Transcript Slides
Credit Derivatives Modelling:
Background Intensity
Sebastien Hitier
The views expressed here are the authors own, and may not represent
those of BNP Paribas group or any of its affiliates.
31 March 2016
Contents
1.
Some context for credit quantitative modelling
2.
Introducing Background Intensity
3.
Five useful applications of background intensity modelling
31 March 2016
2
Some context for credit quantitative modelling
1.
Overview of the mathematical tools used, specificities of the problem
2.
Methodological differences between actuarial and derivatives pricing
3.
Historical perspective in credit quantitative modelling
31 March 2016
3
Overview of mathematical tools used
One Period Modelling
•
Default occurrence is a {0, 1} valued random variable.
•
Recovery upon default is a [0, 1] valued variable,
but little is know about its mean
n
•
Credit Portfolio loss is defined by :
•
For credit portfolio, the goal is to calculate the density of LT
•
Given the marginals are bernouli variables, joint density of loss is determined by
the default copula.
LT 1 Ri 1 i T
i 1
Multi Period Modelling
•
Default event is modelled as the first jump of a poisson process with intensity l
•
As default risk varies over time, intensity can be made stochastic, which leads to a
survival probability formula that looks like a rates discount factor:
P T exp lT
T
P T Gt E exp lu du Gt
t
•
Using stochastic intensity allows to leverage all of the rates modelling
implementations
31 March 2016
4
Specificities of credit
Empirical studies of credit events
•
Default events are rare occurrences in the investment grade world, little statistics
•
Large number of correlated variables
•
Possible regime changes
•
Less market depth and liquidity
Need a parsimonious model for credit correlation
Complexity of multi period modelling
•
Positivity on intensity is a must have, compared to interest rate models, the asset
values have jumps.
•
Just as rates model are made difficult by the discounting being stochastic, credit
models are difficult because of the future payoff is conditional on (stochastic)
survival
•
Difficulties raised by multi period models for multiple credit
•
Theoretical difficulties: What is a correct filtration setup for multiple credit?
Multi period modelling is relatively well known for single name, challenge is multiple
names
31 March 2016
5
Differences between actuarial and derivatives pricing
Actuarial pricing
•
based on historical default realisation
•
parameters are estimated, and realism is more important than parsimony
The goal is to compute an “expectation”. Single period modeling is not an issue.
Derivatives pricing
•
Derivative pricing entail replication by simpler instruments
•
Calculating price as a “risk neutral” expectation implies
Multi period modelling is relatively well known for single name, challenge is multiple
names
31 March 2016
6
Historical perspective in credit quantitative modelling
Initial modeling:
•
Intensity models (used wherever we need a random arrival time)
•
Structural models introduced in the 70s (Merton model) when stochastic calculus
was introduced in finance
•
Tools from actuarial finance: copula (used to model joint accidents)
Progress prompted by market developments:
•
CDS replaced asset swaps as the credit derivative of choice in 98, allow to isolate
bond basis
•
Market develops as new derivatives become available: e.g., credit index
•
Crisis (Enron, Worldcom) push new regulations (observability of derivatives)
which was the first reason to push an index tranche market and base correlation
31 March 2016
7
Introducing background intensity
1.
Motivation for the concept of background intensity
2.
The default realisation marker, background filtration and intensity
3.
Revisiting the H hypothesis, and Kusuoka’s “remark”
4.
Generalised credit dynamics and risk neutral dynamics
31 March 2016
8
Motivation for the concept of background intensity
Default intensity can always be defined (in distribution sense) w.r. to a filtration:
•
General modeling: default time is a
•
Default compensator is defined by:
stopping time.
The default intensity is 0 after default. Therefore, it is typically not independent from default
occurrence.
Background Intensity
•
The concept of background intensity obtained by projection onto a background
filtration
is already well known. But the number of possible subfiltrations is
bewildering. What good properties do we seek?
•
Reduced form models have a natural choice of background filtration, this background
intensity is independent of the default realisation variable.
•
Other models (structural) are often used for multi-name modeling, do they have a
background intensity (defined in distribution sense), and what can be said of their
systemic risk?
Our goal is to clarify what are the useful hypotheses for background intensity in the context
of single name an multi-name pricing.
31 March 2016
9
The default realisation marker, background filtration and
intensity
Approach can be summarized as follow:
•
Orthogonalise
•
The default marker is defined by:
•
The useful assumption concerning background intensity is:
into
31 March 2016
10
Revisiting the H hypothesis, and Kusuoka’s “remark”
How does HH1 compare with usual assumptions and H hypothesis?
