Ch 4.

Binomial Tree Model

I. One-Period Binomial Tree

II. CRR Binomial Tree Model

III. Estimation and Calibration of µ and σ

IV. Dividends and Option Pricing

V. Introduction of Combinatorial Method

Appendix A. Binominal Tree Model for Jump-Diffusion Processes

• This chapter is devoted to introduce the binomial tree model, which is also known as a
kind of lattice model. The lattice models, such as the binomial tree model introduced in
this chapter or the finite difference method introduced in the next chapter, are popular
numerical methods for option pricing, particularly for pricing American-style derivatives.
They are also flexible since only nominal changes of the payoff function are needed for
dealing with pricing complex, nonstandard options.

I. One-Period Binomial Tree

Figure 4-1
22
1

20
( = 21)

18
0

t=0 t = 0.25

(i) Constructing a portfolio: long ∆ shares and short 1 call

4-1
Figure 4-2
22△－1

 20△－ c

18△

(ii) Deciding the value of ∆: if 22∆ − 1 = 18∆ ⇒ portfolio is riskless

⇒ 4∆ = 1 ⇒ ∆ = 0.25.
(iii) Hence, at t = 0.25, the value of the portfolio is 22 · 0.25 − 1 = 18 · 0.25 = 4.5
⇒ 20∆ − c = 4.5 · e−r·0.25 , where ∆ = 0.25 and r = 0.12 ⇒ c = 0.633.
∗ The emergence of the risk free interest rate r is due to the no-arbitrage argument.

Figure 4-3
S0u  cu

S 0  c cu  cd
“=”  =
S 0u  S 0 d

S0 d  cd

t 0 t T






risk free rate r

cu −cd
∆= S0 u−S0 d
cu −cd u −cd
⇒ S0 u−S0 d 0
S − c = (S0 u S0cu−S 0d
− cu )e−rT
u −cd u −cd
⇒ c = cu−d − ( cu−d u − cu )e−rT
u −cd rT (cu −cd )u (u−d)
⇒ c = e−rT ( cu−d e − u−d
+ u−d u
c )
rT −c
⇒ c = e−rT ( (cu −cd )e u u+cd u+ucu −dcu +cd d−cd d
u−d
)
(erT −d)cu (u−d)cd (erT −d)c
= e−rT ( u−d
+ u−d
− u−d
d
)
erT −d erT −d
= e−rT ( u−d
· cu + (1 − u−d
) · cd )
rT −d erT −d
( eu−d and 1 − u−d
are like binomial probabilities, so they are denoted as p and 1 − p.)
−rT
=e (p · cu + (1 − p) · cd ).
∗ For different options, the above equation remains valid, but different payoffs cu and cd
should be considered.

4-2
• It is worth noting that p is not the probability for cu (or for the upward movement of S).
However, p can be regarded as the probability for cu (or for the upward movement of S)
in the risk neutral world. This is because in the real world, if the expected stock return
is µ,
eµT − d
S0 eµT = S0 u · q + S0 d · (1 − q) ⇒ q = .
u−d
Similarly, in the risk neutral world, since the expected returns for all securities are the
same and equal to r,

erT − d
S0 erT = S0 u · p + S0 d · (1 − p) ⇒ p = .
u−d
Therefore, p and 1 − p are termed as risk neutral probabilities in the binomial tree frame-
work.

• The option pricing equation c = e−rT (p · cu + (1 − p) · cd ) in the binomial tree model is

consistent with the RNVR because both the expected growth rate of the underlying asset
and the discount rate of the option payoff are the risk free rate.

• If the upward and downward probabilities in the real world are considered, it is difficult
to identify a proper discount rate to discount the expected payoff of options, i.e.,

c = e−?T (q · cu + (1 − q) · cd ).

In practice, for securities with more risk or uncertainty, it should apply higher discount
rates to future expected payoffs. Furthermore, it is well-known that options are riskier
than their underlying assets due to the high-leverage characteristic for investing in options.
So, it is not suited to use the expected return for the underlying asset, µ, to discount the
expected payoff of the option.
If we reconsider the above numerical example of the one-period binomial tree and
assume µ = 16%, then we can derive the probability q in the real world to be 0.7041.
Next by equalizing e−?T · (q · cu + (1 − q) · cd ) to be 0.633, which is the true option value
in the numerical example, we can derive the discount rate for the option to be 42.58%.
In fact, the proper discount rates for expected payoffs of options depend not only on
the expected returns (µ) and volatilities (σ) of underlying assets, but also on the different
K and T of options. As a consequence, it is almost impossible to derive theoretical option
prices directly in the real world.

