Professional Documents
Culture Documents
Lecture 31
Here VaRt−1 and sVaRt−1 are the VaR and Stressed VaR (10-day 99%),
and VaRavg and sVaRavg are the average of VaR and Stressed VaR (10-day
99%) over the previous 60 days. mc , ms ≥ 3 are Multiplicative Factors.
2 Incremental Default Risk Charge: Capital Requirements making use of the
maximum of capital using “Trading Book Calculation” and “Banking
Book Calculation”.
3 The Comprehensive Risk Measure: Designed to take account of risks in
what is known as correlation book and is applicable to instruments such
as ABS and CDO, which are sensitive to the correlation between the
default risks of different assets.
Basel III: The Highlights
Basel-III regulations are currently in force, which encompasses capital
requirement for Credit Risk and Market Risk and regulates six points:
1 Capital Definitions and Requirements.
2 Capital Conservation Buffer.
3 Countercyclical Buffer.
4 Leverage Ratio.
5 Liquidity Risk.
6 Counterparty Credit Risk.
Credit Risk
1 Credit Risk arises from a default, the most common examples of which are
a lender failing to repay the loan to a bank or a bond issuer failing to
make the coupon and/or face value payment to the bond holder.
2 What is referred to as Non-Performing Asset (NPA) is a consequence of
default and qualifies under Credit Risk.
3 The first basic model for Credit Risk is the Firm-Based models, which are
modelled in the classical Black-Scholes-Merton framework where the
assets of an entity are composed of equities and debts. The advantage of
this approach is that one can adopt the results for the
Black-Scholes-Merton framework, for which the Nobel Prize in Economics
was awarded in 1997.
4 The other model is the Intensity models, which are driven by stochastic
processes, such as homogeneous and in-homogeneous Poisson process, as
well as the Gamma Ornstein-Uhlenbeck process.
WCDR and PD
1 The “Worst-Case-Default-Rate” at 99.9% confidence level, for the i-th
counterparty is given by:
" √ #
N −1 (PDi ) + ρN −1 (0.999)
WCDRi = N √ .
1−ρ
3 The VaR will be the appropriate percentile of the values of vn+1 arranged
in descending order.
Operational Risk
1 Operational Risk encompasses a broad range of risk resulting from failure
in the internal process, people or system of a financial institution, or
resulting from external events.
2 While the high frequency losses can be modelled using parametric
distributions, the low frequency (but high intensity) losses have to be
modelled using the statistical approach of Extreme Value Theory (EVT).
3 Loss frequency can be modeled with Poisson distribution, where the
(λT )n
probability of n losses in T years is given by: e −λT .
n!
4 Suppose the number of losses in one year is N.
5 Then the loss severity for each can be generated using log-normal
distribution, with the losses being L1 , L2 , . . . , LN .
N
X
6 The total loss experience is then given by: Li .
i=1
Price Risk of Bonds and Macaulay Duration
1 The price risk of bonds depends exclusively on interest rate.
2 The approach we will use is to consider bonds as derivatives, whose
underlying variable is the interest rate.
3 A peculiarity of this underlying variable is that it is not directly traded in
the markets (unlike the underlying of plain vanilla stock options, for
example).
4 We consider the main risk measure for bonds, namely, Duration.
5 We explore the notion of Macaulay Duration for measuring interest-rate
risk, introduced in 1938.
6 Although it is based on several restrictive assumptions, its simplicity and
the fact that it captures the main factor that affects the bond price
volatility makes it the fundamental tool for bond portfolio management.
Duration: Definition and Interpretation
1 We first focus on the simplest case. Consider a risk-free pure discount
bond that will pay a nominal amount (say, 100) in exactly one period (for
example, one year) from today.
2 The one-year spot rate is r . Therefore, the price P of this bond is:
100
P= .
1+r
3 An investor who holds the bond only has to worry about one type of risk,
namely, a possible change (more explicitly, an increase) in the interest rate.
