays a periodic premium (spread) on thenotional amount being protected.Quoting is on fair spread, that is, spread that makes a contracthave zero upfront value at inception.
5
2www.riskmetrics.com Risk Management
Standard products
Credit default swap indices (CDX, iTraxx)
Contract is essentially a portfolio of (125, for our purposes)equally weighted CDS on a standard basket of firms.Protection seller compensates for losses (par less recovery) inthe event of a default.Protection buyer pays a periodic premium (spread) on theremaining notional amount being protected.New contracts (series) are introduced every six months.Standardization of premium, basket, maturity has createdsignificant liquidity.Quoting is on fair spread, with somewhat of a twist.Pricing depends only on the prices of the portfolio names. . . almost.
6
2www.riskmetrics.com Risk Management
Standard products
Tranches on CDX
Protection seller compensates for losses on the index in excessof one level (the attachment point) and up to a second level(the detachment point).For example, on the 3-7% tranche of the CDX, protection sellerpays losses over 3% (attachment) and up to 7% (detachment).Protection buyer pays an upfront amount (for most juniortranches) plus a periodic premium on the remaining amountbeing protected.Standardization of attachment/detachment, indices, maturity.Not strictly a derivative on the index, in that payoff does notreference the index pricePricing depends on the distribution of losses on the index, notjust the expectation.
Also, options on CDX, but we will not consider these.7
2www.riskmetrics.com Risk Management
CDX history
Mar05 Sep05 Mar06 Sep06 Mar07 Sep07 Mar080
40
80
120
160
200In
dex
spre
ad (
bp)
8
2www.riskmetrics.com Risk Management
CDX tranche history
Mar05 Sep05 Mar06 Sep06 Mar07 Sep070
10
20
30
40
50
60
Tra
nche
upf
ront
pric
e (%
)
0
100
200
300
400
500
Tra
nche
fair
spre
ad (
bp)
0−3% (LHS)3−7% (RHS)7−10% (RHS)
9
2www.riskmetrics.com Risk Management
What do we want from a model?
Fit market prices, but why?
Extrapolate, i.e. price more complex, but similar, structures
Non-standard attachment pointsCustomized baskets
Hedge risk due to underlying
Dealers provide liquidity in tranches, but want to controlexposure to underlyings.Speculators want to make relative bets on tranches without aview on underlyings.
Risk management
Aggregate credit exposures across many product types.Recognize risk that is truly idiosyncratic.
For anything other than extrapolation, we care about how pricesevolve in time, so we should look at the dynamics.
10
2www.riskmetrics.com Risk Management
Standard pricing models—basic stuff
Price for a tranche is the difference of
Expectation of discounted future premium payments, andExpectation of discounted future losses.
Boils down to the distribution of the loss process on the indexportfolio, in particular things like
E min{(d − a),max{0, Lt − a}}.
Suffices to specify the joint distribution of times to default Ti
for all names in the basket.
CDS (or CDX) quotes imply the marginal distributions fortime to default for individual names Fi (t) = P{Ti < t}.
11
2www.riskmetrics.com Risk Management
Standard pricing models—specific stuff
Dependence structure is a Gaussian copula:
Let Zi be correlated standard Gaussian random variables.Default times are given by Ti = F−1
i (Φ(Zi )).
Correlation structure is pairwise constant . . .
Zi =√ρZ +
√1− ρ εi .
For a single period, just count the number of Zi that fallbelow the default threshold αi = Φ−1(pi ).
12
2www.riskmetrics.com Risk Management
Criticisms of the standard model
Does not fit all market tranche prices on a given day.
No dynamics, so no natural hedging strategy.
Link to any observable correlation is tenuous. At best, modelis “inspired” by Merton framework, so correlation is onequities.
13
2www.riskmetrics.com Risk Management
Model flavors
Most numerical techniques rely on integrating over Z , given whichall defaults are conditionally independent.
Granular model—use full information on underlying spreads,and model full discrete loss distribution.
Homogeneity assumption—assume all names in the portfoliohave the same spread; use index level.
Fine grained limit—continuous distribution, easy integrals
Large pool model—combine homogeneity and fine grainedassumptions.
Also the question of whether to use full spread term structuresor a single point
14
2www.riskmetrics.com Risk Management
Correlation conventions
Start to introduce model abuses, ala the B-S volatility smile.