4-3
II. CRR Binomial Tree Model

• Lognormal property
If X is lognormally distributed, i.e., ln X follows a normal distribution with mean =
E[ln X] and variance = var(ln X), then E[X] = eE[ln X]+ 2 ·var(ln X) , and var(X) =
1

e2·E[ln X]+var(ln X) · (evar(ln X) − 1).

σ2
∵ ln ST ∼ N D(ln S0 + (µ − 2
)T, σ 2 T )
2
⇒ E[ST ] = S0 eµT , and var(ST ) = S02 e2µT (eσ T − 1).
∵ ln St+∆t ∼ N D(ln St + (µ − σ2 )∆t, σ 2 ∆t)

2
⇒ E[St+∆t ] = St eµ∆t .

⇒ var(St+∆t ) = S 2 e2µ∆t (eσ2 ∆t − 1)
t
2 2 x

≈ St (1 + 2µ∆t)(1 + σ ∆t − 1) (because e ≈ 1 + x when x → 0)
2 2 2 2

= St σ ∆t + St (2µ∆t)(σ ∆t)

≈ St2 σ 2 ∆t (because the term with ∆t2 is relatively too small).
(In the next paragraph, I will use this property to derive the parameters u, d, and p

in the CRR binomial tree framework.)

• Deriving u, d, and p in the CRR (Cox, Ross, and Rubinstein (1979)) binomial tree model,
which is the most common and famous binomial tree model.

Figure 4-4

St  u
p

St

1 p
St  d

t

(i) Matching mean: p · St u + (1 − p) · St d = E[St+∆t ] = St er∆t

er∆t −d
⇒p= u−d
.

4-4
 2
(ii) Matching variance: var(St+∆t ) = E[St+∆t ] − E[St+∆t ]2 .
⇒ σ 2 ∆t = p · u2 + (1 − p) · d2 − [p · u + (1 − p) · d]2 (both sides ×(1/St )2 )
= p · u2 + (1 − p) · d2 − p2 · u2 − 2 · p · (1 − p) · u · d − (1 − p)2 · d2
= u2 · (p − p2 ) + [(1 − p) − (1 − p)2 ] · d2 − 2 · p · (1 − p) · u · d
= u2 · p · (1 − p) + (1 − p) · [1 − (1 − p)] · d2 − 2 · p · (1 − p) · u · d
= p · (1 − p) · [u2 − 2 · u · d + d2 ] = p · (1 − p) · (u − d)2

p · (1 − p) = p − p2 = er∆t −d − e2r∆t −2·d·er∆t +d2

u−d (u−d)2

r∆t r∆t 2 −e2r∆t +2·d·er∆t −d2

= e ·u−u·d−e ·d+d(u−d)2


r∆t 2r∆t r∆t 2r∆t
= e (u−d+2·d)−u·d−e = e (u+d)−u·d−e


(u−d)2 (u−d)2

⇒ σ 2 ∆t = er∆t (u + d) − u · d − e2r∆t
⇒ σ 2 ∆t = er∆t (u + u1 ) − u · 1
u
− e2r∆t (by defining d = u1 )
1 σ 2 ∆t+1+e2r∆t
⇒ u+ u
= er∆t
−r∆t 2
=e σ ∆t + e−r∆t + er∆t
−r∆t
e ≈ (1 − r∆t), er∆t ≈ (1 + r∆t), and r · σ 2 · ∆t2 → 0

≈ σ 2 ∆t + 2

⇒ u2 − (σ 2 ∆t + 2)u + 1 = 0
√ √
σ 2 ∆t+2± (σ 2 ∆t+2)2 −4 σ 2 ∆t+2± σ 4 ∆t2 +4σ 2 ∆t+4−4
⇒u= 2
= 2
(σ 4 ∆t2 → 0)
2 √
= σ 2∆t + 1 ± σ ∆t
√
Since ∆t is far larger than ∆t for a small ∆t,

and σ 2 is relatively smaller than σ

⇒ we can ignore the first term σ2 ∆t
2
√
≈ 1 ± σ ∆t
√ √
≈ e+σ ∆t
(because u > 1 ⇒ u 6= e−σ ∆t
).

