Working Paper Series

10/2007
Bayesian analysis and Markov chain Monte
Carlo simulation
Medova, E.A.

These papers are produced by Judge Business School, University of Cambridge.

They are circulated for discussion purposes only. Their contents should be
considered preliminary and are not to be quoted without the authors permission.

Author contact details are as follows:

Dr E A Medova
Centre for Financial Research
Judge Business School
University of Cambridge
[email protected]

Please address enquiries about the series to:

Research Support Manager
Judge Business School
Trumpington Street
Cambridge CB2 1AG, UK
Tel: 01223 760546 Fax: 01223 339701
E-mail: [email protected]

Bayesian Analysis and Markov Chain Monte Carlo Simulation

Elena Medova
Centre for Financial Research, Judge Business School, University of Cambridge
Trumpington Street, Cambridge CB2 1AG, UK
&
Cambridge Systems Associates Limited
5-7 Portugal Place, Cambridge CB5 8AF, UK
Tel: +44 1223 339593 Fax: +44 1223 339652
Email: [email protected] Web: www-cfr.jbs.cam.ac.uk

Overview of main concepts

Bayesian analysis offers a way of dealing with information conceptually different from all other
statistical methods. It provides a method in which observations are used to update estimates of the
unknown parameters of a statistical model. With the Bayesian approach we start with a
parametric model that is adequate to describe the phenomenon we wish to analyze. Then we
assume a prior distribution for the unknown parameters of the model which represent our
previous knowledge or belief about the phenomenon before observing any data. After observing
some data assumed to be generated by our model we update these assumptions or beliefs. This is
done by applying Bayes theorem to obtain a posterior probability density for the unknown
parameters given by

p ( | x ) =

p ( x | ) p()
,
p( x | ) p ()d

where is the vector of unknown parameters governing our model, p ( ) is the prior sampling

density function of and x is a sample drawn from the true underlying distribution with
sampling density p(x | ) that we model. Thus the posterior distribution for takes into account
both our prior distribution for and the observed data x.

A conjugate prior family is a class of densities { p (i ) } which has the feature that given the
sampling density p(x |) the posterior density p (i x) also belongs to the class. The name arises
because we say that the prior p (i ) is conjugate to the sampling density considered as a
likelihood function p ( x ) for given x. The concept of conjugate prior as well as the term was
introduced by Raiffa and Schlaifer [14].
After obtaining a posterior distribution for the parameters we can compute various quantities of
interest such as integrals of the form

f ( y ) g ( y; ) p ( x) dyd ,

(1)

where f is some arbitrary function and g is the probability density function describing a related
parametric model. In general, because we are not assuming independence between the each of the
individual parameters this integral is difficult to compute, especially if there are many parameters.
This is the situation in which Markov chain Monte Carlo (MCMC) simulation is most
commonly used.
The distinguishing feature of MCMC is that the random samples of the integrand in (1) are
correlated, whereas in conventional Monte Carlo methods such samples are statistically
independent. The goal of MCMC methods is to construct an ergodic Markov chain that converges
quickly to its stationary distribution which is the required posterior density or some functional
thereof such as (1).
One can broadly categorize the use of MCMC methods as Bayesian or non-Bayesian. NonBayesian MCMC methods are used to compute quantities that depend on a distribution from a
statistical model that is non-parametric. In a Bayesian application we consider a parametric
model for the problem of interest. We assume some prior distribution on the parameters and try to
compute quantities of interest that involve the posterior distributions. This approach remains
suitable if the data is sparse, for example, in extreme value applications [10].
There are many different types of MCMC algorithms. The two most basic and widely used are
the Metropolis-Hastings algorithm and the Gibbs sampler which we will now review.