Compound correlations
Price each individual tranche with a distinct correlation.Not all tranches are monotonic in correlation.Trouble calibrating mezzanine tranches, especially in 2005
Base correlations
Decompose each tranche into “base” (i.e. 0-x%) tranches.Bootstrap to calibrate all tranches.Base tranches monotonic, but calibration not guaranteed.
15
2www.riskmetrics.com Risk Management
Correlation conventions
Constant moneyness (ATM) correlations
Some movements in implied correlation are due to changing“moneyness” as the index changes.Examine correlations associated with a detachment point equalto implied index expected loss.If base correlations are “sticky strike”, then ATM correlationsare closer to “sticky delta”.
16
2www.riskmetrics.com Risk Management
CDX correlation structure
0 5 10 15 20 25 300
10
20
30
40
50
60
70
80
90
100
Detachment point (%)
Bas
e co
rrel
atio
n (%
)
Mar05Mar06Mar07Mar08
17
2www.riskmetrics.com Risk Management
CDX base correlations, granular model
Sep05 Mar06 Sep06 Mar07 Sep070
20
40
60
80
100B
ase
corr
elat
ion
(%)
18
2www.riskmetrics.com Risk Management
CDX base correlations, large pool model
Mar05 Sep05 Mar06 Sep06 Mar07 Sep070
20
40
60
80
100B
ase
corr
elat
ion
(%)
19
2www.riskmetrics.com Risk Management
CDX base correlations, large pool model, with DJXoption-implied correlation
Mar05 Sep05 Mar06 Sep06 Mar07 Sep070
20
40
60
80
100B
ase
corr
elat
ion
(%)
20
2www.riskmetrics.com Risk Management
CDX Series 4-7, GR model
4 6 8 10 12 14 16 180
10
20
30
40
50
60
70
Detachment plus index EL (%)
Bas
e co
rrel
atio
n (%
)
21
2www.riskmetrics.com Risk Management
CDX Series 4-7, LP model
4 6 8 10 12 14 16 180
10
20
30
40
50
60
70
Detachment plus index EL (%)
Bas
e co
rrel
atio
n (%
)
22
2www.riskmetrics.com Risk Management
CDX Series 8-9, GR model
4 6 8 10 12 14 16 180
10
20
30
40
50
60
70
Detachment plus index EL (%)
Bas
e co
rrel
atio
n (%
)
23
2www.riskmetrics.com Risk Management
CDX Series 8-9, LP model
4 6 8 10 12 14 16 180
10
20
30
40
50
60
70
Detachment plus index EL (%)
Bas
e co
rrel
atio
n (%
)
24
2www.riskmetrics.com Risk Management
Time series properties
Examine the correlation between various implied correlationsand the index.
We would like this to be low. Why?
For risk, we have identified idiosyncratic risk correctly.
For hedging, we have captured what we are able to from theunderlying.
Start with statistics on daily changes.
25
2www.riskmetrics.com Risk Management
CDX 0-3%, large pool model, base correlations
4 5 6 7 8 9 All0
10
20
30
40
50
60
Series
Cor
rela
tion
to in
dex
(%)
Flat curveFull curve
26
2www.riskmetrics.com Risk Management
CDX 0-3%, base correlations
4 5 6 7 8 9 All−20
−10
0
10
20
30
40
50
60
Series
Cor
rela
tion
to in
dex
(%)
Large poolGranular
27
2www.riskmetrics.com Risk Management
CDX 0-3%, large pool model
4 5 6 7 8 9 All−60
−40
−20
0
20
40
60
Series
Cor
rela
tion
to in
dex
(%)
Base CorrATM Corr
28
2www.riskmetrics.com Risk Management
CDX 0-3%
4 5 6 7 8 9 All−30
−20
−10
0
10
20
30
40
50
60
Series
Cor
rela
tion
to in
dex
(%)
LP, baseGran, ATM
29
2www.riskmetrics.com Risk Management
CDX 3-7%
4 5 6 7 8 9 All−20
−10
0
10
20
30
40
50
60
Series
Cor
rela
tion
to in
dex
(%)
LP, baseGran, ATM
30
2www.riskmetrics.com Risk Management
Time series properties
Now look at statistics on weekly changes.