4-5
• Implementation of the CRR binomial tree model:

Figure 4-5

S (n, 0)  S0u n d 0


S (2, 0)  S0u 2
S ( n, j )  S 0u n  j d j
S (1, 0)  S0u
S (i, j )  S0u i  j d j 

S (2,1)  S0 
S (0, 0)  S 0


S (1,1)  S0 d
S (2, 2)  S0 d 2 


1. S (i, j )  S0ui  j d j , for 0  j  i  n and T / n  t
2. c( n, j )  payoff for S0u n  j d j at maturity 
3. c(i , j )  e  rt [ p  c(i  1, j )  (1  p )  c(i  1, j  1)] (backward induction)
4. For American options, c(i , j )  max(c(i, j ), exercise value for S0ui  j d j )
5. c(0,0) is the option price today S ( n, n )  S 0 u 0 d n

√
r∆t r∆t −σ ∆t
• Problems of the CRR model: 1) p = e u−d−d = eeσ√∆t−e √
−e−σ ∆t
is not necessary in [0, 1] unless
∆t is small enough. 2) The approximations used to derive var(St+∆t ) as well as u and d
are vaild only when ∆t is very small.

4-6
• Another binomial tree model proposed in Jarrow and Rudd (1983): by considering the
σ2
√
logarithmic stock price space and the constraint of p = 1/2, u = e(r− 2
)∆t+σ ∆t
and
2 √
(r− σ2 )∆t−σ ∆t
d=e can thus be obtained.

Figure 4-6

ln St  ln u
p

ln St

1 p
ln St  ln d

t

σ2
∵ ln St+∆t ∼ N D(lnSt + (r − 2
)∆t, σ 2 ∆t)
σ2
∴ E[ln St+∆t ] = ln St + (r − 2
)∆t, var(ln St+∆t ) = σ 2 ∆t
Matching mean and variance of ln St+∆t :
2
p(ln St + ln u) + (1 − p)(ln St + ln d) = ln St + (r − σ2 )∆t

(1)
⇒ 2
p(ln St + ln u)2 + (1 − p)(ln St + ln d)2 − (ln St + (r − σ2 )∆t)2 = σ 2 ∆t (2)
In this model, p is fixed to be 0.5, and only u and d are left as unknowns.
From Eq. (1), we can derive
σ2
0.5 ln u + 0.5 ln d = (r − 2
)∆t
σ2
⇒ ln u = 2(r − 2
)∆t − ln d
σ2
k define D = ln d and µ = (r − 2
)∆t
⇒ ln u = 2µ − D.
Replacing ln u with 2µ − D in Eq. (2) yields
0.5(ln St + 2µ − D)2 + 0.5(ln St + D)2 − (ln St + µ)2 = σ 2 ∆t
⇒ (ln St + 2µ − D)2 + (ln St + D)2 − 2(ln St + µ)2 = 2σ 2 ∆t
⇒ D2 − 2µD + µ2 − σ 2 ∆t = 0
√ √
2µ± 4µ2 −4(µ2 −σ 2 ∆t)
⇒D= 2
= µ ± σ ∆t = ln d
√
⇒ ln u = µ ∓ σ ∆t.
( √ σ2
√
u = eµ+σ ∆t = e(r− 2 )∆t+σ ∆t
Because u > d ⇒ √ σ2
√ .
d = eµ−σ ∆t = e(r− 2 )∆t−σ ∆t

4-7
• Advantages of this alternative binomial tree model: 1) there is no approximation; 2) p
maintains to be a positive number between 0 and 1; 3) the convergence rate is generally
better than the CRR model for pricing plain vanilla options.

• Disadvantages of this alternative binomial tree model is due to S0 ud 6= S0 : 1) Since there

is no stock price layer coinciding with the specified barrier price, it is less effecient to
price the family of barrier options. 2) Since there is no stock price layer equal to S0 , it
is impossible to apply a quick method, known as the extended tree model, to estimate
some Greek letters, like ∆, Γ, and Θ.

