Vous êtes sur la page 1sur 22

Computational Management Symposium 2004

Long Term Policy Making for Operational Risk management in the Basel II framework
University of Neuchatel
Room ALG Session 16 Duc PHAM-HI, Head of IT Dept., prof. Computational Finance,

E.C.E engineering school, Paris, France

Focus is on Risk management in Banks


Definitions (New Basel Accord) "risk of loss from failed or inadequate process, systems or people, resulting from internal or external events" Basel II means profound changes quantification of op risks for mandatory Capital reserves involvement of top management, held accountable very large compliance projects Operational Risk in banks 7 event categories quantitative approach

31/03/2014

Duc Pham-Hi

Operational risk: Advanced Model Approach


Loss distribution approach hope for lower Capital requirements risk control performances watched by Rating agencies

Probability

Loss occurences

Probability

Probability

Monte Carlo Simulation etc.


individual Loss amount

Mean

Threshold

Total loss

31/03/2014

Duc Pham-Hi

Basel II, computation-wise


Basel II for the Top banks : Advanced Model Approach Pillar 1: capital at risk calculation use of all 4 inputs mandatory, but internal data scarce allocation between part and whole? source of risk is absent from model Pillar 2: control and supervision governance and management role are not modelised consistency and rationality: no objective way to judge General: procyclicity look-back etc Conclusion : crying lack for a theory of dynamic risk control to build a rational framework
31/03/2014 Duc Pham-Hi 4

Levy process for modeling risk state


2 kinds of risks
Ordinary business losses Negligible losses Extraordinary catastrophic losses

"Business losses" are frequent, and small, and : rather regular both positive and negative "Catastrophic losses" are rare and very large and unpredictable only negative ("no free lunch" effect)
31/03/2014 Duc Pham-Hi 5

Bank activity variables for modeling business


Small and frequent losses
X1 X 1,0 exp t Bt

dX1

X 1 dBt
X 2,0 exp Lt

Rare but very large losses


dX 2 X2 b
R*

X2
R*

(e z 1 z1 z 1 ) (dz) dt

~ (e z 1) N (dt, dz)

after application of Ito's formula, where : ~ N (dt, dz) N (dt, dz) dt (dz) is the compensated Poisson random measure, (dz) is the Lvy measure, with condition
e z 1 (dt , dz )

z 1

All operational losses Revenue income process Wealth is resultant of losses and gains
31/03/2014 Duc Pham-Hi

dX
dW

dX1 dX2

dR

dt

dR dX
6

Parameters for modeling Management


Key Risk Indicators and Scorecards on Business Environment : indicates level of danger/volatility t Scorecards on Internal Control t How effective does Management transform budgets into Loss reduction Approximation Assmussen & Rosinski Cumulated large losses as finite number of jumps, in practical
N (t )

Xt
1

xj

31/03/2014

Duc Pham-Hi

Risk state description: effects of management / environment


Impacts on system
Wealth W Environt var. Transition Probability Control variables , Wealth W+dW

Reducing small losses through better process management


F ( t)
t

( )

where t is the loss reduction factor whose cost is expense (more fraud detection, personnel, etc.) Reducing impact of catastrophic events through insurance, or N (t ) recovery plans, at cost
G ( Lt ) K (x j )
0

where losses xj may be capped or reduced


K ( x j ) inf H ( ), x j K (x j )
31/03/2014

where H ( ) H( ) 1

fixed amount

H ( ).x j

where 0
Duc Pham-Hi

Introducing rationality and control

Wealth evolution is :
N (t )

dW

Rt

) dt

X 1 ( ) ( )dBt
j 1

K (x j )
( ), t , , )

State of wealth at time : Economic value depending on policy

W
0

dW (t ,

J
where is the given of a pair ( (t) , (t)) .
0

exp( rt ) U W (t ) dt

Objective is to maximise:
V max E
0

exp( rt ) U W (t ) dt

31/03/2014

Duc Pham-Hi

Other related problems :


Refer to Cramer-Lundberg : Ruin theory
Goal to stay afloat becomes constraint

31/03/2014

Duc Pham-Hi

10

How to solve : 3 choices of methods


Pure HJB, G-HJB : Galerkin + Viscosity consistent with events classification Neural / Learning Adaptive: Reinforcement Learning Neural, supervised QMC or MCMC exploratory Has to back test and stress test anyway Scenario exploration thru parameters : Markov chaining events instead of Experts panel imagination. Classic dilemma : explore or exploitation : genetic programming (splice & dice)
31/03/2014 Duc Pham-Hi 11

rk

Using Hamilton Jacobi Bellman to model risk ?