Metropolis-Hastings algorithm
The Metropolis-Hastings algorithm [11, 8, 4] has been used extensively in physics but was little
known to others until Mller [12] and Tierney [19] expounded the value of this algorithm to
statisticians. The algorithm is extremely powerful and versatile and has been included in a list of
top 10 algorithms [5] and even claimed to be most likely the most powerful algorithm of all
time [1].
The Metropolis-Hastings algorithm can draw samples from any target probability density for
the uncertain parameters requiring only that this density can be calculated at . The algorithm
makes use of a proposal density q ( t , ) which depends on the current state of the chain t to
generate each new proposed parameter sample . The proposal is accepted as the next state
of the chain ( t +1 := ) with acceptance probability ( t , ) and rejected otherwise. It is the
specification of this probability that allows us to generate a Markov chain with the desired
target stationary density . The Metropolis-Hastings algorithm can thus be seen as a generalized
form of acceptance/rejection sampling with values drawn from approximate distributions which
are corrected in order that they behave asymptotically as random observations from the target
distribution.
The algorithm in step-by-step form is as follows:
a) Given the current position of our Markov chain t , generate a new value from the
proposal density q (see below).
b) Compute the acceptance probability

( ) q ( , t )
,
( , ) := min 1,
( t ) q ( t , )

(2)

where is the density of the target distribution.

c)

With probability (t , ) , set t +1 := , else set t +1 := t .

d) Return to step a).

( )

This algorithm generates a discrete time ergodic Markov chain t

corresponding to , i.e. as t

t 0

with stationary distribution

P ( t B ) ( B )
for all suitably (Borel) measurable sets B

Some important points to note [4]:

We need to specify a starting point 0 , which may be chosen at random (and often is).

Preferably 0 should coincide with a mode of the density .

We should also specify a burn-in period to allow the chain to reach equilibrium. By

this we mean that we discard the first n values of the chain in order to reduce the
possibility of bias caused by the choice of the starting value 0 .

The proposal distribution should be a distribution that is easy to sample from. It is also

desirable to choose its density q to be close or similar to the target density , as this
will increase the acceptance rate and increase the efficiency of the algorithm.

We only need to know the target density function up to proportionality that is, we

do not need to know its normalizing constant, since this cancels in the calculation (2) of
the acceptance function .
The choice of the burn-in period still remains somewhat of an art, but is currently an active area
of research. One can simply use the eyeballing technique which merely involves inspecting
visual outputs of the chain to see whether or not it has reached equilibrium.
When the proposal density is symmetric, i.e. q (t , ) = q (, t ) (the original Metropolis
algorithm), the computation of the acceptance function is significantly faster. In this case from
(2) a proposal is accepted with probability =()/( t), i.e. its likelihood () relative to that
of ( t) (as originally suggested by Ulam for acceptance/rejection sampling).

Random walk Metropolis

If q(t , ) := f (| |) for some density f and norm || then this case is called a random walk
chain because the proposed states are drawn according to the process following = t + ,

where F , the distribution corresponding to f. Note that since this proposal density q is
symmetric the acceptance function is of the simple Metropolis form described above. Common
choices for q are the multivariate normal, multivariate t or the uniform distribution on the unit
sphere.
If q (t , ) := q ( ) then the candidate observation is drawn independently of the current state of
the chain. Note however that the state of the chain t +1 at t + 1 does depend on the previous state
t because the acceptance function (t , ) depends on t .

In the random walk chain we only need to specify the spread of q , i.e. a maximum for | | at a
single step. In the independence sampler we need to specify the spread and the location of q .

Choosing the spread of q is also something of an art. If the spread is large, then many of the
candidates will be far from the current value. They will therefore have a low probability of being
accepted, and the chain may remain stuck at a particular value for many iterations. This can be
especially problematic for multi-modal distributions; some of its modes may then not be explored
properly by the chain. On the other hand, if the spread is small the chain will take longer to
traverse the support of the density and low probability regions will be under-sampled. The
research reported in [13] suggests an optimal acceptance rate of around 0.25 for the random walk
chain. In the case of the independence sampler it is important [2] to ensure that the tails of q
dominate those of , otherwise the chain may get stuck in the tails of the target density. This
requirement is similar to that in importance sampling.

