mfdc9d

download mfdc9d

of 16

Transcript of mfdc9d

  • 7/27/2019 mfdc9d

    1/16

    9 Pricing Derivatives via Monte Carlo Simu-

    lation

    Suppose, we have a contract at date 0 that

    gives us the right but not the obligation to

    sell one share of a stock at date T for a

    price equal to the maximum during the pe-riod from 0 to T (a lookbackoption).

    In such a case the exercise price is random,

    depending on the path of St, 0 t T.

    Another example is the right to buy, say cur-

    rency, at an average rate during the contract

    period (an Asian option).

    1

    In these cases (particularly for the look back

    options) the risk neutral approach offers a

    considerably simpler alternative to the more

    traditional PDE.

    Consider the example of the right to sell at

    maximum (put option).

    The current price (day 0) of the put option

    (H) is

    H0

    = e

    rTEP max0tT St ST= erTEP

    max0tT St

    erTEP [ST]

    = erTEP

    max0tT

    St

    S0.

    (1)

    Remark 9.1: Why is erTEP[ST] =S0?

    2

  • 7/27/2019 mfdc9d

    2/16

    In the Monte Carlo simulationm (usually thou-

    sands) sample baths are generated under the

    risk neutral process, and the average of the

    maximums are calculated.

    The sample average is used to estimate

    EP

    max0tT St

    .

    Denote the average as Smax.

    Then

    H0=erTSmax S0.(2)

    3

    Discrete Approximation of a Continuous Pro-

    cess

    In practice the realizations of continuous pro-

    cess are evaluated at discrete time points.

    Usually, the solution of the SDE is not known

    explicitly, but must be discretely approximated.

    The GBM is one rare, though important,

    case where the solution of the SDE is known.

    That is, if

    dS=Sdt + SdW(3)

    then

    S=S0e(122)t+W,(4)

    and we can directly generate sample paths ofS (at discrete time points).

    4

  • 7/27/2019 mfdc9d

    3/16

    In the general case, consider the stochastic

    process

    dSt=a(St, t)dt + b(St, t)dWt.(5)

    Using the discrete approximation to calculate

    St, divide the sample interval [0, T] into n

    equal length subintervals

    0 =t0< t1< . . . < tn1< tn=T .(6)

    such that tk =tk tk1=T /n h.

    In the numerical approximations of St, the

    idea is to replace dSt with the discrete differ-

    ences St=Stk Stk1.

    5

    Euler Method

    The Euler scheme is:

    Stk+1 =Stk+ akh + bkWtk+1,(7)

    where ak =a(Stk, tk), bk = (Stk, tk), and

    Wtk+1 =Wtk+1 Wtk N(0, h)are independent for k = 0, 1, . . . , n 1.

    The procedure is started from a given S0

    .

    Due to the discretisation of dS, the error of

    the Euler method is of order O(h).

    6

  • 7/27/2019 mfdc9d

    4/16

    Milstein Method

    The Milstein scheme is:

    Stk+1 = Stk+ akh + bkWtk+1

    +12bkbk

    (Wtk+1)

    2 h

    ,(8)

    whereb =b(S, t) = b(S, t)/S, evaluated atStk.

    The order of error is O(h2).

    7

    Runge-Kutta Method

    Yet another scheme is the Runge-Kutta, which

    avoids calculations of the derivatives b(S, t).

    The scheme is:

    Sk = Stk+ akh + bkhStk+1 = Stk+ akh + bkWtk+1

    + 12

    h(bk bk)

    (Wtk+1)

    2 h

    ,

    (9)

    where bk =b(Sk, tk).

    Again the order of error is O(h2).

    8

  • 7/27/2019 mfdc9d

    5/16

    Which ever method is user, the key is to sim-

    ulate the sample paths. In this we can utilize

    the property

    Wtk+1 =

    hZk+1,(10)

    where Zk+1 are independent N(0, 1) random

    deviates.

    The (approximated) sample paths for St are

    generated by first generating n independent

    N(0, 1) random deviates, Z1, . . . , Z n, and use

    them in the equation of the selected method

    [(7) (9)] to calculate the (discretely approx-

    imated) realization for St.

    9

    In many software packages N(0, 1) random

    number generators are readily available.

    An easy, though inefficient, method in Excel

    to generate normal random deviates is to use

    the inverse function method, such that

    Z= 1(U),(11)

    where 1(U) (in Excel: = NormSinV(z)) isthe inverse of the standard normal cumula-

    tive distribution function, (z), andU U(0, 1)is a uniformly distributed random variable in

    the interval (0, 1) (in Excel: = rand()).

    10

  • 7/27/2019 mfdc9d

    6/16

    A far more efficient way is to use for exam-

    ple the Box-Muller method: Let U1 and U2be two independentU(0, 1) random variables,

    then

    Z1 =2 log(U1) cos(2U2)

    Z2 = 2 log(U1) sin(2U2)(12)

    are two independentN(0, 1) random variables.

