Modelling of successive default events Nicole El Karoui, CMAP, Ecole Polytechnique Monique Jeanblanc, Universit´ e d’ ´ Evry, France Ying Jiao, Ecole Sup´ erieure d’Ing´ enieur L´ eonard de Vinci Princeton, Credit Risk, May 2008 Financial support from Fondation du Risque and F´ ed´ eration des Banques Fran¸caises 1
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Modelling of successive default events
Nicole El Karoui, CMAP, Ecole PolytechniqueMonique Jeanblanc, Universite d’Evry, FranceYing Jiao, Ecole Superieure d’Ingenieur Leonard de Vinci
Princeton, Credit Risk, May 2008
Financial support from Fondation du Risque and Federation des Banques Francaises 1
The aim of this talk is
• to give a general framework for multi-defaults modelling
• to obtain the dynamics of derivative products in a multinamesetting.
2
The basic tool is the conditional law of the default(s) with respect to areference filtration
3
Single Name
Single Name
4
Single Name
Notation
• G is the global market filtration,• τ is a default time,• Ht = 11τ≤t is the default processes,• H is the natural filtration of H, with H ⊂ G,• F is a reference filtration, with F ⊂ G .
5
Single Name
On the set {τ > t}: ”before-default”
(F, G, τ) satisfy the minimal assumption (MA) if∀ t ≥ 0 and ZG ∈ Gt, ∃ ZF ∈ Ft such that
ZG ∩ {τ > t} = ZF ∩ {τ > t}.
• If Gτ := F ∨ H, then (F, Gτ , τ) enjoys MA.
• Under MA, for any G∞-measurable (integrable) r.v. Y G,
11{τ>t}E[Y G|Gt] = 11{τ>t}E[Y G11{τ>t}|Ft]
P(τ > t|Ft)a.s.
on the set A := {ω : P(τ > t|Ft)(ω) > 0}.
6
Single Name
On the set {τ ≤ t}: ”after-default”
1. We assume that G = Gτ = F ∨ H
2. J-Hypothesis (Jacod for enlargement of filtration purpose):
We assume that there exists a family of Ft ⊗ B(R+) r.vsαt(θ) such that
P(τ ∈ dθ|Ft) = αt(θ)dθ dθ ⊗ dP − a.s.
and for any θ the process αt(θ), t ≥ 0 is a right-continuousmartingale
Under the above hypotheses, we can compute Gt-conditionalexpectations on the set {τ ≤ t}A ”weak version” of J-hypothesis consists of the existence of thedensity only for 0 ≤ t ≤ θ. This is useful for before-default studies, butnot sufficient for the after-default ones.
7
Single Name
Density and F-martingales
We assume J-hypothesis and introduce the cadlag (super) martingales
• Conditional survival process (St = P(τ > t|Ft), t ≥ 0) (Azemasuper-martingale)
• Conditional probability process (St(θ) = P(τ > θ|Ft), t ≥ 0).Note that, for t ≤ θ, one has St(θ) = E[Sθ|Ft],
• Family of martingales (for θ ∈ R): the densities (αt(θ), t ≥ 0) ofSt(θ)
8
Single Name
Decompositions of Survival process S (super-martingale)
• The Doob-Meyer decomposition of S is St = MFt − AF
t with
– the F-martingale
MFt = −
∫ t
0
(αt(u) − αu(u))du = St − S0 +∫ t
0
αu(u)du
– the increasing process AFt =
∫ t
0αu(u) du
• The multiplicative decomposition isSt = LF
t DFt
where
– The F-martingale LF is given as dLFt = e
∫ t0 λF
sdsdMFt
– The decreasing process DF is DFt = exp
(− ∫ t
0λF
sds)
where
λFt = αt(t)
Ston {St > 0}.
9
Single Name
Links with the classical intensity approach
• The G-intensity is the G-adapted process λG such that(11{τ≤t} −
∫ t
0λG
s ds, t ≥ 0) is a G-martingale.
• Under the weak version of J-Hypothesis,
λGt = 11{τ>t}λF
t = 11{τ>t}αt(t)St−
a.s..
• For any θ ≥ t,αt(θ) = E[λG
θ |Ft] a.s..
Note that the intensity approach does not contain enough informationto study the after-default case (i.e. for θ < t).
10
Single Name
A characterization of G-martingales
Any Gt-measurable r.v. X can be written as
X = Xt11{τ>t} + Xt(τ)11{τ≤t}where Xt and Xt(θ) are Ft-measurable.
The process MX is a G-martingale if its decomposition as
MXt := Xt11{τ>t} + Xt(τ)11{τ≤t}
satisfies
• (XtSt +∫ t
0Xs(s)αs(s) ds, t ≥ 0) is an F-martingale
• For any θ, (Xt(θ)αt(θ), t ≥ θ) is an F-martingale
Remark: The first condition is equivalent to :XtL
Ft − ∫ t
0(Xs − Xs(s))LF
sλFs ds is an F-martingale
11
Single Name
Immersion hypothesis
Immersion holds if (and only if) any F-martingale is aG-martingale.
