Monte Carlo Simulations

Dr HK Pradhan
XLRI Jamshedpur
 Asset Price Regimes

Fixed deterministic stochastic

 Return Generating Process
• Variance of the daily returns isn’t fixed but rather a
random sample from a Normal Distribution(randomize
volatility scaled by the sqrt of time)
 Return generation sequence

 
 S 
• Return rt  ln  t 
 S 
 0 

ln( St )  ln( S 0 )  (  * t ,  t )  rt
 &  are respectively the mean and the variance
rt 1  t   t  t 1 , with  t 1 ~ N (0,1)
The random variable  is an innovation term with (i.i.d.).
This is also known as a Wiener process because it is a random normal
variable
 Asset Prices
• Asset prices (exchange rates, interest rates, stock prices,
commodity prices.) reflect the complex interactions between
economic forces, market dynamics and investors' sentiments, all
of which are influenced by news, which is inherently
unpredictable and random
• They follow a time-continuous stochastic process that generates
outcomes which are like random innovations over time
• Since speculators have incorporated all available knowledge into
their trades, one expects that the result will be prices showing
unpredictable fluctuations, independent of their past history.
• The future value of an asset is determined by random factors, past
values are not so important (except its distribution parameters)
• This is what is modernly called a "Brownian motion" "random
walk" or, in certain contexts, a "martingale".
 Asset Prices as stochastic processes
• A univariate Brownian motion is defined as a stochastic
process satisfying
– The process is defined for times t ≥ 0
– Realizations are continuous functions of time t.
– These are assumed to be iid with mean 0 and variance t
– Random variables (independent)

• Basic Weiner process (i.e. Brownian motion) applied to changes in

asset prices can be represented by the following:
dS   d t
• We can generate simulated random return distributions and then
calculate risk (VaR)
 Brownian/Wiener Process
• Robert Brown observed in
1827 how particles
suspended in water moved
erratically on a microscopic
scale.
• Louis Bachelier was the
first to use Brownian motion
to model securities in his
1900 thesis(later formalized
by Norbert Wiener).
 Brownian motion & Wiener process
• We consider a variable Δz (dS), whose value changes
continuously
• The change Δz occurs in a small interval of time Δt
• The variable Δz follows a Wiener process:

z   t where  is  (0,1)
• The values of Δz for any 2 different (non-overlapping) periods
of time are independent.
• i.e. the Wiener process has a drift rate (i.e. average change per
unit time) of 0 and a variance rate of 1.
• If assets have different degree of randomness or volatility, we
can incorporate a vol term:
dS    dt
 Simulating Asset Prices
• In a risk neutral world the process for a stock price is (by
simulating a path by choosing time steps of length t )

S t  S 0 * exp (  t   t )
• where  is a random sample from N(0,1)
• Risky assets have a positive expected return, and a random
volatility component
•
 Simulating Asset Prices
S t  S 0 * exp (  t  Z t )
St  35 * exp (0.15 * 0.004  0.3 * 0.1165 * 0.004 )  35 .0985
15

10

10 20 30 40
-5

-10

-15
 Asset Prices Formulations

• Changes in asset prices are continuous,

which can be represented by a basic dS   dt
Weiner process (i.e. Brownian motion):

• Different assets have different degree of dS    dt

randomness or volatility
• Prices can have a random
component, with or without drift, i.e. dS   d t    d t
the risky assets have a positive
expected return

• Asset prices are not independent of

the level of asset price (Ito Process) dS   S d t   S  d t
 Monte Carlo Simulations for VaR
• Price paths are represented by
the parameters , & , and a
random component (Wieners
Process)
• Determine , &  of returns
• Generate the standard normal
variable: N(0,1)
• Simulate the desired price path
• Generate portfolio value changes
• Set the desired percentiles
• Calculate Var & expected
shortfalls
 Montecarlo Steps
Get daily return data Find the distribution which
as input best fits the distribution of
daily returns

Step 1 Step 2

Simulate daily returns

with the same Find VaR by reading it off
distribution that best the distribution of generated
fits the input data simulated returns

Step 3 Step 4
 Generating Random Numbers
• Excel’s Pseudo-random-number generator function
produces uniformly distributed random numbers:

=RAND()
– Uniform (0, 1) distribution falling between interval 0 & 1
– (what happens if you multiply 10 & add 50 to each?)