4 Therefore, the natural measure of the risk of holding the bond is the
derivative of the bond price with respect to the interest rate:
dP 100 P
=− =− .
dr (1 + r )2 1+r
5 We see that the effect on the price of the bond, of an increase in the level
of interest rate will be larger (in absolute value), the larger the price of
the bond.
Duration: Definition and Interpretation (Contd ...)
1 We can easily extend the same type of analysis to a pure discount bond
that has several periods (say, T ) to maturity. Suppose that the T -period
spot rate is r . Then the price of the bond is:
100
P= .
(1 + r )T
which is similar to the previous equation, with the additional factor of the
time left to maturity.
3 That is, the price of a bond is more sensitive to a change in interest rates,
if its maturity is longer.
Duration of Coupon Bonds
1 Consider a bond that pays the coupons of size Ci , at the end of each
period, for T periods (we assume that the final coupon CT includes the
principal).
2 Suppose further that the corresponding interest rates are denoted by
ri , i = 1, 2, . . . , T so that the price of the bond is:
T
X Ci
P= .
i=1
(1 + ri )i
3 Comparing the equations, it is clear that the yield y depends only on the
rates ri that affect the price of the bond.
4 For example, if all the rates ri go up, the yield y also goes up. Then,
similar to our analysis of risk for pure discount bonds, we consider the
derivative of the coupon bond with respect to its yield:
T T Ci
dP X Ci P X (1+y )i
=− i× = − i ,
dy i=1
(1 + y )i+1 1 + y i=1 P
where the second equality is simply the result of multiplying and dividing
by P.
Macaulay Duration (Contd ...)
1 The second factor of the equation is called Duration or Macaulay
Duration.
2 Definition: The duration of a bond is a measure of the sensitivity of the
bond to interest rate movements.
3 More explicitly, the duration D of a bond that pays coupons Ci and has
yield y and maturity T is given by:
T Ci
(1+y )i
X
D := i .
i=1
P
5
X 8 100
P = + = 112.99,
i=1
(1.05)i (1.05)5
5 8 100
X (1.05)i (1.05)5
D = i +5 = 4.36.
i=1
112.99 112.99
2 Now consider the above example with 30 years left until maturity. Then:
30
X 8 100
P = + = 146.12,
i=1
(1.05)i (1.05)30
30 8 100
X (1.05)i (1.05)30
D = i + 30 = 14.82.
i=1
146.12 146.12
3 Finally consider the previous example, but with an yield of 10% instead of 5%. Then:
30
X 8 100
P = + = 81.15,
i=1
(1.1)i (1.1)30
30 8 100
X (1.1)i (1.1)30
D = i + 30 = 10.65.
i=1
81.15 81.15
MA 592: Mathematical Finance
Lecture 34
in cash.
5 In order to guarantee that we will be able to meet the payment in two
years (approximately), we have to invest the money on a portfolio of bonds
with the duration equal to the duration of the portfolio of short positions.
6 In this case, there is only one liability with maturity in two years,
equivalent to a short position in a pure discount bond, with maturity in
two years and therefore, the duration is equal to two years.
7 We want to invest in a portfolio with duration of two, as well.
Hedging by Immunization: An Example (Contd ...)
1 There are only two pure discount securities, with maturities (and,
therefore, durations) equal to one and four, respectively.
2 We now take advantage of the property:
If a bond portfolio comprises of n bonds with weights w1 , w2 , . . . , wn and
durations of D1 , D2 , . . . , Dn , respectively, then the duration of the
portfolio is given by:
X n
D= wi Di .
i=1
1 ∂2P
C := .
P ∂y 2
∂P 1 ∂2P D 1
∆P = ∆y + (∆y )2 = − P∆y + CP(∆y )2 .