•
Decreasing background survival with t
•
Any
martingale is a
martingale (
)
Kusuoka’s remark (as seen by french bankers)
•
Extremly relevant due to the fundamental difference between finance and
measure theory. Finance is about replication after measure change.
•
Measure change introduces drifts for brownians, for poisson process means
change of intensity (see Cont)
•
H hypothesis not necessary hold under a non Ft adapted measure change
•
P intensity is the natural default intensity, Q intensity is the market price of
instantaneous protection. What is the condition on dQ/dP for the H
hypothesis to hold under risk neutral pricing?
31 March 2016
11
Generalised credit dynamics and risk neutral dynamics
Definition of hazard rates:
Survival Probability Diffusion and Martingality:
Risk Neutral Intensity
•
Market price of risk:
•
Hedging a T maturity derivative requires T credit delta and instant protection:
and
31 March 2016
12
Generalised credit dynamics and risk neutral dynamics
Under what condition can we talk of a Risk Neutral Intensity?
•
Valuing a T zero coupon: change from P to QT changes intensity
•
Change from Qb to QT does not change intensity unless IR term structure
jumps on default
•
Radon-Nikodym derivative needs to be in the background filtration for the H
hypothesis to hold under QT when it holds under P
What is the subfiltration of choice?
•
General dynamics are valid under any subfiltration of Gt
•
If the subfiltration does not verify HH1, the hazard rates are warped by the
projection. We want to study the market dynamics, so we use a subfiltration
that verifies HH1
31 March 2016
13
Five useful applications of background intensity modelling
1.
General credit dynamics formula, Introducing conditionally independent
default
2.
Diversification effect: results on forward loss distribution
3.
Stronger results for spot loss: conditional independence
4.
Introducing the canonical copula for portfolio loss
5.
Properties of the portfolio loss copula
31 March 2016
14
General credit dynamics formula, Introducing conditionally
independent default
General credit dynamics formula
Introducing conditionally independent default HH2
•
Assumption required for arbitrarily large sequence of exchangeable names:
Other names hazard rates do not jump on default
•
When this assumption is not verified, projection on the largest filtration
independent to default markers yields pseudo-intensities, not market
intensities
•
Our concern beyond martingality and intensity is to get dynamics that
describe moves of market implied probability.
31 March 2016
15
Diversification effect: defining systemic information
•
Background information is a good step towards the definition of systemic
information
•
If there are c arbitrarily large sequences of exchangeable names, and each
name hazard rate is driven by a brownian motion, we can define the systemic
filtration Fc
•
Hazard rate still have each an idiosyncratic evolution, but it is expected to
diversify away.
31 March 2016
16
Diversification effect: results on forward loss distribution
Main result:
31 March 2016
17
Stronger results for spot loss: conditional independence
Loss convergence
31 March 2016
18
Market info on forward loss distribution
•
The previous result describe the dynamics of the forward loss density.
•
The tranche market is a loss option market
What information do we get on loss density from the tranche market?
For instance, Dupire formula allows to extract local volatility information from
option prices given:
Here we do not have a term structure of options on the same underlying, but
options on the spot loss Lt where t varies
In the case of the coterminal option, we could get information on systemic
volatility, in the case of spot loss options, we get only information on
systemic intensity.
31 March 2016
19
Conclusion on diversification
•
We defined systemic information and showed diversification effects that
apply to the forward loss E(LT Gt)
•
We showed that the right variable conditional on which we obtain
diversification is the projection onto Fc, which is a martingale
•
We see that the right variable for the spot loss diversification is the same
projection indexed by t, which is a predictable process drifting by the
systemic intensity.
31 March 2016
20
Introducing the canonical copula for portfolio loss
Conditional independence is much stronger than diversification
It allows to price finite discrete portfolio by conditioning on the copula factor.
31 March 2016
21
Properties of the portfolio loss copula
31 March 2016
22
Conclusion on Canonical Copula
•
The De Finetti theorem already granted us similar results for the distribution
of the sum of excheangable bernouli variables (default indicators)
•
What we explicited is the link between survival dynamics and the one factor
copula variable obtainable with De Finetti theorem
•
We also obtained results concerning the copula for non exchangeable
variables
The large pool and exchangeable assumption helped us to
•
Rule out contagion
•
Explicit what happens when no single name is supposed to have a systemic
role in the economy
But ultimately, we can relax the exchangeability assumption and still obtain
results for discrete, heterogeneous portfolios.
31 March 2016
23