Tải bản đầy đủ - 0trang
1222



Part Six numerical methods and programs



Theta = -0.5 * Vol ^ 2 * S(i) ^ 2 * Gamma - _

Int_Rate * S(i) * Delta + Int_Rate * VOld(i) _

’ Black-Scholes

VNew(i) = VOld(i) - dt * Theta ’ Update option value

Next i

VNew(0) = VOld(0) * (1 - Int_Rate * dt) ’ Boundary condition at S=0

VNew(NAS) = 2 * VNew(NAS - 1) - VNew(NAS - 2) ’ Boundary condition at S=infinity

For i = 0 To NAS ’ Replace Old with New

VOld(i) = VNew(i)

Next i

If EType = "Y" Then ’ Check for early exercise

For i = 0 To NAS

VOld(i) = Application.Max(VOld(i), Payoff(i))

Next i

End If

Next k

For i = 1 To NAS - 1

Dummy(i, 3) = VOld(i) ’ Third column of Dummy is option value

Dummy(i, 4) = (VOld(i + 1) - VOld(i - 1)) / 2 / dS ’ Fourth column of Dummy _

is delta

Dummy(i, 5) = (VOld(i + 1) - 2 * VOld(i) + VOld(i - 1)) / dS / dS

’ Fifth column of Dummy is gamma

Dummy(i, 6) = -0.5 * Vol ^ 2 * S(i) ^ 2 * Dummy(i, 5) - _

Int_Rate * S(i) * Dummy(i, 4) + Int_Rate * VOld(i) ’ Sixth column _

of Dummy is theta

Next i

Dummy(0, 3) = VOld(0)

Dummy(NAS, 3) = VOld(NAS)

Dummy(0, 4) = (VOld(1) - VOld(0)) / dS ’ End points need to be treated _

separately for delta, gamma and theta

Dummy(NAS, 4) = (VOld(NAS) - VOld(NAS - 1)) / dS

Dummy(0, 5) = 0

Dummy(NAS, 5) = 0

Dummy(0, 6) = Int_Rate * VOld(0)

Dummy(NAS, 6) = -0.5 * Vol ^ 2 * S(NAS) ^ 2 * Dummy(NAS, 5) - _

Int_Rate * S(NAS) * Dummy(NAS, 4) + Int_Rate * VOld(NAS)

Option_Value_2D_US = Dummy ’ Output array

End Function



Table 77.4 shows the results from this code, same parameters but now 40 asset steps.



77.16 BILINEAR INTERPOLATION

Suppose that we have an estimate for the option value, or its derivatives, on the mesh points;

how can we estimate the value at points in between? The simplest way to do this is to use



finite-difference methods for one-factor models Chapter 77



Table 77.4 Call option values and greeks output by the explicit finitedifference code.

Stock

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

125

130

135

140

145

150

155

160

165

170

175

180

185

190

195

200



Payoff



Option



Delta



Gamma



0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

5.000

10.000

15.000

20.000

25.000

30.000

35.000

40.000

45.000

50.000

55.000

60.000

65.000

70.000

75.000

80.000

85.000

90.000

95.000

100.000



0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.001

0.003

0.016

0.058

0.173

0.437

0.953

1.832

3.174

5.044

7.461

10.403

13.816

17.628

21.764

26.149

30.722

35.430

40.235

45.107

50.023

54.969

59.935

64.913

69.900

74.892

79.886

84.883

89.881

94.880

99.878

104.877



0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.002

0.005

0.016

0.038

0.078

0.139

0.222

0.321

0.429

0.536

0.636

0.723

0.795

0.852

0.896

0.928

0.951

0.968

0.979

0.986

0.991

0.994

0.997

0.998

0.999

0.999

0.999

1.000

1.000

1.000

1.000



0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.001

0.003

0.006

0.010

0.015

0.019

0.021

0.022

0.021

0.019

0.016

0.013

0.010

0.007

0.005

0.004

0.003

0.002

0.001

0.001

0.001

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000



Theta

0.000

0.000

0.000

0.000

0.000

0.000

0.000

0.000

−0.001

−0.005

−0.023

−0.086

−0.256

−0.617

−1.235

−2.113

−3.168

−4.255

−5.224

−5.962

−6.425

−6.625

−6.614

−6.459

−6.225

−5.965

−5.712

−5.488

−5.301

−5.153

−5.039

−4.954

−4.892

−4.848

−4.817

−4.796

−4.781

−4.770

−4.761

−4.754

−4.754



1223



1224



Part Six numerical methods and programs



V1



V2



A3



A4



Want to estimate value here



A2



A1



V4



Figure 77.14



V3



Bilinear interpolation.