• One possible trinomial tree structure proposed in Hull (2011):

r r
2
√ 1 ∆t σ 1 2 ∆t σ2 1
u = eσ 3∆t , d = , pu = 2
(r − ) + , p m = , p d = − 2
(r − )+ ,
u 12σ 2 6 3 12σ 2 6
through matching (i) mean of St+∆t , (ii) variance of St+∆t , (iii) pi = 1, (iv) d = u1 ,
P
(v) pm = 23 , where (iv) and (v) are arbitrarily imposed constraints.
(Refer to Boyle (1986), Kamrad and Ritchken (1991), and Tian (1993) for more trinomial
tree structures.)
(Kamrad and Ritchken (1991) and Tian (1993) show that faster error convergence rates
can be attained if all of the three branching probabilities in the trinomial tree are close
or equal to 13 .)

4-8
 III. Estimation and Calibration of µ and σ

E[ln(ST /S0 )] = (µ − σ2 )T
2
6= ln(E[ST /S0 ])
↑ because k assume
2
ln ST − ln S0 ∼ N D((µ − σ2 )T, σ 2 T ) KT

Assume S1 = S0 eη (T = 1 year) ⇒ E[ST /S0 ] = eKT

(η: continuonsly compounding return per annum) ⇒ E[ST ] = S0 eKT
(S1 is stochastic ⇒ η follows a distribution) (E[ST ] is a known number ⇒ K is constant)
σ2
⇒ η = ln SS01 ∼ N D(µ − 2
2 ,σ ) ⇒ K = µ = expected growth rate according to
the lognormal property of ST on page 4-5
2
σ
⇒µ− 2 is the expected value of the continuously
compounding return per annum

i. Considering two trading days, t and t + 1, i. Considering two trading days, t and t + 1,

we can derive St+1 = St · eηd and ηd = ln SSt+1

t
then E[St+1 /St ] = eKd and Kd = ln E[St+1 /St ]

ii. Calculating the average of ln SSt+1

t
for n days, ii. Calculating the average of St+1
St for n days and
σd2
the result is the estimation of µd − 2 . taking the natural log, we can estimate Kd = µd
S1
1
n (ln S0 + ln SS21 + · · · + ln SSn−1
n
) ln( n1 ( SS10 + S2
S1 + ··· + Sn
Sn−1 ))

= n1 (ln R1 + ln R2 + · · · + ln Rn ) = ln( n1 (R1 + R2 + · · · + Rn ))

= n1 (ln(R1 · R2 · · · Rn ))
1
= ln(R1 · R2 · · · Rn ) n
⇒ The geometric average of daily returns ⇒ The arithemtic average of daily returns
estimate the continuously compounding estimates the daily expected growth rate µd
2
return µd − σ2d
St+1
⇒ The standard deviation of the series of daily ⇒ The standard deviation of the series of St
ln SS01 , ln SS12 , . . . , ln SSn−1
n
generates the estimation is NOT the estimation of σd
of σd

∗ Note that the estimated results based on the daily data, i.e., µd and σd , need to be annualized to derive
the corresponding annual results.

4-9
• Implied volatility (the calibration (校準) of σ)
For any option pricing function c(S0 , K, r, σ, T ), S0 is the stock price today, K and T can
be found in the option contract, and r is the risk-free rate corresponding to the time to
maturity T . As for σ, it is commonly estimated based on the historical stock prices.
However, the market price of an option reflects the consensus of the forward-looking
σ of market participants and may not equal the theoretical option value based on the
historical σ. Through equalizing c(S0 , K, r, σ, T ) and the market option price, it is possible
to calibrate σ from the forward-looking viewpoint.
The value of σ ∗ satisfying f (σ ∗ ) ≡ c(S0 , K, r, σ ∗ , T )−market option price = 0 is called
the implied volatility. Here two root-finding algorithms are introduced to solve for the
implied volatility.

• Bisection Method
First find [an , bn ] such that f (an )f (bn ) < 0. The iterative two steps to find [an+1 , bn+1 ]
are as follows.
bn −an
(i) Calcuate xn = an + 2
(ii) If f (an )f (xn ) < 0 ⇒ an+1 = an , bn+1 = xn
else ⇒ an+1 = xn , bn+1 = bn

• Newton’s Method
f (xn )
xn+1 = xn − f 0 (xn )

Based on the first-order Taylor series: f (x) = f (xn ) + f 0 (xn )(x − xn )

(Find the root of f (x), i.e., solve f (x) = 0.)