Multiple-block updates
When the number of dimensions is large it can be difficult to choose the proposal density q so
that the algorithm converges sufficiently rapidly. In such cases it is helpful to break up the space
into smaller blocks and to construct a Markov chain for each of these smaller blocks [8]. Suppose

that we split into two blocks (1 , 2 ) and let q1 1t t2 , 1 and q2 t2 1t , 2 be the proposal

densities for each block. We then break each iteration of the Metropolis-Hastings algorithm into 2
steps and at each step we update the corresponding block. To update block 1 we use the
acceptance function given by

(
(

) (
) (

)
)

(
(

) (
) (

) .
)

1 t2 q1 1 t2 , 1t
, 1 := min 1,
1t t2 q1 1t t2 , 1

t
1

t
2

(3)

and to update block 2 we use

2 1t q2 2 1t , t2
, 2 := min 1,
t2 1t q2 t2 1t , 2

t
2

t
1

(4)

If the blocks each consist of just a single variable, then the resulting algorithm is commonly
called the single-update Metropolis-Hastings algorithm. Suppose in the single-update algorithm it
turns out that each of the marginals of the target distribution (i ~ i ) can be directly sampled
from. Then we would naturally choose q (i ~ i ) := (i ~ i ) since all candidates will then be
accepted with probability 1. This special case is the well-known Gibbs sampler [2].

Gibbs sampler
Gibbs sampling is applicable in general when the joint parameter distribution is not known
explicitly but the conditional distribution of each parameter given the others is known. Let
P ( ) = P ( 1 ,..., k ) denote the joint parameter distribution and let p ( i | ~ i ) denote the
conditional density for the i -th component i given the other k 1 components, where

~ i := j : j i for i = 1,..., k . Although we do not know how to sample directly from P we do

know how to sample directly from each p ( i | ~ i ) . The algorithm begins by picking the

arbitrary starting value 0 = 10 ,..., 0 . It then samples randomly from the conditional densities
p ( i | ~ i ) for i = 1,..., k successively as follows:

(
p (

)
| , ,..., )

Sample 11 from p 1 | 02 , 30 ,..., 0k

Sample 12 from

1
1

0
3

0
k

Sample 1k from p k | 11 , 12 ,..., 1k 1 .

This completes a transition from 0 to 1 and eventually generates a sample path 0 , 1 ,..., t ,
of a Markov chain whose stationary distribution is P.
In many cases we can use the Gibbs sampler which is significantly faster to compute than the
more general Metropolis-Hastings algorithm. In order to use Gibbs however we me must know
how to directly sample from the conditional posterior distributions for each parameter,
i.e. p (i ~i , x) , where x represents the data to time t.

Use of MCMC in capital allocation for operational risk

Due to lack of reported data on operational losses Bayesian Markov chain Monte Carlo
simulation is well suited for the quantification of operational risk and operational risk capital
allocation. In [9] a framework for evaluation of extreme operational losses has been developed
which assumes that market and credit risks may be managed separately but jointly impose a value
at risk limit uVaR on these risks.

It is assumed that losses beyond the uVaR level belong to the operational risk category. In most
cases, due to overlapping between risk types a detailed analysis of operational loss data is
required to support the assumption that the uVaR level approximately equals the unexpected loss
threshold. This approach to capital allocation for operational risk takes into account large but
rare operational losses, is naturally based on extreme value theory (EVT) [6,7] and focusses on

tail events and modelling the worst-case losses as characterized by loss maxima over regular
observation periods.
According to regulatory requirements [20] operational risk capital calculation requires two
distributions a severity distribution of loss values and a frequency distribution of loss
occurrences. In the approach described here a unified resulting asymptotic model known as the
peaks over threshold (POT) model [15,16,18] is applied. It is based on an asymptotic theory of
extremes and a point process representation of exceedances over a threshold given by the POT
model. The following is assumed.
Given an i.i.d. sequence of random losses X1,, Xn drawn from some distribution we are
interested in the distribution of the excess Y := X u over the threshold u. The distribution of
excesses is given by the conditional distribution function in terms of the tail of the underlying
distribution function F as

Fu ( y ) := P( X u y | X > u ) =

F (u + y ) F (u )
for 0 y .
1 F (u )

(5)

The limiting distribution G, ( y ) of excesses as u is known as the generalized Pareto

distribution (GPD) with shape parameter and scale parameter given by

y
G, ( y ) = 1 1 +

1/

y
1 exp

0 where y [ 0, ] 0 or y [ 0, / ] < 0

(6)

= 0.