a two-dimensional interpolation method called bilinear interpolation. This method is most

easily explained via the schematic diagram in Figure 77.14.

We want to estimate the value of the option, say, at the interior point in the figure. The

values of the option at the four nearest neighbors are called V1 , V2 , V3 and V4 , simplifying

earlier notation just for this brief section. The areas of the rectangles made by the four corners

and the interior point are labeled A1 , A2 , A3 and A4 . But note that the subscripts for the areas

correspond to the subscripts of the option values at the corners opposite. The approximation

for the option value at the interior point is then

4

i=1 Ai Vi

.

4

i=1 Ai



77.17 UPWIND DIFFERENCING

The constraint (77.4) can be avoided if we use a one-sided

difference instead of a central difference for the first derivative of the option value with respect to the asset. As I said

in Chapter 6 the first S derivative represents a drift term. This

drift has a direction associated with it, as t decreases, moving

away from expiry, so the drift is towards smaller S. In a sense,

this makes the forward price of the asset a better variable to

use. Anyway, the numerical scheme can make use of this by using a one-sided difference. That

is the situation for the Black–Scholes equation. More generally, the approximation that we use



finite-difference methods for one-factor models Chapter 77



for delta in the equation could depend on the sign of b. For example, use the following

V k − Vik

∂V

(S, t) = i+1

∂S

δS



if b(S, t) ≥ 0 then

but if

b(S, t) < 0 then



k

V k − Vi−1

∂V

(S, t) = i

.

∂S

δS



This removes the limitation (77.4) on the asset step size, improving stability. However, since

these one-sided differences are only accurate to O(δS) the numerical method is less accurate.

The use of one-sided differences that depend on the sign of the coefficient b is called upwind

differencing.3 There is a small refinement of the technique in the choice of the value chosen

for the function b:

if b(S, t) ≥ 0 then b(S, t)

but if

b(S, t) < 0 then b(S, t)



V k − Vik

∂V

(S, t) = bk 1 i+1

i+ 2

∂S

δS



k

V k − Vi−1

∂V

(S, t) = bk 1 i

.

i− 2

∂S

δS



Notice how I have used the mid-point value for b.4

Below is a Visual Basic code fragment that uses a one-sided difference, depending on the

sign of the drift term. This code fragment can be used for interest rate products, since it is

solving

∂V

∂ 2V

∂V

+ 12 σ (r)2 2 + (µ(r) − λ(r)σ (r))

− rV = 0.

∂t

∂r

∂r

Note the arbitrary σ (r) = Volatility(IntRate(i)) and µ(r) − λ(r)σ (r) =

RiskNeutralDrift(IntRate(i)).

This fragment of code is just the time stepping; above it would go declarations and setting up

the payoff. Below it would go the outputting. It does not implement any boundary conditions

at i = 0 or at i = NoIntRateSteps; these would depend on the contract being valued.

For i = 1 To NoIntRateSteps - 1

If RiskNeutralDrift(IntRate(i)) > 0 Then

Delta(i) = (VOld(i + 1) - VOld(i)) / dr

RNDrift = RiskNeutralDrift(IntRate(i) + 0.5 * dr)

Else

Delta(i) = (VOld(i) - VOld(i - 1)) / dr

RNDrift = RiskNeutralDrift(IntRate(i) - 0.5 * dr)

End If

Gamma(i) = (VOld(i + 1) - 2 * VOld(i) + VOld(i - 1)) / (dr * dr)

VNew(i) = VOld(i) + Timestep * (0.5 * Volatility(IntRate(i)) _

* Volatility(IntRate(i)) * Gamma(i) + RNDrift * Delta(i) - IntRate(i) * VOld(i))

Next i

3



That’s ‘wind’ as in breeze, not as in to wrap or coil.