⇒ −f (xn ) = f 0 (xn )(x − xn ) ⇒ x = xn+1 = xn − ff0(x n)

(xn )

Quadratical convergence:
f 00 (ξ)


According to the second-order Taylor series, f (x) = f (xn ) + f 0 (xn )(x − xn ) + 2
2 (x − xn ) ,


where ξ is between xn and x.

Solve f (x) = 0:
f (xn ) f 00 (ξ)

⇒ x − xn + f 0 (xn )
= − 2f 0 (xn ) (x − xn )2
f 00 (ξ)
⇒ x − xn+1 = − 2f 0 (xn ) (x − xn )2



f 00 (ξ) f 00 (ξ)
(Suppose | 2f 0 (xn ) | is bounded by a finite number M , i.e., | 2f 0 (xn ) | ≤ M < ∞.)


⇒| x − xn+1 | ≤ M | x − xn |2

(The error is smaller than the product of a finite number and the square of the error of

the n-th iteration.)

4-10
IV. Dividends and Option Pricing

• This section introduces how to modify the option pricing models if the dividend yield q
or known cash dividends D at the pre-specified time point t are considered:

Table 4-1 Three Different Models for Dividend Payments

Suppose the time Black-Scholes Model Binomial Tree Model

point today is 0 European European American
Model 1: ← or
∆
−
Dividend yield ← ∆
− =
= −
−
Model 2:
Known cash
dividends at as a not available Figure 4-8 Figure 4-8
percentage of

Model 3:
← −
Known cash ← − Figure 4-10
or Figure 4-10
dividends at

• Model 1: dividend yield q


S0 + paying dividend yield q
It is known that the distributions of ST are the same under .
S0 e−qT + no dividend payment

Figure 4-7

4-11
 Analytical formula obtained by replacing S0 with S0 e−qT in the Black-Scholes formula:
⇒ c = S0 e−qT N (d1 ) − Ke−rT N (d2 ), p = Ke−rT N (−d2 ) − S0 e−qT N (−d1 ),
2
ln(S0 /K)+(r−q+ σ2 )T
2
ln(S0 /K)+(r−q− σ2 )T √
where d1 = √ , d2 = √ , and d2 = d1 − σ T .
σ T σ T

 The PDE for (S0 + paying dividend yield q):

dS
Given S = (µ − q)dt + σdZ, we have dS = (µ − q)Sdt + σSdZ.
(Note that S denotes the ex-dividend underlying asset price.)
If f (S, t) is the price for any derivative, according to the Itô’s Lemma,
2
df = ( ∂f
∂t +
∂f
∂S (µ − q)S + 1∂ f 2 2 ∂f
2 ∂S 2 σ S )dt + ∂S σSdZ.
Construct a portfolio π:
−1 derivative
∂f
+ ∂S shares
∂f ∂f ∂f
⇒ π = −f + ∂S S ⇒ dπ = −df + ∂S dS + ∂S Sqdt (including received dividends)
∂f ∂f
(where ( ∂S S)qdt is the dividend from holding ( ∂S S) in [t, t+dt])
2
= (− ∂f
∂t −
1∂ f 2 2
2 ∂S 2 σ S + ∂f
∂S Sq)dt.
Due to the no-arbitrage argument,
dπ = rπdt
2
⇒ (− ∂f
∂t −
1∂ f 2 2
2 ∂S 2 σ S + ∂f
∂S Sq)dt = r(−f + ∂f
∂S S)dt
2
∂f ∂f ∂ f
⇒ ∂t + (r − q)S ∂S + 21 σ 2 S 2 ∂S 2 = rf .

You can check that the formulas shown at the top of this page are the solutions of the
above PDE given the boundary conditions being max(ST − K, 0) and max(K − ST , 0) for
the call and put options, respectively.

∂f ∂f
∗ When the dividend yield q is present, it can be found that (r − q)S ∂S replaces rS ∂S
in the original PDE (without dividend yield) on p. 2-2. However, it does not mean to
replace r with r − q in the PDE, because the right-hand side of the above PDE remains
to be rf . The truth is that for the underlying asset S, the expected growth rate is from
r to become r − q, but the discount rate for the derivative f is still r.