∂y 2 ∂y 2 1+y 2
Value-at-Risk (VaR)
1 Ideally we look for a number (or a set of numbers) that expresses the
potential loss with a given level of confidence, enabling the risk manager
to adjudge whether the risk is acceptable or not.
2 Value-at-Risk (VaR) is a standard benchmark for measuring financial risk
a
.
a The main drawback, however, is that it does not provide information about the
Quantiles
1 Let (Ω, F, P) be a probability space and let X : Ω → R be a random
variable.
2 The cumulative distribution function FX : R → [0, 1], defined by:
qα (X ) = inf{x|α ≤ FX (x)},
Proof of Lemma 1
Since the given condition on FX ensures that it is invertible, the inverse
function α → FX−1 (α) is continuous and α < FX (x) is equivalent to
FX−1 (α) < x. This gives:
Lemma 2
Let X be a random variable. If f : R → R is right-continuous and
non-decreasing, then:
q α (f (X )) = f (q α (X )).
Measuring Downside Risk
1 We work on a single step financial market model, in which we invest at
time t = 0 and terminate our investment at time t = T .
2 We denote by X , the discounted value of the position at time T .
3 Definition: For α ∈ (0, 1), we define the Value-at-Risk (VaR) of X at
confidence level 100(1 − α)% as:
Theorem 1
N
X
Assume that X is a discrete random variable with P(X = xi ) = pi , pi = 1
i=1
and x1 < x2 < · · · < xN . Then:
VaR α (X ) = −xkα ,
kα−1
X
where kα ∈ N is the largest number such that pi ≤ α.
i=1
Proof of Theorem 1
1 Since X has discrete distribution and x1 < x2 < · · · < xN , we can see that:
k
X
P(X ≤ xk ) = pi .
i=1
( k
) ( k−1
)
X X
2 We shall use the fact: min k|α < pi = max k| pi ≤ α .
i=1 i=1
3 Therefore:
VaR α (X ) = −q α (f (Z )) = −f (q α (Z )) = −f (N −1 (α)),
−1
h i
= S(0) 1 − e m−rT +σN (α) .
Example on Calculating VaR Using Historical Simulation
1 Suppose that an investor owns, on September 25, 2008, a portfolio worth
$10 million consisting of investments in four stock indices:
(A) Dow Jones Industrial Average (DJIA) in the US.
(B) Financial Times Stock Exchange (FTSE) 100 in the UK.
(C) Cotation Assistée en Continu (CAC) 40 in France.
(D) Nikkei 225 in Japan.
2 The value of the investment in each index on September 25, 2008, is
shown in the following Table.
Table: Scenarios generated for September 26, 2008, using the data in the
previous Table
Example on Calculating VaR Using Historical Simulation (Contd ...)
1 Scenario 1 (the first row in the previos Table shows the values of market
variables on September 26, 2008, assuming that their percentage changes
between September 25 and September 26, 2008, are the same as they
were between August 7 and August 8, 2006.
2 Scenario 2 (the second row in the previous Table) shows the value of the
market variables on September 26, 2008, assuming these percentage
changes are the same as those between August 8 and August 9, 2006 and
so on.
3 In general, Scenario i assumes that the percentage changes in the indices
between September 25 and September 26, 2008 are the same as they were
between Day (i − 1) and Day i for 1 ≤ i ≤ 500.
4 The 500 rows of the previous Table are the 500 scenarios considered.
5 The DJIA was 11,022.06 on September 25, 2008. On August 8, 2006, it
was 11,173.59, down from 11,219.38 on August 7, 2006.
6 Therefore the value of the DJIA under Scenario 1 is:
11, 173.59
11, 022.06 × = 10, 977.08.
11, 219.38
Example on Calculating VaR Using Historical Simulation (Contd ...)
1 Similarly, the values of the FTSE 100, the CAC 40 and Nikkei 225 are
5,187.46, 4,236.71 and 12,252.62, respectively.