• Generating Normally Distributed Random Numbers

=NORMSINV(RAND(), 0,1)

• Convert Uniform to Normally distributed numbers (with =0, =1)

 VaR Comparisons
• Parametric
Estimates VaR with equation that specifies
parameters such as volatility, correlation,
delta, and gamma as input.
• Historical simulation
Estimates VaR by reliving history; takes actual
historical rates and revalues positions for
each change in the market.
• Monte Carlo simulation
Estimates VaR by simulating random
scenarios and revaluing positions in the
portfolio.
 Two Asset Case
Bivariate Normal Distribution
• We are talking about generating data pairs
(X,Y) from a bivariate normal distribution
with the following:
• X~N( µx, σx)
• Y~N( µy, σy)
• Correlation ρxy.
• Need to create a correlation structure for
random variables(Gaussian Copula )
Close Automatic recalculation Tools>Options>Calculation>Manual
 Two Asset Case
• Need to generate Portfolio Var, in a Two Asset Case
• We need to define a covariance Metrix of two normal variable X1
and X2
– Need to transform a vector of uncorrelated (i.e. independent)
normally-distributed random variates into a vector of correlated
(i.e. dependent) normally-distributed random variates
– The correlated random variates can be used in a Monte Carlo
simulation
• We transform the variable X1 to a new variable Z1 that has a
standard normal distribution on a “percentile-to-percentile” basis.
• We transform the variable X2 to a new variable Z2 that has a
standard normal distribution on a “percentile-to-percentile” basis.
• Z1 and Z2 are assumed to have a bivariate normal distribution
17
 Cholesky decomposition
[Link]

• Cholesky decomposition is a procedure that creates an

empirical distribution with random normals
• With normal variables with zero mean and unit variance, by
Cholesky decomposition, we can generate new multivariate
normal samples with zeros mean and covariance
matrix . An example with n = 3 is shown below:

• Here the covariance between different return series is

retained by Cholesky decomposition.
 Cholesky decomposition
[Link]

• Using Cholesky decomposition let us create a Lower

triangular matrix of a bivariate distribution
• Lets asume the two standard normal random variates with
correlation Rho, with the correlaiton matrix is

• If we apply Cholesky decomposition to this matrix , we get

• Let us define uncorrelated random normal numbers

• We convert these into correlated numbers as follows

• or, or,
 Steps
• Use Excel Random Number Generator to generate two
independent samples e & e from a univariate standard
1 2
normal distribution
• Calculate 1, 2
• Calculate an intermediate value(incorporating the correlation
structures)
• Standardize them by subtracting their sample means and
dividing by their sample sd.
• Next, linearly transform these standardized random variables
into two random variables x1 & x2 while maintaining the
correlation structure

We now have two random variables x1 and x2, such that

x1  N (  x ,  x ), x2  N (  y ,  y ) with  xy
 Multivariate Normals
• Cholesky’s decomposition can be used to generate the
required Variance-Covariance matrix of the
multivariate random normal samples
 12  12 
V  2  is the Co-variance Matrix
 12 2 
• If

• and if V=LU, where L is the Cholesky

• Then L(z) has the covariance V
 So by now we have three methods
• Historical simulation relies on history to forecast
what could happen in the future.
• Monte Carlo simulation does not necessarily rely on
history to forecast what could happen in the future.
• Monte Carlo simulation will most likely be more
accurate in estimating the risk in newly developed
capital markets.
• If extensive historical data is available, historical
simulation is more accurate than Monte Carlo
simulation
Can you comment on the three methods
of VaR that you have computed

MC shows lower than

HS, but higher
than Parametric

Historical simulation
shows most
conservative figures

© RiskMetrics
What are the limitations of HS?

– Sensitive to lag length chosen

– Problems when data availability for very
shorter periods
– Assumes no change in volatility or correlations
through time, or the distribution is stationary
 Questions
• Which method of VaR is easy to calculate?
• What are the advantages of different approaches?
• What are the disadvantages?
 VaR Measurement Summary
• Parametric • Limitations
• Very fast • Constrained by Normal
• Immediate convergence assumption
• Linear approximation
• Historical Simulation
• No distributional assumption • Highly dependant on data period
• Full re-pricing • Will be influenced by extreme
events
• Monte Carlo Simulation
• Provides full range of potential • Depends on the choice of model
outcomes • Simulation error (can be
• Incorporates random minimized by larger trials)
components of stock prices
• powerful and flexible (non-
linear, not normal)
THANK YOU