Upload
hoangkhanh
View
215
Download
0
Embed Size (px)
Citation preview
Systemic Risk and the Mathematics of Falling Dominoes
Reimer Kuhn
Disordered Systems GroupDepartment of Mathematics, King’s College London
http://www.mth.kcl.ac.uk/∼kuehn/riskmodeling.html
Teachers’ Conference, KCL, June 22, 2011
The Laws of Falling Dominoes
• A domino falls, if kicked sufficiently vigorously.
• A domino can be toppled by another domino.
• Avalanches can occur, if dominoes are set too closely.
2/23
Risk and Falling Dominoes
Operational Risk
Domino Theory & Spread
of Communism
Financial CrisisBlackouts in Power Grids
3/23
Fundamental Problem of Risk Analysis
• Estimation of risk
– Market: potential negative fluctuation of portfolio-value
(stock-prices, exchange rates, interest rates, economic in-
dices)
– Credit: potential change of credit quality, including de-
fault (asset values of firms, ratings, stock-prices)
– Operational: potential losses incurred by process failures
(human errors, hardware/software-failures, lack of com-
munication, fraud, external catastrophes)
– Liquidity: potential losses incurred by rare fluctuations
in cash-flows, requiring short term acquisition of funds to
maintain liquidity
4/23
Task
• Estimate PDF of
– Portfolio value at time T , PV (T), given PV (0)
– Market value of assets of obligors at time T , Ai(T), given
Ai(0)
– Losses L(T) due to process failures incurred during risk
horizon T
– Losses L(T) incurred during risk horizon T by need to
maintain liquidity in situations of stress
5/23
• Quantity of interest: Value at Risk
P(L)
L
1−q
EL
VaRq,T = (Qq[L(T)]−EL) e−rT Prob(L(T) ≤ Qq[L(T)]) = q
Money to set aside now, to cover losses in excess of expected
loss at time T , which are not exceeded with probability q.
• Requires good knowledge of tails of loss distributions.
6/23
Our Main Interest and Concern: Interactions
• Traditional approaches treat risk elements as independent or
at best statistically correlated
• Misses functional & dynamic nature of relations:
terminal–mainframe/input errors–results/manufacturer–supplier
relations . . .
• Effect of interactions between risk elements
- Possibility of avalanches of risk
events (process failures, defaults)
- Fat tails in loss distributions
- Volatility clustering in markets
(intermittency)
7/23
The Case of Operational Risks — Interacting Processes
• Conceptualise organisation as a network of processes
• Two state model: processes either up and running (ni = 0)
or down (ni = 1)
• Reliability of processes and degree of functional interdepen-
dence heterogeneous across the set of processes (quenched
disorder); connectivity functionally defined
⇒ model defined on random graph
• losses determined (randomly) each time a process goes down
(annealed disorder)
8/23
Dynamics – Mathematics of Falling Dominoes
• processes need support to keep running (energy, human re-
sources, material, information, input from other processes,
etc.)
• hit total support received by process i at time t
hit = ϑi −∑
j
Jijnjt + ηit
with ηit random (e.g. Gaussian white noise).
• process i will fail, if the total support for it falls below a
critical threshold (domino falls, if kicked too strongly)
nit+∆t = Θ
∑
j
Jijnjt − ϑi − ηit
9/23
Probability that a Domino Falls
• Probability of failure/probability of domino falling
Prob(
nit+∆t = 1∣
∣
∣n(t))
=∫
∑
jJij njt − ϑi
−∞dη p(η) ≡ Φ
∑
j
Jij njt − ϑi
p(η)
∑
jJij njt − ϑi
η
• unconditional and conditional probability of failure
pi = Φ(−ϑi)
pi|j = Φ(Jij − ϑi)
⇒ ϑi = −Φ−1(pi) , Jij = Φ−1(pi|j) − Φ−1(pi)
10/23
All-to-All Interactions — Mean Field Theory
• look at case with all-to-all couplings, (N ≫ 1)
Jij =J0
N, ∀ (i, j) ⇒
∑
j
Jijnjt =J0
N
∑
j
njt = J0mt
• Dynamics
nit+∆t = Θ
∑
j
Jijnjt − ϑi − ηit
= Θ(J0mt − ϑi − ηit) .
Thus (by LLN)
mt+∆t = =1
N
∑
i
Θ(J0mt − ϑi − ηit)
=1
N
∑
i
Φ(J0mt − ϑi) =⟨
Φ(J0mt − ϑ)⟩
ϑ
11/23
Mean Field Theory — Graphical Analysis
• Iterative dynamics mt+∆t =⟨
Φ(J0mt − ϑ)⟩
ϑ
mt
m t+1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
2 4 6 8 10 12 14 16 18
m
J0
Left: Graphic representation of iterated dynamics for three different values of J0 (curves
from left to right correspond to decreasing J0). Right: Location of fixed points as a
function of J0, for nonrandom ϑ = 2.25 ↔ pi ≃ 0.015. Backbending part of curve is
unstable fixed point.