31
2www.riskmetrics.com Risk Management
CDX 0-3%, large pool model, base correlations
4 5 6 7 8 9 All−60
−40
−20
0
20
40
60
80
Series
Cor
rela
tion
to in
dex
(%)
Flat curveFull curve
32
2www.riskmetrics.com Risk Management
CDX 0-3%, base correlations
4 5 6 7 8 9 All−60
−40
−20
0
20
40
60
80
Series
Cor
rela
tion
to in
dex
(%)
Large poolGranular
33
2www.riskmetrics.com Risk Management
CDX 0-3%, large pool model
4 5 6 7 8 9 All−80
−60
−40
−20
0
20
40
60
80
Series
Cor
rela
tion
to in
dex
(%)
Base CorrATM Corr
34
2www.riskmetrics.com Risk Management
CDX 0-3%
4 5 6 7 8 9 All−80
−60
−40
−20
0
20
40
60
80
Series
Cor
rela
tion
to in
dex
(%)
LP, baseGran, ATM
35
2www.riskmetrics.com Risk Management
CDX 3-7%
4 5 6 7 8 9 All−60
−40
−20
0
20
40
60
80
Series
Cor
rela
tion
to in
dex
(%)
LP, baseGran, ATM
36
2www.riskmetrics.com Risk Management
Time series conclusions
Looks like granular model produces “more idiosyncratic”implied correlations.
The ATM approach looks appears to improve things, butseems to overcorrect in the most volatile periods.
High correlations overall suggest index moves may have somepredictive power for implied correlations.
Differences are much more pronounced at a one-day horizonthan a one-week horizon.
37
2www.riskmetrics.com Risk Management
Hedging experiment
On day zero, we have all prices (spreads and tranches) plushistory.
At a future date, we are told how the underlying spreads(index or single-name CDS) have moved, and are asked toguess what the new tranche prices should be.
Compare the predicted tranche price moves to the actualones, over time and across different modeling approaches.
Another approach would be to look at delta hedgeperformance, but this mixes in the error from linearization.
38
2www.riskmetrics.com Risk Management
Hedging experiment candidates
Models
Regression—fit linear relationship between tranche pricechanges and index changes, using prior six months of data.Large pool model—calibrate based on current index levels.Granular model—calibrate based on current single-name CDSlevels.
Correlation approaches
Base—use most recent calibrated correlation.ATM—move along the most recent correlation structureaccording to change in the index-implied expected loss.Regression—fit linear relationship between base correlationchanges and index changes, using prior six months of data.
39
2www.riskmetrics.com Risk Management
CDX 0-3%, Regression
Sep05 Mar06 Sep06 Mar07 Sep070
10
20
30
40
50
60
70
80
Tra
nche
pric
e (%
)
−8
−4
0
4
8
Pre
dict
ion
erro
r (%
)
40
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model, base correlations
Sep05 Mar06 Sep06 Mar07 Sep070
10
20
30
40
50
60
70
80
Tra
nche
pric
e (%
)
−8
−4
0
4
8
Pre
dict
ion
erro
r (%
)
41
2www.riskmetrics.com Risk Management
CDX 0-3%, Regression
−8 −6 −4 −2 0 2 4 6 8−8
−6
−4
−2
0
2
4
6
8
Tranche price change (%)
Pre
dict
ed c
hang
e (%
)
Ser 4Ser 5Ser 6Ser 7Ser 8Ser 9
42
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model, base correlations
−8 −6 −4 −2 0 2 4 6 8−8
−6
−4
−2
0
2
4
6
8
Tranche price change (%)
Pre
dict
ed c
hang
e (%
)
Ser 4Ser 5Ser 6Ser 7Ser 8Ser 9
43
2www.riskmetrics.com Risk Management
Statistics on hedging approaches
Examine standard deviation of tranche forecast error.
Daily horizon
Bars are for correlation approaches: Base (blue), ATM(green), Regression (red).
Curve is for simple regression model.
44
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model
4 5 6 7 8 9 All0
0.005
0.01
0.015
0.02
0.025
0.03S
TD
of p
redi
ctio
n er
ror
Series
45
2www.riskmetrics.com Risk Management
CDX 0-3%, GR model
4 5 6 7 8 9 All0
0.005
0.01
0.015
0.02
0.025
0.03S
TD
of p
redi
ctio
n er
ror
Series
46
2www.riskmetrics.com Risk Management
CDX 3-7%, LP model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series47
2www.riskmetrics.com Risk Management
CDX 3-7%, GR model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series48
2www.riskmetrics.com Risk Management
CDX 7-10%, LP model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series49
2www.riskmetrics.com Risk Management
CDX 7-10%, GR model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series50
2www.riskmetrics.com Risk Management
Statistics on hedging approaches
Examine correlation of tranche forecast error with actualtranche move.