Under immersion hypothesis,
• αt(θ) = αt∧θ(θ)
• S is a non-increasing process
• LF is a constant
• the processMX
t := Xt11{τ>t} + Xt(τ)11{τ≤t}is a G-martingale if
(a) Xt(θ) is a F-martingale on [θ,∞).
(b) Xt −∫ t
0(Xs − Xs(s))λF
s ds is an F-martingale
12
Single Name
Toy Exemple: Cox process model
Let τ = inf{t : Λt :=∫ t
0λF
sds ≥ Θ} whereΛ is an F-adapted increasing process, Λ0 = 0, limt→∞ Λt = +∞Θ is a G-measurable r.v. independent of F∞, Θi ∼ exp(1).F is immersed in G = F ∨ H.The conditional distribution of τ is⎧⎨⎩ P(τ > θ| Ft) = E[e−Λθ | Ft], for θ > t
P(τ > θ| Ft) = e−Λθ , for θ ≤ t
and the density is⎧⎨⎩ αt(θ) = E[λθe−Λθ | Ft], for θ > t
= λθe−Λθ , for θ ≤ t
13
Single Name
Modelling the density process
Two possible solutions:
• Model St(θ) := P(τ > θ|Ft) and then take derivatives w.r.t. θ
• Model the density αt(θ) as a family of strictly positive martingalessuch that
∫∞0
αs(θ)dθ = 1
Remarks :
• for fixed θ, both processes are positive F martingales
• reference to the interest models
• distinction between θ ≥ t (classical part) and θ < t (non-classicalpart)
14
Single Name
F-martingale representation and HJM framework
Model the family of F-martingales St(θ) in the HJM framework
SupposedSt(θ)St(θ)
= Φt(θ)dMt, t, θ ≥ 0
where M is a continuous multi-dimensional F-martingale, thenSt(θ) = St(t) exp
(− ∫ θ
tλt(u)du
)where λt(θ) is the forward intensity
and
* St(θ) = S0(θ) exp(∫ t
0Φs(θ)dMs − 1
2
∫ t
0|Φs(θ)|2d〈M〉s
);
* St = exp(− ∫ t
0λF
sds +∫ t
0Φs(s)dMs − 1
2
∫ t
0|Φs(s)|2d〈M〉s
).
* λt(θ) = λ0(θ) −∫ t
0ϕs(θ)dMs +
∫ t
0ϕs(θ)Φs(θ)d〈M〉s.
15
Single Name
For any θ ≥ 0, assume that λ0(θ) is a family of positive probabilitydensities.
Let b(θ) be a given family of non-negative F-adapted processes. Define
ϕt(θ) = −bt(θ)λ0(θ) exp(∫ t
0
bs(θ)dWs − 12
∫ t
0
bs(θ)2ds
)and let
αt(θ) = λt(θ) exp(−∫ t
0
λt(v)dv
)where λt(θ) = λ0(θ) −
∫ θ
0ϕs(θ)dWs +
∫ t
0ϕs(θ)Φs(θ)ds.
Then the family (αt(θ), t ≥ 0) is a density process.
16
Single Name
Examples of martingale density process
• Compatibility between martingale and probability properties
– the r.v. St(θ) is [0, 1]-valued
– for any t, the map θ → St(θ) is non-increasing
• A Generalized exponential model: ∀ t, θ ≥ 0, let
St(θ) = exp(− θMt − 1
2θ2 〈M〉t
)where M is an F-martingale.
• Exponential law S0(θ) = P(τ > θ) = exp(−θM0).
• Probability condition : Mt + 12θ 〈M〉t ≥ 0
17
Single Name
Comparison with interest rate modelling
• Zero-coupon B(t, T ) = E[e−∫ T
trsds|Ft].
Short rate rt = −∂T |T=t log B(t, T ).
• Defaultable zero-coupon without actualization
E[11{τ>T}|Gt] = 11{τ>t}E[ST /St|Ft] := 11{τ>t}Bτ (t, T ).
Intensity λFt = −∂T |T=t log Bτ (t, T ).
18
Single Name
A general construction of density process. I.
1. Start with P0 with immersion hypothesis, and τ with densityprocess (α0
t (θ), t ≥ 0) constant in time after θ
2. Let ZGt = Zt11{τ>t} + Zt(τ)11{τ≤t} a positive (G, P0)-martingale
with expectation 1
3. Define dP = ZGT dP0 on GT . The RN density of P w.r.t. P0 on Ft is
ZFt = ZtSt +
∫ t
0Zs(s)α0
s(s)ds.
4. Then the density process of τ under P is
αPt (θ) = α0
θ(θ)Zt(θ)ZF
t
for θ < t
αPt (θ) =
E[Zθ(θ)α0θ(θ)|Ft]
ZFt
for θ ≥ t.
19
Single Name
A general construction of density process. II.
1. Start with P0 with τ independent of F∞ with density f
2. Let q∞(u) a family of F∞-measurable r.v. such that∫∞0
q∞(u)f(u)du = 1
3. Define dP = q∞(τ)dP0 on G∞.
4. Then, setting qt(u) = E0(q∞(u)|Ft), the RN density of P w.r.t. P0