This choice won’t make much difference to the result but it does help to make the numerical method ‘conservative,’

meaning that certain properties of the partial differential equation are retained by the solution of the difference equation.

Having a conservative scheme is important in computational fluid dynamics applications, otherwise the scheme will exhibit

mass ‘leakage.’

4



1225



1226



Part Six numerical methods and programs



To get back the O(δS 2 ) accuracy of the central difference with a one-sided difference you

can use the approximations described in Section 77.6.

We have seen that the explicit finite-difference method suffers from restrictions in the size of

the grid steps. The explicit method is similar in principle to the binomial tree method, and the

restrictions can be interpreted in terms of risk-neutral probabilities. The terms A, B and C are

related to the risk-neutral probabilities of reaching the points i − 1, i and i + 1. Instability is

equivalent to one of these being a negative quantity, and we can’t allow negative probabilities.

More sophisticated numerical methods exist that do not suffer from such restrictions, and I will

describe these next.

The advantages of the explicit method

• It is very easy to program and hard to make mistakes

• When it does go unstable it is usually obvious

• It copes well with coefficients that are asset and/or time-dependent

• It is easy to incorporate accurate one-sided differences



The disadvantage of the explicit method

• There are restrictions on the time step so the method can be slower than

other schemes



77.18 SUMMARY

The diffusion equation has been around for a long time. Numerical schemes for the solution of

the diffusion equation have been around quite a while too, not as long as the equation itself but

certainly a lot longer than the Black–Scholes equation and the binomial method. This means

that there is a great deal of academic literature on the efficient solution of parabolic equations

in general. I’ve introduced you to the subject with the explicit method, and in the next chapter

I’ll show you how the numerical methods can get more sophisticated.



FURTHER READING

• For general numerical methods see Johnson & Riess (1982) and Gerald & Wheatley (1992).

• The first use of finite-difference methods in finance was due to Brennan & Schwartz (1978).

For its application in interest rate modeling see Hull & White (1990).

• An excellent, well written, book on numerical methods in finance is Ahmad (2006).



CHAPTER 78



further

finite-difference

methods for

one-factor models

In this Chapter. . .

•

•

•

•

•



implicit finite-difference methods including Crank–Nicolson

Douglas schemes

Richardson extrapolation

American-style exercise

exotic options



78.1



INTRODUCTION



We continue with one-factor numerical methods, discussing the more difficult to program

implicit methods. The extra complexity of the methods is outweighed, though, by their superior

stability properties. I also show how to extend the finite-difference method to cope with early

exercise and path-dependent contracts.



78.2



IMPLICIT FINITE-DIFFERENCE METHODS



The fully implicit method uses the points as shown in Figure 78.1 to calculate the option value.

The scheme is superficially just like the explicit method using finite-difference estimates of the

option value, its delta and gamma but now at the time step k + 1. The relationship between the

option values on the mesh is simply

Vik − Vik+1

δt

+ aik+1



k+1

k+1

Vi+1

− 2Vik+1 + Vi−1

δS 2



1228



Part Six numerical methods and programs



S



Calculate option

values at these

points



from the option

value at this point



t



Figure 78.1 The relationship between option values in the fully implicit method.



+ bik+1



k+1

k+1

− Vi−1

Vi+1

2 δS



+ cik+1 Vik+1 = 0.

(It doesn’t matter much whether the coefficients a, b and c are evaluated at the time step k + 1

or k.) The method is still accurate to O(δt, δS 2 ).

This can be written as

k+1

k+1

Ak+1

Vi−1

+ (1 + Bik+1 )Vik+1 + Cik+1 Vi+1

= Vik

i



(78.1)



where

= −ν 1 aik+1 − 12 ν 2 bik+1 ,

Ak+1

i

Bik+1 = 2ν 1 aik+1 − δt cik+1

and

Cik+1 = −ν 1 aik+1 + 12 ν 2 bik+1

where

ν1 =



δt

δS 2



and ν 2 =



δt

.