12/23
Analysis for a Stochastic Setting
• Interactions on a random graph
Jij = cij
(
J0
c+
J√c
xij
)
with
P(cij) =c
Nδcij,1 +
(
1 − c
N
)
δcij,0 ,
symmetric (cij = cji), and
xij = 0 , x2ij = 1 , xijxji = α .
Study the limit
c ≫ 1 , N ≫ 1 .
Processes/organizations: small J ! 13/23
Results for α = 0
• Use limit theorems (LLN, CLT) to show∑
j Jijnjt is Gaussian.• If noise ηt ∼ N (0, σ) (also Gaussian), then
mt+∆t =
⟨
Φ
J0mt − ϑ√
σ2 + J2mt
⟩
ϑ
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
-2 0 2 4 6 8 10 12 14 16
m
J0
2
4
6
8
10
12
14
16
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11
J 0c
p
N at 0.035,8C at 0.025,5O at 0.015,3
O
C
N
(K Anand and RK, Phys Rev E75 (2007))
Left: Stationary fraction of down-processes for pi = 0.01, α = 0, and J = 0.2. Right: phase
diagram.14/23
Stationary Solution at α 6= 0
• Have to work much harder.
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16
m
J0
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10 12 14 16m
J0
(K Anand and RK, Phys Rev E75 (2007))
Stationary fraction of down-processes for α = 0.5 (left) α = 1.0 (right) at J = 0.2
15/23
Spontaneous Blackouts
• If mutual dependence of processes exceeds a certain level,
spontaneous blackouts can occur.
0 10000 20000 30000 40000 50000time
0
100
200
300
loss
es
0 10000 20000 30000 40000 50000time
0
100
200
300
loss
es
Loss record for a system of N = 50 interacting processes. Left panel: Coexisting low-loss
and high-loss phases @ pmaxi = 0.02 and (pi|j/pi)
max = 2.6. Right panel: Low-loss phase
simulated for the same pmaxi but (pi|j/pi)
max = 3. The low-loss situation is only meta-stable
and spontaneously deays to a high-loss situation via a bubble-nucleation process.
16/23
Key Features
• For sufficiently strong cooperative interactions — “dominoes
too densely placed” — get a (first order) phase transition:
coexistence of ‘functioning state’ and state of ‘catastrophic
breakdown’
• There is a critical Jc inside the coexistence region, beyond
which the system — while still working — prefers breakdown,
• Critical point at sufficiently high pi.
• Of special relevance: resilience to (external) stress can be
tested.
17/23
The Case of Credit Risks — Interacting Companies
• Risk arising from the possibility of obligors going bankrupt
or from changes in ‘credit quality’ (⇒ credit trading)
• Here only influence of defaults
• Two state model: company up and running (ni = 0) or down
(ni = 1)
• Probabilities of default and mutual impacts of defaults het-
erogeneous across the set of firms (quenched disorder); con-
nectivity functionally defined
⇒ model defined on random graph
• Losses determined (randomly: recovery process) when a com-
pany defaults (annealed disorder)
18/23
Dynamics — Falling Dominoes Again
• Companies need “orders” (support, cash inflow) to maintain
wealth and avoid default
• hit total wealth of company i at time t,
hit = ϑi −∑
j
Jijnjt − ηit
with ηit random (e.g. Gaussian white noise).
• company i defaults, if the total wealth falls below zero
nit+1 = nit + (1 − nit)Θ
∑
j
Jijnjt − ϑi + ηit
• No recovery within ‘risk horizon’ T : ni = 1 is absorbing state.
Time unit: 1 month; T = 12 ⇔ 1 year.
19/23
Global Analysis
• Analysis in a stochastic setting on random graphs as for OR,
• Loss distribution through fluctuating macro-economic
conditions. Noise model (minimal Basel II):
ηit =√
ρi η0t +√
1 − ρi ξit
with η0t slow economy-wide component, and ξit idiosyncratic
noise, and ρi = ρ(ϑi)..
〈nt+1(ϑi)〉 = 〈nt(ϑi)〉 +(
1 − 〈nt(ϑi)〉)
Φ
J0mt +√
ρi η0 − ϑi√
1 − ρi + J2mt
mt+1 = mt +
⟨
(
1 − 〈nt(ϑ)〉)
Φ
J0mt +√
ρ η0 − ϑ√
1 − ρ + J2mt
⟩
ϑ
20/23
Results CR — Defaulted Fraction and Loss Distribution
0
0.005
0.01
0.015
0.02
0.025
0 2 4 6 8 10 12
mt
t
1e-07
1e-06
1e-05
1e-04
0.001
0.01
0.1
1
-20 0 20 40 60 80 100 120 140 160 180
P(L
)
L
(JPL Hatchett and RK, J Phys A39 (2006))
Fraction of defaulted firms for neutral macro-economic conditions η0 = 0 at (J0, J) =
(0,0), (1,0), (0,1) and (1,1) (left, bottom to top; Loss distribution for a system with (J0, J) =
(0,0) and (1,1) and ℓ(ϑ) = 1/(ε + pd(ϑ)) (right)
21/23
Key Features
• Interpretation of parameters ϑi and Jij in terms of uncondi-
tional and conditional default probabilities
• For sufficiently strong interactions — in particular coopera-
tive interactions — get possibility of collective acceleration
of default rates in times of economic stress
• Typical behaviour less sensitive to interactions than rare events
⇒ fat tails in loss distribution; important for risk analysis
where rare event asymptotics is relevant.