The identification of an appropriate threshold u is again somewhat of an art and requires a data
analysis based on a knowledge of EVT [3, 6, 7].
The capital provision for operational risk over the unexpected loss threshold u is given in [10] as

u E ( X u | X > u ) = u

u + u
,
1

(7)

where E ( X u | X > u ) =

u + u
is the expectation of excesses over the threshold u (which is
1

defined for 1 and must be replaced by the median for >1), u := + ( u ) and the
exceedances form a Poisson point process with intensity

(u )
u := 1 +

1/

(8)

usually measured in days per annum.

The accuracy of our model depends on accurate estimates of the , , and parameters. To
address this, hierarchical Bayesian Markov chain Monte Carlo simulation is used to determine the
parameter estimates of interest through intensive computation. The empirical estimation
efficiency of this method when back-tested on large data sets is surprisingly good.
Hierarchical Bayesian parameter estimation considers the parameters to be random variables
possessing a joint probability density function. The prior density f | of the random parameter
vector is parametric with a vector of random hyper-parameters and is conjugate prior to the
sampling density f X | so that the calculated posterior density f | X

1 ,..., X n ,

:= f |+ is of the same

form with the new hyper-parameters + determined by and the observations X 1 ,..., X n . In
the hierarchical Bayesian model the hyper-hyper parameters are chosen to generate a vague
prior due to the lack of a prior distribution for the hyper-parameters before excess loss data is
seen. Hence, we can decompose the posterior parameter density f | X , with the observations X
and the initial hyper-hyper parameters as

f | X , f X | ( X | ) f | ( | ) f ( | )
f X | ( X | ) f | ( | , )
f X | ( X | ) f ( | + ) .

Here the Bayesian update of the prior parameter density f

f | f

is performed in 2 stages.

First by updating the hyper-hyper parameters to + conditional on then evaluating the

corresponding posterior density for this given the observations X .
Hierarchical Bayesian MCMC simulation for the parameters is based on the Metropolis-Hasting
algorithm described briefly above and in detail in [17]. The idea is that the state of the chain for
the parameter vector := {j, log j, j : j = 1, 2,, J} converges to a stationary distribution
which is the Bayesian posterior parameter distribution f | x , given the loss data x and a vector

of hyperparameters{ m , s2 , mlog , s2log , m , s2 }. The hyperparameters are sampled from a

conjugate prior gamma-normal (GM) distribution and are used to link the parameters {j, j, j :

j = 1, 2,, J} of each individual risk [10].

The aim of the model is to estimate the parameters of interest {j, j, j : j = 1, 2,, J}
conditional on both the data x and the hyperparameters { m , s2 , mlog , s2log , m , s2 }. The

posterior distributions of the parameters are normally distributed:

j N(m , s2), log j N(mlog , s2log) and j N(m , s2).

A schematic summary of the loss data, parameters and hyperparameters is given in Table 1.

10

Table 1: Bayesian hierarchical model

Type 1

Type 2

Business Unit 1

x11

x12

x1J

Business Unit 2

x21

x22

x2J

xn,1

xn,2

xn,J

Data

Business Unit n

Type J

Hyperparameters

Parameters
Mean ()

Mean m

Scale (log)

Log 1

log 2

log J

Mean mlog Variance s2log

Shape ()

Mean m

Variance s2
Variance s2

Illustrative example
The data is assumed to represent the operational losses of a bank attributable to three different
business units. The data starts on 03.01.1980 and ends on 31.12.1990. The time span is calculated
in years hence the parameters will also be measured on a yearly basis. The data has been
generated from the Danish insurance claims data [3] by two independent random multiplicative
factors to obtain the three sets of loss data summarized in Table 2.
A typical analysis of such data includes time series plots, log histogram plots, sample mean
excess plots, QQ plots for extreme value analysis against the GPD, Hill estimate plots of the
shape parameter and plots of the empirical distribution functions. All these tests have been
performed for the three data sets to conclude that data are heavy tailed and that the POT model is
valid.

11

Table 2: Summary statistics for data

Data

X1 (Danish)

X2

X3

Min:

1.000

0.8

1.200

1st Qu.