2 Therefore the value of the portfolio under Scenario 1 is (in $000s) is:
10977.08 5187.46 4236.71 12252.62
4, 000× +3, 000× +1, 000× +2, 000× = 10, 021.502.
11022.06 5197.00 4226.81 12006.53
Ni
2 So we expect: f (xi ) = .
M∆x
Ni
3 Accordingly, plotting should give an approximation to the density
M∆x
function values of f (xi ).
Monte-Carlo Method
1 Recall: Valuation of an European option reduces to the computation of an
expectation.
2 We will accomplish this through the usage of pseudo-random numbers.
3 To begin with, we consider the case of a general random variable X ,
whose expected value E (X ) = a and variance Var (X ) = b 2 are not known.
4 Suppose:
(A) We are interested in computing and approximation to “a” (and
possibly “b”).
(B) We are able to take independent samples of X using a
pseudo-random number generator.
5 Let X1 , X2 , . . . , XM denote independent random variables with the same
distribution as X .
M
1 X
6 Then, we might expect: aM := Xi .
M i=1
7 It can be shown that aM is an unbiased estimate of E (X ).
Monte-Carlo Method (Contd ...)
M
2 1 X
1 To estimate Var (X ) we can use bM := (Xi − aM )2 .
M − 1 i=1
2
2 It can be shown that bM is an unbiased estimate of Var (X ).
M
X
3 Now, by the Central Limit Theorem Xi behaves like N (Ma, Mb 2 )
i=1
random variate.
b2
4 So: aM − a is approximately N 0, .
M
5 Equivalently: aM − a is approximately N (0, 1) random variable scaled by
b
√ .
M
1.96b 1.96b
6 This leads to: P a − √ ≤ aM ≤ a + √ = 0.95 a .
M M
a If Z ∼ N (0, 1), then P(µ − 1.96σ ≤ Z ≤ µ + 1.96σ) = 0.95
Monte-Carlo Method (Contd ...)
1.96b 1.96b
1 We thus have: P aM − √ ≤ a ≤ aM + √ = 0.95.
M M
2 Replacing the unknown b by the approximation bM , we see that the
1.96b 1.96b
unknown expected value a lies in the interval aM − √ , aM + √ ,
M M
with probability 0.95.
3 In other words, this gives an approximate 95% confidence interval for “a”.
4 This analysis leads us to the basic Monte-Carlo method for approximating
“a”.
5 We compute M independent samples and determine aM .
6 In order to monitor the error, we also compute the variance approximation
2
“bM ”.
Monte-Carlo for European Option Valuation
1 We are now in a position to use Monte-Carlo for option valuation.
2 We consider the payoff at time T of the form f (S(T )).
3 Then the time 0 price of the option is given by e −rT E∗ (f (S(T ))).
4 Recall that: 1 2
S(T ) = S(0)e (r − 2 σ )T +σW (T ) .
5 Consequently, the price of the option is given by:
1 2
√
e −rT f S(0)e (r − 2 σ )T +σz T , z ∼ N (0, 1).
Monte-Carlo for European Option Valuation (Contd ...)
1 Then the resulting Monte-Carlo algorithm can be summarized as:
for i = 1 : M
Compute an N (0, 1) sample ξi
√
1 2
Set Si = S(0) exp µ − σ T + σ T ξi
2
−rT
Set Vi = e f (Si )
end
M
1 X
Set aM := Vi
M i=1
M
2 1 X
Set bM := (Vi − aM )2 .
M − 1 i=1
2 The output provides an approximate option price aM and an approximate
95% confidence interval.
MA 592: Mathematical Finance
Lecture 37
∂V (S, t) V (S + h, t) − V (S, t)
= + O(h), h → 0.
∂S h
3 Choosing a small value of h, we obtain the finite difference approximation:
∂V (S, t) V (S + h, t) − V (S, t)
≈ + O(h).
∂S h
4 This produces an approximation to Delta, at a single point, based on
option values at two points with slightly different arguments.