• Due to initial conditions ni0 ≡ 0 and absorbing state:
⇒ no equilibrium dynamics
22/23
Summary
• Standard risk models based on statistically correlated risk el-
ements miss dynamically generated functional correlations
• Interacting processes capture functional dependencies in OR.
• Similarly: economic interactions ⇒ credit contagion in CR
• Physics analogy: lattice gas model on (random) graph.
• Describes bursts and avalanches of risk events (dominoes)
• Important: coexistence with phases of catastrophic break-down — no noticeable precursors !
• Generalized Credit risk model to include derivative trading (CDS).
• Neuro-dynamics model for intermittent dynamics & fat-tailed loss distributions in MR
23/23
THANK YOU
Supplementary Material
Results CR — Value at Risk
0.95
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
< L
>/<
L >
0 V
aR/V
aR0
R
0.95
1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
1.45
1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
< L
>/<
L >
0 V
aR/V
aR0
R
0.9
1
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
< L
>/<
L >
0 V
aR/V
aR0
R
(JPL Hatchett and RK, J Phys A39 (2006))
Ratios of value at risk (upper curves) and average losses (lower curves) for systems with
and without functional interaction evaluated along straight lines in the J0-J plane, with
R =√
J20 + J2. Left: J0 = 0 ; right: J = 0; lower: J0/J = 1.
Credit Risk with CDS
Fee payment
Reimbursement
Credit event
Protection buyer
Protection seller
Reference entity
paymentInterest
Credit
Fee payment
Reimbursement
Credit event
Protection buyer
Protection seller
Reference entity Bank
Credit
Interest payment
Left: Mechanics of a CDS used for hedging. Right: Mechnanics of a speculative CDS.
0.0001
0.001
0.01
0.1
1
10
-0.2 0 0.2 0.4 0.6 0.8 1 1.2
P(L
)
L
0.0001
0.001
0.01
0.1
1
10
100
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
P(L
)
L
Baseline-scenario (red full line), and either one third (green long-dashed) or two thirds (blue
short-dashed) of the avarage exposure to firms are hedged by CDS, with insurance sector ,
Right: baseline-scenario (red full line) compared with situations where banks have completely
hedged their exposures through CDS and used this to either double (green long-dashed) or
triple (blue short-dashed) the size of their loan books.
Functional Dependencies in Market Risk
• Standard approach: Geometric Brownian motion (GBM) for
Log-returns of risk elements i (stocks, bonds . . . )
1
Si(t)
dSi(t)
dt= µi + σiηi(t)
• Statistical dependencies via correlated Gaussian noises
〈ηi(t)ηj(t′)〉 = ρijδ(t − t′)
• Is basis of JP Morgan’s Risk MetricsTM tool!
• But: no functional dependencies, no collective behaviour, no
market bubbles, crashes
28/23
Functional Dependencies in Market Risk — iGBM
• Functional dependencies between risk elements e.g. via
recommendations of analysts, economic dependencies.
• GBM in terms of hi(t) = log(Si(t)/Si0)
dhi(t)
dt= µi −
1
2σ2
i + σiηi(t)
• Minimal interacting generalisation (iGBM): stabilisation and
non-linear feedback (e.g. g(x) = tanh(x))
dhi(t)
dt= −κhi(t) + µi −
1
2σ2
i +∑
j
Jijg(hj(t)) + σiηi(t)
• ⇔ dynamics of of graded-response neurons
many meta-stable states; transitions between them ⇒ inter-
mittent dynamics.
Functional Dependencies in Market Risk – iGBM
• Solve in stochastic setting as for OR/CR
• Here MC study; trigger transitions by ‘unexpected news’
0 1000 2000 3000 4000 5000t/τ
−1500
−500
500
1500
∆I(t
)
−6 −4 −2 0 2 4 6∆h/σ
0.0001
0.001
0.01
0.1
1
10
σ P
(∆h/
σ)
τ=25τ=50
(RK and P Neu, J Phys A41 (2008))
Change of index I(t) = N−1∑
iSi(t) over time increment τ = 25 (left) and normalised
distribution of log-returns for τ = 25, and τ = 50 (right).