δS



Again, Equation (78.1) does not hold for i = 0 or i = I , the boundary conditions supply the

two remaining equations.



further finite-difference methods for one-factor models Chapter 78



There is a world of difference between this scheme and the explicit finite-difference scheme.

The two main differences concern the stability of the method and the solution procedure.

The method no longer suffers from the restriction on the time step. The asset step can be

small and the time step large without the method running into stability problems.

The solution of the difference equation is no longer so straightforward. To get Vik+1 from

k

Vi means solving a set of linear equations; each Vik+1 is directly linked to its two neighbors

and thus indirectly linked to every option value at the same time step.

I am not going to pursue the fully implicit method any further since the method can be

significantly improved upon with little extra computational effort.



78.3



THE CRANK–NICOLSON METHOD



The Crank–Nicolson method can be thought of as an average of the explicit method and the

fully implicit method. It uses the six points shown in Figure 78.2.

The Crank–Nicolson scheme is

Vik − Vik+1

δt

+



aik+1

2



k+1

k+1

− 2Vik+1 + Vi−1

Vi+1

δS 2



+



bik+1

2



k+1

k+1

− Vi−1

Vi+1

2 δS



+

+



aik

2



bik

2



k

k

− 2Vik + Vi−1

Vi+1

δS 2

k

k

− Vi−1

Vi+1

2 δS



+ 12 cik+1 Vik+1 + 12 cik Vik = O(δt 2 , δS 2 ).



S



Calculate option

values at these

points



from the option

values at these

points



t



Figure 78.2 The relationship between option values in the Crank–Nicolson method.



1229



1230



Part Six numerical methods and programs



This can be written as

k+1

k+1

k

k

Vi−1

+ (1 − Bik+1 )Vik+1 − Cik+1 Vi+1

= Aki Vi−1

+ (1 + Bik )Vik + Cik Vi+1

,

−Ak+1

i



(78.2)



where

Aki = 12 ν 1 aik − 14 ν 2 bik ,

Bik = −ν 1 aik + 12 δt cik ,

and

Cik = 12 ν 1 aik + 14 ν 2 bik .

These equations only hold for 1 ≤ i ≤ I − 1. The boundary conditions again supply the two

missing equations. They are harder to handle than in the explicit method and I will discuss

them in depth shortly.

Although this looks a bit messy, and it is harder to solve, the beauty of this method lies in

its stability and accuracy. As with the fully implicit method there is no relevant limitation on

the size of the time step for the method to converge. Better yet, the method is more accurate

than the two considered so far. The error in the method is O(δt 2 , δS 2 ). So now we can use

larger time steps, and still get an accurate solution. To see that the error due to the size of the

time step is now O(δt 2 ) expand Equation (78.2) about the point S, t − 12 δt (this is the point

‘x’ in Figure 78.2).

The Crank–Nicolson method can be written in the matrix form

 k+1 

V0

 k+1 





 V

 1



−Ak+1

1 − B1k+1 −C1k+1

0

.

.

.

1

 . 











 . 

1 − B2k+1 .

.

.

.

0

−Ak+1

2









 . 

.

0

.

.

.

0

.









 . 

k+1

k+1









.

.

.

. 1 − BI −2

−CI −2

0



 . 





k+1

 k+1 

1 − BIk+1

.

.

.

0 −Ak+1

 V



I −1

−1 −CI −1

 I −1 







Ak1





 0





= .



 .



.



1 + B1k



C1k



0



.



.



.



Ak2

0



1 + B2k

.



.

.



.

.



.

0



.

.



CIk−2



0



1 + BIk−1



CIk−1



.



.



.



.



.



0



1+



BIk−2



AkI −1



V0k





 V1k



.





.





.







 .

 .



 Vk

 I −1



VIk+1



































(78.3)



VIk

The two matrices have I − 1 rows and I + 1 columns. This is a representation of I − 1

equations in I + 1 unknowns.



further finite-difference methods for one-factor models Chapter 78



The two equations that we are missing come from the boundary conditions. Using these

conditions, I am going to convert this system of equations, (78.3), into a system of equations

involving a square matrix. The aim is to write the system of equations in the form

k+1

+ rk = MkR vk ,

Mk+1

L v



for known square matrices Mk+1

and MkR , and a known vector rk , and where details of the

L

boundary conditions have been fully incorporated.