Daily horizon
Bars are for correlation approaches: Base (blue), ATM(green), Regression (red).
Curve is for simple regression model.
51
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
52
2www.riskmetrics.com Risk Management
CDX 0-3%, GR model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
53
2www.riskmetrics.com Risk Management
CDX 3-7%, LP model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
54
2www.riskmetrics.com Risk Management
CDX 3-7%, GR model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
55
2www.riskmetrics.com Risk Management
CDX 7-10%, LP model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
56
2www.riskmetrics.com Risk Management
CDX 7-10%, GR model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
57
2www.riskmetrics.com Risk Management
Statistics on hedging approaches
Examine standard deviation of tranche forecast error.
Weekly horizon
Bars are for correlation approaches: Base (blue), ATM(green), Regression (red).
Curve is for simple regression model.
58
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model
4 5 6 7 8 9 All0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045S
TD
of p
redi
ctio
n er
ror
Series
59
2www.riskmetrics.com Risk Management
CDX 0-3%, GR model
4 5 6 7 8 9 All0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045S
TD
of p
redi
ctio
n er
ror
Series
60
2www.riskmetrics.com Risk Management
CDX 3-7%, LP model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series61
2www.riskmetrics.com Risk Management
CDX 3-7%, GR model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series62
2www.riskmetrics.com Risk Management
CDX 7-10%, LP model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series63
2www.riskmetrics.com Risk Management
CDX 7-10%, GR model
4 5 6 7 8 9 All0
0.5
1
1.5
2
2.5
3
3.5x 10
−3S
TD
of p
redi
ctio
n er
ror
Series64
2www.riskmetrics.com Risk Management
Statistics on hedging approaches
Examine correlation of tranche forecast error with actualtranche move.
Weekly horizon
Bars are for correlation approaches: Base (blue), ATM(green), Regression (red).
Curve is for simple regression model.
65
2www.riskmetrics.com Risk Management
CDX 0-3%, LP model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
66
2www.riskmetrics.com Risk Management
CDX 0-3%, GR model
4 5 6 7 8 9 All−1
−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0C
orr(
Act
,Err
)
Series
67
2www.riskmetrics.com Risk Management
CDX 3-7%, LP model
4 5 6 7 8 9 All−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1C
orr(
Act
,Err
)
Series
68
2www.riskmetrics.com Risk Management
CDX 3-7%, GR model
4 5 6 7 8 9 All−0.9
−0.8
−0.7
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1C
orr(
Act
,Err
)
Series
69
2www.riskmetrics.com Risk Management
CDX 7-10%, LP model
4 5 6 7 8 9 All−1
−0.5
0
0.5C
orr(
Act
,Err
)
Series
70
2www.riskmetrics.com Risk Management
CDX 7-10%, GR model
4 5 6 7 8 9 All−1
−0.5
0
0.5C
orr(
Act
,Err
)
Series
71
2www.riskmetrics.com Risk Management
Conclusions
Overall, hedging errors are surprisingly large.
For the equity tranche, a simple regression works quite well,with its underhedging problems reduced at a slightly longerhorizon.
There is a benefit to using the granular model, but is it worththe cost?
For more senior tranches, the ATM approach appears tocapture some of the link to credit spreads, but does notmarkedly reduce the hedging error.
Any candidate for a new standard model should be able to dobetter in this experiment.
72
2www.riskmetrics.com Risk Management
Further investigations
Can we use these results to learn more about how our modelmight be misspecified?
Richer correlation structure?Different copula?Better dynamics?
How much worse (or better) are compound correlations?
Can the regression be improved by including the HiVol index?
Can the LP model be improved by adding a simple correctionfor heterogeneity?
Does any of this change at or close to defaults?
73
2www.riskmetrics.com Risk Management
Further reading
Couderc, F. (2007) Measuring Risk on Credit Indices: On theUse of the Basis. RiskMetrics Journal.
Finger, C. (2004) Issues in the Pricing of Synthetic CDOs.RiskMetrics Journal.
Finger, C. (2005) Eating our Own Cooking. RiskMetricsResearch Monthly. June.
All are available at www.riskmetrics.com
74