    11

    Example 9.1: Consider the Cox-Ingersoll-Ross process

    used in valuation of interest rate derivatives.

    drt=( rt)dt + rtdWt,(13)where indicates the long run mean, indicates thespeed of the mean reversion, and is the volatilityparameter.

    Thus the drift and volatility parameters are

    a(r, t) =( r)and

    b(r, t) =

    r.

    12

  • 7/27/2019 mfdc9d

    7/16

    Let = .50, = 5%, and = 3. Furthermore, let

    T = 10 years, h= 1/120 (one month), and supposethat r0 = 5%.

    The Euler scheme is

    rtk+1 =rtk+ ( rtk)h +

    rtk

    hZk+1,(14)

    whereZk+1 are independent N(0, 1) random variables.

    In the Milstein approximation

    b =

    2

    r.(15)

    As seen from the figure below the differences between

    the approximations are at least in this case immate-

    rial.

    Cox-Ingersoll-Ross SDE

    dr = alpha (mu-r) dt + sigma sqrt(r) dW

    alpha = 0.5, mu = 5% (p.a.), sigma = 3%

    0

    1

    2

    3

    4

    5

    6

    7

    8

    9

    10

    0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

    105

    110

    115

    120

    month

    Interestrate

    Euler

    Milstein

    Runge-Kutta

    Figure 9.1: Euler, Milstein, and Runge-Kutta approx-

    imations of the CIR PDE.

    13

    Simulating the Geometric Brownian Motion

    The geometric Brownian motion (GBM) SDE

    (3) has the solution (4).

    Thus we do not need to use the approx-

    imations described above, but can directlysimulate the sample path of St at the time

    points tk, k = 0, 1, . . . n 1 with step lengthtk+1 tk =h, for example by the formula

    Stk+1 =Stke(122)h+

    hZk+1,(16)

    where Zk+1 are independent N(0, 1) random

    variables and S0 = St0 is given.

    The approximation can be made arbitrarily

    precise by increasing n (and therefore de-

    creasing h).

    Unfortunately there are no general rules how

    big n should be to make the approximation

    adequate.

    14

  • 7/27/2019 mfdc9d

    8/16

    Remark 9.2: If we are only interested about the end

    values of at the maturity T, as is the case with theEuropean options, we can directly simulate ST, when and are constants.

    This is because

    ST =S0e(1

    22)T+WT,(17)

    whereWT N(0, T), so that we can write WT = T Zwith Z N(0, 1).

    However, in the case of the path-dependent options

    we need to simulate the whole sample paths.

    Remark 9.3: If the parameters (usually volatility) istime varying and we are pricing European options, itis usually computationally more efficient to simulatefirst the returns

    Rtk = ( 1

    2tk )h + tk

    hZk,(18)

    whereZk N(0, 1) fork = 1, . . . , nand instead of (16)use

    ST =S0 exp

    nk=1

    Rtk

    .(19)

    Note that tk is non-anticipating.

    15

    Example 9.2: Consider the look-back option

    discussed above.

    Suppose that the current price of the stock is

    S0 = 90, the risk-free interest rate is r = 5%,

    = 25%, and the exercise day is 90 days

    from now.

    We simulate (16) with replaced by the risk-

    free rate, r = 0.05, (see Ch. 8).

    16

  • 7/27/2019 mfdc9d

    9/16

    For illustration purposes, let m= 10, then the below

    Excel sheet excerpt gives an estimate

    Smax= 98.78.(20)

    thus the price estimate is

    H0=e(0.0590/360)Smax 90 = 7.56.(21)

    Figure 9.2 Simulation example of a look-back option.

    17

    How Many Simulations

    The simulation sample size m can be deter-

    mined in the usual manner of determining

    sample size in a survey research.

    H0=erT1

    m

    m

    j=1 Y

    jn S0,(22)where Yjn = max0knSjk , and{Sjk}nk=0 isthe jth sample path.

    18

  • 7/27/2019 mfdc9d

    10/16

    The Central Limit Theorem (CLT) implies

    m

    H0 H0 a N0, 2y (n),(23)

    where

    2y (n) = Var

    erTYjn

    .(24)

    Therefore for large m an asymptotic 95%

    confidence interval for H0 is

    H0 1.96y(n)

    m .(25)

    19

    Hence, if, for example, we want a H0 that is

    within H0 $0.001 with 95% confidence, mmust be chosen so that:

    1.96y(n)

    m 0.001.(26)

    That is,

    m 1.96

    0.001

    22y (n).(27)

    Usually Var[Yjn] is not known, but can be

    estimated by

    Var[Yjn] = 1mm

    j=1(Yjn Yn)

    2(28)

    with

    Yn= 1

    m

    mj=1

    Yjn.(29)

    20

  • 7/27/2019 mfdc9d

    11/16

    The standard error of the Monte Carlo esti-

    mator H0 is

    H=y(n)

    m(30)

    is inversely proportional to the square root of

    the number of replications m.

    As a result a 50% reduction in standard error

    requires four times more replications.