1.321

1.057

1.585

Mean

3.385

2.708

4.062

Median

1.778

1.423

2.134

3rd Qu

2.967

2.374

3.560

Max

263.250

210.600

315.900

2167

2167

2167

StdDev

8.507

6.806

10.209

50

100

150

200

250

Figure 1: Time series of log danish data X1

030180

030182

030184

030186
Time

12

030188

030190

Figure 2: Histogram of log danish data

Percent of Total

30

20

10

0
0

log(danish)

Inputs for the MCMC model:

Threshold u = 30
Initial parameters:

log 1

log 2

log 3

20

21

22

3.2

2.8

0.5

0.4

0.7

The tables below are a summary of the posterior mean estimates of the parameter values i and i
based on the MCMC posterior distribution mean parameter values.

For j = 1 (Unit 1)

Code

1000 loops
2000 loops

Mean(1)

Mean(log1)

Mean(1)

Expected Excess

37.34

3.14

0.77

18.46

1.41

180.70

36.89

3.13

0.80

18.35

1.39

211.75

The number of exceedances above threshold is 15.

13

For j = 2 (Unit 2)

Code

Mean (2)

Mean(log2)

Mean(2)

Expected Excess

36.41

3.16

0.77

19.04

1.34

186.22

35.76

3.13

0.8

18.5

1.30

218.40

1000 loops
2000 loops

The number of exceedances above threshold is 11.

For j = 3 (Unit 3)

Code

1000 loops
2000 loops

Mean (3)

Mean(log3)

Mean(3)

39.55

3.05

0.79

14.21 1.71

180.52

39.23

3.03

0.82

13.83 1.70

213.50

Expected Excess

The number of exceedances above threshold is 24.

The plots in Figure 3 below for the results of 2000 simulation loops show that convergence has
been reached for the marginal posterior distributions of all parameters for Unit 1 and that the
estimates of these parameters are distributed approximately normally. (Those of are thus
approximately lognormal.) Similar results hold for the other two units.
The capital provision for operational losses is calculated using expression (7). The probability of
such losses is given by the choice of threshold u for extreme operational losses. This threshold
must be obtained from an analysis of the historical operational loss data and should agree or
exceed the threshold level uVaR of unexpected losses due to market and credit risk. The
probability of crossing the combined market and credit risk threshold uVaR is chosen according to
the usual value at risk (VAR) risk management procedures. The level of losses u due to
operational risks is exceeded with probability , so that u uVaR . The probability of
exceeding u depends on the shape of the tail of the loss distribution but is in general very much
smaller than .
Assuming that three types of losses are the bank business unit losses from operational risk over a
period of 11 years the bank should hedge its operational risk for these units by putting aside
944.60 units of capital (1.39 x 211.75 + 1.30 x 218.40 + 1.70 x 213.50) for any one year period.

14

Although in this illustrative example unexpected losses above the combined VaR level (30 units
of capital) occur with probability 2.5% per annum, unexpected operational risk losses will exceed
this capital sum with probability less than 0.5%. In practice lower tail probabilities might be
chosen, but similar or higher probability ratios would obtain.
Note that in this example the loss data for each business unit was generated as independent and
the total capital figure takes the resulting diversification effect into account. On actual loss data
the dependencies in the realized data are taken into account by the method and the diversification
effect of the result can be analyzed by estimating each unit separately and adding the individual
capital figures (which conservatively treats losses across units as perfectly correlated) [10].
Although the results presented here are based on very large (2167) original sample sizes, the
simulation experiments on actual banking data reported in [10] verify the high accuracy of
MCMC Bayesian hierarchical methods for exceedance sample sizes as low as 10 and 25, as in
this example.

15

Figure 3: Simulation results of MCMC for 200k iterations

40
30
20

0.04

0.08

results[(3 * j) - 2, ]

50

Plot for parameter mu vs no.ofloops

0.0

density(results[(3 * j) - 2, ])$y

Density plot for parameter mu

20

30

40

50

200

400

600

1000

3.5
3.0
2.5

0.4

0.8

results[(3 * j) - 1, ]