I will consider three cases separately, depending on the form of the boundary condition that

is to be implemented. I will also only deal with a boundary condition at i = 0; obviously

boundary conditions at i = I are treated similarly.

78.3.1



Boundary Condition Type I: V0k+1 Given



Sometimes we know that our option has a particular value on the boundary i = 0, or on i = I .

For example, if we have a European put we know that V (0, t) = Ee−r(T −t) . This translates to

knowing that V0k+1 = Ee−r(k+1)δt . We can write

 k+1 

V0









 V k+1 

k+1

k+1





1

−Ak+1

1

−

B

−C

0

.

.

.

1

1

1







 . 

k+1

k+1







1 − B2

.

.

.

.

0

−A2

 . 









 . 



.

0

.

.

.

0

.







 . 



k+1

k+1

 . 



.

.

.

. 1 − BI −2

−CI −2

0









 k+1 

k+1

k+1

1

−

B

−C

.

.

.

0 −Ak+1





I −1

I −1

I −1

 VI −1 

VIk+1

as







1 − B1k+1





 −Ak+1

2

























+















−C1k+1



0



.



.



1 − B2k+1



.



.



.



0



.



.



.



0



.



.



.



1 − BIk+1

−2



−CIk+1

−2



.



0



−Ak+1

I −1



1 − BIk+1

−1



.

k+1

−Ak+1

1 V0

0

0

.

.

0

.



























V1k+1

.

.

.

.

.

VIk+1

−1

































 = Mk+1 vk+1 + rk .

L











All I have done here is to multiply out the top and bottom rows of the matrix. The matrix ML

is square and of size I − 1. The vector rk , which is of length I − 1, only has non-zero elements



1231



1232



Part Six numerical methods and programs



at the top (and bottom) and is completely known because it only depends on the function A and

the prescribed value of V at the boundary.

78.3.2



Boundary Condition Type II: Relationship Between V0k+1 and V1k+1



If we have a barrier option for which a grid point does not coincide with the barrier we must

use the approximation explained in Section 77.10:

V0k+1 =



1

f − (1 − α)V1k+1 .

α



This relationship between V0k+1 and V1k+1 for a ‘down’ barrier or between VIk+1 and VIk+1

−1 for

an ‘up’ barrier is seen in other contexts. Perhaps we know the slope of the option value for

large or small S; this gives us a gradient boundary condition, which is also of this form. If we

use a one-sided difference for the derivative then the boundary condition can also be written

as a relationship between the last grid point and the last but one. More generally, suppose

we have

V0k+1 = a + bV1k+1 .

This time the left-hand side of (78.3) can be written as





B1k+1



1−



 −Ak+1

2















−C1k+1

1−



0



.



.



0



.

.



.

.



.

.



0



.



.



.



1 − BIk+1

−2



−CIk+1

−2



.



.



0



−Ak+1

I −1



1 − BIk+1

−1



B2k+1







 

k+1

−Ak+1

)

 V1k+1

1 (a + bV1



 .  



 

0











 .  



0











.



+

 . 

.













 . 

.











 



 .  

0

k+1

VI −1

.



Again, this comes from multiplying out the top row of the matrix. By absorbing the V1k+1 term

into the matrix we can write this as



 k+1  

−aAk+1



 V1

1

k+1

k+1

k+1

1 − B1 − bA1

−C1

0 .

.





 

 .   0 







 





1 − B2k+1 . .

.

−Ak+1



 

2





 .   0 





0

.

.

.

0



.







 .  +

.















k+1

k+1 



.

.

.

.

.

1

−

B

−C









I

−2

I

−2





 



 .   0 

k+1

k+1

.

. 0 −AI −1

1 − BI −1

.

VIk+1

−1

Again this is of the form

k+1

Mk+1

+ rk

L v



but for different Mk+1

and rk from before. This matrix and this vector are both known.

L