Monte-Carlo for Greeks (Contd ...)
1 Using the risk-neutral discounted expected payoff, we can approximate the
time-zero Delta via Monte-Carlo estimation, as follows:
E∗ (f (S(T ))|S(0) = S0 + h) − E∗ (f (S(T ))|S(0) = S0 )
e −rT .
h
2 Observation:
Error for each of the two Monte-Carlo estimation is
1
O √ and hence after dividing the difference by h, the overall error
M
1
can be expected to be √ .
h M
3 Hence: To keep a good approximation for small h, we need to make M
large (more samples will be required than the basic Monte-Carlo
approach).
Monte-Carlo for Greeks (Contd ...)
1 Note: The two random variables in the two expectations are clearly
“correlated”.
2 Implication: For any particular asset path, the payoff starting from
S(0) = S0 is likely to be close to the payoff starting from S0 + h.
3 Intuitively: The corresponding sample mean errors shall be close, provided
that we use the same paths for the two simulations.
4 We thus apply Monte-Carlo to the equivalent problem:
for i = 1 : M
Compute a N (0,
1) sample ξ i
√
1 2
Set Si = S0 exp µ − σ T + σ T ξi
2
√
h 1
Set Si = (S0 + h) exp µ − σ 2 T + σ T ξi
2
e −rT f (Sih ) − f (Si )
Set ∆i =
h
end
M
1 X
Set aM := ∆i
M i=1
M
2 1 X
Set bM := (∆i − aM )2 .
M − 1 i=1
Monte-Carlo for Exotic Options
1 The basic Monte-Carlo for European option can easily be extended to
handle path dependency.
2 Accordingly, an extra step is involved, in terms of the temporal grid points
T −0
tj = j∆t, for 0 ≤ j ≤ N, such that ∆t = .
N
3 Then the asset price path is given by:
√
1
Sj+1 = Sj exp r − σ 2 ∆t + σ ∆tξj , ξj ∼ N (0, 1).
2
Example of Exotic Option: Algorithm for Up-And-Out Barrier Call Option
for i = 1 : M
for j = 0 : N − 1
Compute an N (0, 1)
sample ξj
√
1
Set Sj+1 = Sj exp r − σ 2 ∆t + σ ∆tξj
2
end
Set Simax = max Sj
0≤j≤N
If Simax < B
Set Vi = exp(−rT ) max(SN − K , 0)
Else Set Vi = 0
end
M
1 X
Set aM := Vi
M i=1
M
2 1 X
Set bM := (Vi − aM )2 .
M − 1 i=1
Example of Exotic Option: Algorithm for Floating Strike Lookback Call Option
for i = 1 : M
for j = 0 : N − 1
Compute an N (0, 1)
sample ξj
√
1
Set Sj+1 = Sj exp r − σ 2 ∆t + σ ∆tξj
2
end
Set Simin = min Sj
0≤j≤N
Set Vi = exp(−rT ) max(SN − Simin , 0)
end
M
1 X
Set aM := Vi
M i=1
M
2 1 X
Set bM := (Vi − aM )2 .
M − 1 i=1
Example of Exotic Option: Algorithm for Average Price Asian Put Option
for i = 1 : M
for j = 0 : N − 1
Compute an N (0, 1)
sample ξj
√
1
Set Sj+1 = Sj exp r − σ 2 ∆t + σ ∆tξj
2
end
X N
Set Smeani = ∆t Sj
j=1
Set Vi = exp(−rT ) max(K − Smeani )
end
M
1 X
Set aM := Vi
M i=1
M
2 1 X
Set bM := (Vi − aM )2 .
M − 1 i=1
Some Simulation Results: Pricing of Exotic Options
1 We ran the simulations for the four different path dependent options using
MatLabTM (Code: Exotic_Options.m).
2 The number of sample paths and time intervals taken are M = 5000 and
N = 1000 respectively.