    21

    Variance Reduction Techniques

    A number of variance-reduction techniques

    have been developed to improve the efficiency

    of the simulation estimators.

    A simple technique is to replace estimatorswith their population counterparts whenever

    possible.

    For example, when simulating risk-neutral as-

    set returns, the sampling variation can becorrected by adding the difference between

    the riskless rate and the sample mean to each

    observation in the replication.

    That is, suppose

    {rjk

    }nk=1 be the jth sample

    with the sample mean rj = 1n nk=1 rjk, thenthe corrected path is rjk =rjk (r rj) withrj =r, which has no sampling error.

    22

  • 7/27/2019 mfdc9d

    12/16

    The efficiency gain depends on the extend

    how much the estimation error in the sam-

    ple mean contributes to the overall sampling

    variation.

    23

    Antithetic Variate Methods

    Symmetry, like population distribution with

    respect to mean (risk-free return), is another

    population information that can be exploited.

    Producing m/2 sample paths the other m/2sample paths can be produced by mirroring

    the first ones through the mean. This is a

    simple example of more general antithetic-

    variates methods.

    Other methods are stratified sampling, im-

    portance sampling, etc.

    24

  • 7/27/2019 mfdc9d

    13/16

    As an illustration consider the antithetic-variate

    estimation.

    For each simulated price path{Sjk}nk=0 an-other path can be obtained by reverting the

    sign of each randomly generated IID standard

    normal variates on which the price series isbased, yielding a second path {Sjk}nk=0whichis negatively correlated with the first.

    25

    The resulting antithetic-variates estimate of

    H0 is

    Ha0 =erT 1

    2m

    mj=1

    Yjn +

    mj=1

    Yjn

    S0(31)or

    Ha0 =e2rT1

    m

    mj=1

    Yjn + Yjn2

    S0,(32)

    where

    Yjn = max0kn

    Sjk and Y

    jn = max0kn

    Sjk .(33)

    26

  • 7/27/2019 mfdc9d

    14/16

    Then

    Var[Ha0] =e2rT1

    mVar

    Yjn + Y

    jn

    ,(34)

    which reduces to

    Var[Ha0] = 1

    2m2y (n)(1 + ),(35)

    where(36)

    2y (n) = Var[erTYjn] = Var[erTYjn],

    and

    = Corr erTYjn, erTYjn.(37)The reduction of variance comes from two

    sources:

    (a) The number of replications is doubled

    from m to 2m

    (b) The factor 1 +, which should be less

    than one if the correlation of the antithetic

    variates is negative.

    27

    Path Dependent Options: PDE Method

    Simulation approach is an easy procedure for

    European type path dependent option. How-

    ever, it is time consuming and hard to imple-

    ment if Early exercise is allowed.

    Several path dependent options can be easily

    implemented in the PDE frame work.

    28

  • 7/27/2019 mfdc9d

    15/16

    Assume generally

    dS=a(S, t)dt + (S, t)dW .(38)

    Suppose that the path dependent quantity

    can be represented by an integral of some

    function of the asset

    I(t) = t

    0f(S, u)du,(39)

    such that

    dI=f(S, t)dt.(40)

    Typical examples are averages: A(T) =I(T)/T

    Arithmetic Average: f(S, t) =S.

    Geometric Average: f(S, t) = log(S).

    Remark 9.4: f(S, t) is assumed to be measurable w.r.t

    the information set and non-anticipating.

    29

    The Pricing Equation

    LetF(S , I , t) denote the value of the contact.

    The hedge portfolio is

    P =F(S , I , t)

    S,(41)

    where is the position making the portfolio

    risk-free (delta hedge).

    Using Ito

    dP =F

    t +

    1

    22(S, t)

    2F

    S2

    dt +

    F

    IdI+

    FS

    dS,(42)

    where 2(S, t) = ((S, t))2.

    30

  • 7/27/2019 mfdc9d

    16/16

    Choosing

    =F

    S(43)

    the portfolio becomes risk-free.

    No-arbitrage implies,

    dP =rP dt.(44)

    Thus, because dI=f(S, t)dt, we get finally

    F

    t +

    1

    22(S, t)

    2F

    S2 +

    F

    If(S, t)

    dt = r(F SF

    S)dt(45)

    or

    rF =F

    t + rS

    F

    S + f(S, t)

    F

    I +

    1

    22(S, t)

    2F

    S2(46)

    This is solved subject to the payoff condition

    at the expiry

    F(S, I) =G(S, I).(47)

    31

    Example 9.3: SupposeSt follows the geometric Brow-

    nian motion. Asian option has aPayoff that dependson the average of the asset price of some period. Ifthe period is form zero to expiry and the average isarithmetic, then f(S, t) =S and

    I=

    t0

    S du.(48)

    The payoff function at the expiry for the call option

    is

    G(S, I) = max[S I/T, 0](49)and the PDE to be solved is

    rF =F

    t + rS

    F

    S + S

    F

    I +

    1

    22S2

    2F

    S2.(50)

    32