Introducing processes Optimal control gives the big picture Modeling processes Process decomposition Equation for risk genesis Introducing Value Optimal control equations Solving for strategies Solving for price of risk Feature based reasoning

31/03/2014

Duc Pham-Hi

12

Drawing help from Techniques of process modeling


Classification as Top down or Bottom up Or, as either Process / Factor : process analysis causal networks connectivity matrix factors identification risk indicators - and predictive models CAPM like and volatility models actuarial techniques empirical loss distribution distributional form parametricized by historical data extreme value theory Hamilton Jacobi Bellman is related to process but transition matrix in Markovian case can be related to Bayesian networks.
31/03/2014 Duc Pham-Hi 13

Banking processes and losses separation


Need for a separate process model and a Losses model... Risks identification is different from risk evaluation one is cost, other is production capital at risk evaluation is not (always) risk evaluation Risk evaluation cannot be done properly without risk reduction which cannot be treated if generation mechanism not seen parameters for risk source easier to catch at op level De facto schizoid treatment separating identification that combines back together at time dimension cost reduction, management from Y to Y+1 at investment cost dimension allocate CaR according to marginal efficiency of capital, not at highest ex post costs therefore justify event type cats as axis

31/03/2014

Duc Pham-Hi

14

Modeling hypotheses and issues


Standard approach (not AMA) and line of business Basel II is profit center orientated, not back office... but operational risk stems from back office modeling correlation errors if share back office processes markovian hypothesis homogeneity issue on local risk-adjusted returns on capital across different business lines Elements for the process model profit generation rate estimates of risk level factor stochastic variable is a LDA type : Extreme value theory to help here.

31/03/2014

Duc Pham-Hi

15

Long history of solving HJB


Known cases Linear Quadratic in Merton's problem and Black Scholes context New : with Levy processes, but with HARA and CARA utility function : explicit solutions But still, classic obstacles Unknown P(x,y) --> POMDP in Markovian case Curse of dimensionality Too large sets {y} for each x Too large sets { } for each x Too strong nonlinearities

31/03/2014

Duc Pham-Hi

16

Learning : quick sampling of state space wrt. rewards


Neural techniques on features
Mixing SDE with NN-based volatilities

To use Q-learning, philosophy is Action-Reward : use of at :

at=
Then

(xt) ,
E r ( x, a)

R( x, a)

Taking null terminal value, the value function is the total of what can be expected in the future (here discount is not present).

V ( x, ( xt ), t )

t 0

r ( x, ( xt ), t )dt

Introducing a discount rate and taking the expected value :

V ( x)

E
t 0

r ( xt , ( xt )) x0

31/03/2014

Duc Pham-Hi

17

Solving for strategies or for value of risk ?


Rationality of the risk center is to seek maximum of value, starting from state t x0 , to "learn" policy maximising V over set A of admissible actions * satisfying

V ( x0 )

min V ( x0 )
A

where V (x) is the the consequence of following policy


V ( x) R( x, ( x)) E
t 0 t

from initial situation x

r ( xt , ( xt ))

Value for a given strategy is sum of immediate reward and discounted flow of possible future rewards, depending on the transition : Vt 1 ( x) R( x, t 1 ) . Px , y .Vt ( y )
y

Solving for optimal policy requires dealing with nonlinear equation


*

( x)

arg min R( x, ( x)) E


t 0

r ( xt , ( xt ))

31/03/2014

Duc Pham-Hi

18

Solving for risk with Temporal differences


Solving for strategies is non-linear, we turn to solving for value (G is terminal value) V ( x, t ) min V ( f ( x), t 1) G( x, u, t )
u U

by reasoning in terms of discrete time. Alternately, in terms of discrete states y, as possible outcomes of state x, and introducing action at :
V* min r (a, x)
a y

P( x, y )V ( y )

P ( x, y ) 1
y

x , P ( x, y ) 0

We iterate on V since the problem is linear. let t be the proxy for V at time t ; we iterate thus :
Vt 1 ( x ) min r ( x, ) .
y

P ( x, y )

( y)

Q-learning is particularly easy to set up and is model-free

Qt 1 (s, a)
31/03/2014

g ( s, a )

p(s, y).Qt ( y, a)
19

Duc Pham-Hi

Optimal control gives the big picture


Control using investment mechanism for Capital at Risk deduction justifiable. give keys for allocation Joining bottom-up and top-down management clear rules of engagement for both operational and senior management on budget Remedy scarce data : How ? e.g by reinforcement learning, greedy algorithm: not necessarily CPU consuming by justifying online sampling compensate scarce "experience points" with observations-weight modifications (ANN like techniques - filtering) allow use of Monte Carlo by process even where no history was available Regulator oversight advantage secure model validation

31/03/2014

Duc Pham-Hi

20

Even if not solving to 100% it's setting the framework


Separate roles of parameters from variables See role of Business Environment & Internal Control Adaptive learning Online sampling

31/03/2014

Duc Pham-Hi

21

Adressing gaps in today's risk management theory


Temporal dimension : growth related risk (agressive strategies) planning Allocation problem : J = Jk Home Host business line arbitrage Adaptive : online sampling learning Operational Real Time hedging : future scanning with high quality inhomogenous Poisson time identification of chains of events as patterns
31/03/2014 Duc Pham-Hi 22

Vous aimerez peut-être aussi