3 The parameters used were S(0) = 100, r = 6%, σ = 30%, K = 120 and
B = 250.
4 The expiration time was T = 1 and hence ∆t = T /N = 10−3 .
Option Type Price Variance 95% CI
European put 20.3997 364.4944 [19.8705,20.9289]
Up-and-out barrier call 6.8101 239.8059 [6.3808,7.2393]
Floating strike lookback call 23.7967 532.7144 [23.1570,24.4365]
Average price Asian put 17.8306 187.0266 [17.4515,18.2097]
4 Recall
that{Ui } are i.i.d. normal random variates with mean
1 2
µ − σ ∆t and variance σ 2 ∆t.
2
5 Thus, collating historical asset price data and obtaining the values of
log-returns (Ui ’s) is equivalent to sampling from:
1
N µ − σ 2 ∆t, σ 2 ∆t .
2
Parametric Estimation Using Historical Data (Contd ...)
1 Suppose that the current time is tn .
2 Let the M + 1 most current asst prices be
S(tn−M ), S(tn−M+1 ), . . . , S(tn−1 ), S(tn ).
3 Using the corresponding log-return data {Un+1−i }M
i=1 , the sample mean
and the sample variance estimate become:
M
1 X
(A) aM := Un+1−i .
M i=1
M
2 1 X
(B) bM := (Un+1−i − aM )2 .
M − 1 i=1
4 Thus we may estimate µ and σ from:
bM
(A) σ 2 ∆t = bM2
⇒ σ = σ∗ = √ .
∆t
2
aM + 21 bM
1
(B) µ − σ ∆t = aM ⇒ µ = µ∗ =
2
.
2 ∆t
Accuracy of the Sample Variance Estimator
bM
1 In order to get idea about the accuracy of the estimate σ ∗ = √ , we
∆t
need to recognize the fact that we are essentially using Monte-Carlo
2
simulation to compute bM as an approximation to the expected value of
2
variable(U − E (U))
the random , where
1
U∼N µ − σ 2 ∆t, σ 2 ∆t .
2
2 Equivalently, after dividing throughout by ∆t, we are using Monte-Carlo
b2
simulation to compute σ ∗2 = M as an approximation to the expected
∆t 2
value of the random variable Ub − E (U)b , where
√
1
Ub ∼ N µ − σ2 ∆t, σ 2 .
2
Accuracy of the Sample Variance Estimator (Contd ...)
1 Hence, an approximate 95% confidence interval for σ 2 is given by:
1.96v
σ ∗2 ± √ ,
M
2
where v 2 is the variance of the random variable Ub − E (U)
b .
6 Now:
√ √ Z1 √ √
Ui 1−Ui
(A) E e e = e x+ 1−x dx.
0
√ √ Z1 √ √
(B) Cov e Ui , e 1−Ui = e x+ 1−x dx − 22 .
0
Antithetic Variates: Example of Uniform Distribution (Contd ...)
1 Thus:
Z1
e2 15 1 √ √
x+ 1−x
Var (Ybi ) = − + + e dx.
4 4 2
0
3 Hence:
Var (Yi )
≈ 181.2485.
Var (Ybi )
Analysis of the Uniform Case
1 To understand the antithetic variate technique works, consider the more
generalized case of approximating:
4 Thus:
f (Ui ) + f (1 − Ui ) 1
Var = Var (f (Ui )) + Cov (f (Ui ), f (1 − Ui )) .
2 2
Analysis of the Uniform Case (Contd ...)
1 The success of the new scheme is contingent on whether
1
Var (f (Ui ) + f (1 − Ui )) is smaller than Var (f (Ui )).
2
2 This, then boils down to making Cov (f (Ui , f (1 − Ui ))) as negative as
possible.
3 We want: f (Ui ) to be big (relative to its mean) when f (1 − Ui ) is small
(relative to the mean).