1.2

4.0

Plot for parameter logsig vs no.ofloops

0.0

density(results[(3 * j) - 1, ])$y

Density plot for parameter logsig

log1

800

counter

density(results[(3 * j) - 2, ])$x

2.5

3.0

3.5

4.0

200

800

1000

60
50
40
30

sig

0.04

10

0.0

20

0.02

probability

600

Plot for parameter sig vs no.loops

0.06

Density Plot of sig

400
counter

density(results[(3 * j) - 1, ])$x

20

30

40

50

60

200

400

600

800

sig

no.loop

Density Plot of xi

Plot for parameter xi vs no.loops

1000

1.5
xi

0.5

1.0

1.0
0.5
0.0

probability

1.5

2.0

10

0.0

0.5

1.0

1.5

xi

2.0

200

400

600
no.loop

16

800

1000

Conclusion
In this chapter we have introduced Markov chain Monte Carlo (MCMC) concepts and techniques
and shown how to apply them to the estimation of a Bayesian hierarchical model of
interdependent extreme operational risks. This model employs the peaks over threshold (POT)
model of extreme value theory (EVT) to generate both frequency and severity statistics for the
extreme operational losses of interdependent business units which are of interest at the board
level of a financial institution. These are obtained respectively in terms of Poisson exceedences
of an unexpected loss level for other risks and the generalized Pareto distribution (GPD).
The model leads to annual business unit capital allocations for unexpected extreme risks which
take account of the statistical interdependencies of individual business unit losses.
The concepts discussed in this chapter are illustrated by an artificially created example involving
three business units but actual banking studies are described in [10] and in forthcoming work
relating to internally collected operational loss data.

References
[1] Beichl I., Sullivan F. (2000). The Metropolis algorithm, Computing in Science & Engineering
2(1): 6569.

[2] Casella G., Edward I.G. (1992). Explaining the Gibbs sampler, The American Statistician 46:
167-17.
[3] Castillo E. (1988). Extreme Value Theory in Engineering. Academic Press, Orlando.
[4] Chib S., Greenberg E. (1995). Understanding the Metropolis-Hastings algorithm, The

American Statistician 49(4): 327335.

[5] Dongarra J., Sullivan F. (2000). The top 10 algorithms, Computing in Science and

Engineering 2(1): 2223.

[6] Embrechts P., Kluppelberg C. & Mikosch T. (1997). Modelling Extremal Events, Springer,
Berlin.
[7] Galambos J. (1978). The Asymptotic Theory of Extreme Order Statistics, Wiley, New York.
[8] Hastings W. (1970). Monte Carlo sampling methods using Markov chains and their
applications, Biometrika 57(1): 97109.
[9] Medova E.A. (2001). Operational risk capital allocation and integration of risks, in Advances

in Operational Risk: Firmwide issues for financial institutions. Risk Books pp.115-127.

17

[10] Medova E.A., Kyriacou M.N. (2002). Extremes in operational risk measurement, in Risk

Management: Value At Risk And Beyond, M.A.H. Dempster, ed., Cambridge University
Press, pp. 247-274.
[11] Metropolis N., Rosenbluth A., Rosenbluth M., Teller A., Teller E. (1953). Equations of state
calculations by fast computing machines, Journal of Chemical Physics 21(1): 10871092.
[12] Muller P. (1993). A generic approach to posterior integration and Gibbs sampling,
Technical Report, Purdue University.
[13] Roberts G., Gelman A., Gilks W. (1994). Weak convergence and optimal scaling of random
walk Metropolis algorithms, Technical Report, University of Cambridge.
[14] Raiffa H., Schlaifer R. (1961). Applied Statistical Decision Theory, Harvard University
Press.
[15] Leadbetter M., LindGreen G., Rootzen H. (1983). Extremes and Related Properties of

Random Sequences and Processes, Springer, Berlin.

[16] Leadbetter M. (1991). On a basis for Peaks over Threshold modeling, Statistics and

Probability Letters 12: 357-362.

[17] Smith A, Roberts G. (1993). Bayesian computation via the Gibbs sampler and related
Markov chain Monte Carlo methods, J. Royal Statistical Society, B55: 3-23.
[18] Smith R. (2001). Measuring risk with extreme value theory, Chapter 8 in Risk Management:

Value At Risk And Beyond, M.A.H. Dempster, ed., Cambridge University Press.
[19] Tierney L. (1994). Markov chains for exploring posterior distributions, Annals of Statistics
22:17011762.

[20] The New Basel Capital Accord (2001). Bank of International Settlements Press Release.

18