4 Intuitively: This approach will work when f is monotonic.
5 Loosely speaking: Antithetic variate attempts to compensate for samples
that are above the mean, by adding samples which are below the mean,
and vice versa.
6 We can now convert this intuition into a mathematical result.
7 (A) f is monotonically increasing: x1 ≤ x2 ⇒ f (x1 ) ≤ f (x2 ).
(B) f is monotonically decreasing: x1 ≤ x2 ⇒ f (x1 ) ≥ f (x2 ).
(C) If f and g are both monotonically increasing or both monotonically
decreasing functions, then (f (x) − f (y ))(g (x) − g (y )) ≥ 0.
Lemma
If f and g are both monotonic increasing functions or both monotonic
decreasing function, then for any random variable X :
Cov (f (X ), g (X )) ≥ 0.
3 Since the N (0, 1) distribution is symmetric about the origin, rather than
1
about , the antithetic estimate used is “−Zi ” rather than “1 − Zi ”.
2
4 Further “−Zi ” is also a N (0, 1) random variable.
5 The analysis done earlier can now be repeated to give:
f (Zi ) + f (−Zi ) 1
Var ≤ Var (f (Zi )),
2 2
when f is monotonic.
MA 592: Mathematical Finance
Lecture 39
Z = X + E (Y ) − Y
satisfies:
E (Z ) = E (X ) + E (Y ) − E (Y ) = E (X ).
Variance Reduction by Control Variates (Contd ...)
1 Hence we can apply Monte Carlo to Z instead of X .
2 In this context Y is called the “control variate”.
3 Further:
Var (Z ) = Var (X − Y ).
4 Hence to get some benefit from this approach, we would like X − Y to
have a smaller variance than X , which is what is meant by “close”.
5 Note: It may be more expensive to sample Z than to sample X .
6 If Var (Z ) = R1 Var (X ) for some R1 < 1 and the cost of sampling Z is R2
times that of sampling X , then we get an overall gain in efficiency, if
R1 R2 < 1.
Variance Reduction by Control Variates (Contd ...)
1 Generalized case:
Zθ = X + θ(E (Y ) − Y ),
for any θ ∈ R.
2 Since we still have E (Zθ ) = E (X ), therefore we may apply Monte-Carlo to
Zθ .
3 Accordingly:
Cov (X , Y )
θmin = .
Var (Y )
5 It can be shown that Var (Zθ ) < Var (X ) if and only if θ lies between 0
and 2θmin .
Variance Reduction by Control Variates (Contd ...)
1 On a general problem we typically do not know Cov (X , Y ) and hence
cannot find θmin .
2 However, it is possible to estimate Cov (X , Y ) and hence θmin during a
Monte-Carlo simulation.
3 The name “control variate” comes from the fact that the term
“E (Y ) − Y ” controls the Monte-Carlo process.
Variance Reduction by Control Variates (Contd ...)
1 Suppose: The covariance is positive, that is
E [(X − E (X ))(Y − E (Y ))] > 0 and θ > 0.
(A) In this case, when X > E (X ), then we would “expect” Y > E (Y ).
(B) Generally, adding the negative amount θ(E (Y ) − Y ) helps to correct
the overestimate of E (X ) from the sample of X .
2 Suppose: The covariance is positive, that is
E [(X − E (X ))(Y − E (Y ))] > 0 and θ > 0.
(A) In this case, when X < E (X ), then we would “expect” Y < E (Y ).
(B) Generally, adding the positive amount θ(E (Y ) − Y ) helps to correct
the underestimate of E (X ) from the sample of X .
3 A similar argument holds when Cov (X , Y ) < 0 and θ < 0.
Control Variates: Arithmetic Average Price Asian Call Option
1 Arithmetic Average Price Asian Call Option: The payoff is:
" n
#
1X
max S(ti ) − K , 0 .
n i=1
max S(ti ) − K , 0 .
i=1