4-12
  For the binomial tree model
Method 1: replace S0 with S0 e−qT + no dividend payment
Method 2: change the risk neutral probability p

e(r−q)∆t − d
p · St u + (1 − p) · St d = St e(r−q)∆t ⇒ p =
u−d
∗ However, during the backward induction phase, we still use r to discount the expected
option value at the next time point, i.e., f = e−r∆t [p · fu + (1 − p) · fd ] (For European
options, both above methods generate correct results, but for American options, only the
second method can generate correct results.)

• Model 2: known cash dividends as a percentage δ of the stock price at the time point t
(In practice, it is rare for companies to distribute cash dividends in this way.)
 For the Black-Scholes formula, it is unavailable to deal with this problem.
 For the binomial tree model, it is simple to deal with this problem (see Figure 4-8).

Figure 4-8

S0u 4 (1   )

S0u 3 (1   )

S0 u
S0u 2 (1   ) S0u 2 (1   )

S0 S0u (1   )

S0 (1   ) S0 (1   )
S0 d
S0 d (1   )

S0 d 2 (1   ) S0 d 2 (1   )

S0 d 3 (1   )

S0 d 4 (1   )

4-13
 • Model 3: known cash dividends D at time point t
 For the Black-Scholes formula, replace S0 with S0 − De−rt .
 For the binomial tree model,
Method 1: replace S0 with S0 − De−rt . However, this method only works for European
options, but it cannot be applied to pricing American options.

S0 + paying D dollars at t
(This is because the distributions of Sτ (τ ≥ t) are the same under .)
S0 − De−rt + no dividend payment
Method 2: deduct the known dividend D from all stock prices on the divided payment
date. This method cannot work because it makes the tree to be non-recombined.

Figure 4-9

S 0u 2  D
S0 u

S0
S0  D

S0 d

S0 d 2  D

4-14
Method 3: this method can maintain the recombined feature of the binomial tree
Figure 4-10
S0u 5

S0u 4

S0u 3 S0u 3

S0u 2 S0u 2

S0u S0u S0u

S0 S0
S0

S0 d S0 d S0 d

S0 d 2 S0 d 2

S0 d 3 S0 d 3

S0 d 4

S0 d 5

t0  0 t1 t2 t3 t t4 t5

 r ( t t0 )  r ( t t1 )  r ( t  t2 )  r ( t t3 )

 De + De  De  De

Define S0∗ = S0 − De−r(t−t0 ) = S0 − De−rt

Step 1: build S ∗ -tree following the traditional way, but it should be noted that in
theory σ ∗ = σ · S0 /(S0 − De−r(t−t0 ) ), where σ ∗ is the volatility for the
stochastic part of the stock price and σ is the total volatility of the stock
price.
(For a different volatility σ ∗ , we should derive new u∗ , d∗ , and p∗ in theory.
In practice, however, it is rare to calculate σ ∗ . The first reason is that the
difference between σ ∗ and σ is minor. The second reason is that since σ
is an estimated value based on the historical data, performing the above
adjustment will not necessarily lead to a better estimation for σ ∗ in the
forward-looking sense.)
Step 2: for ti < t, replace St∗i with St∗i + De−r(t−ti )
(Since S0∗ = S0 − De−r(t−t0 ) = S0 − De−rt in Step 1, S ∗ -tree is the same as
the tree in Method 1 after t, which formulates the stock price process after
the dividend payment appropriately. However, the stock prices before t are
not consistent with the stock price process before the dividend payment,
i.e., S0∗ 6= S0 and E[St∗i ] 6= E[Sti ] for ti < t. Therefore, the adjustment in
Step 2 should be performed.)

4-15
• Currency option
 Replace the dividend yield q with the foreign risk-free rate rf
(Because the underlying asset St is a dollar of the foreign currency in domestic dollars,
and holding the foreign currency can earn the foreign risk-free rate, a foreign currency
is analogous to a stock paying a known dividend yield.)

 c = S0 e−rf T N (d1 ) − Ke−rT N (d2 ),

2
ln(S0 /K)+(r−rf + σ2 )T √
where d1 = √ and d2 = d1 − σ T .
σ T

• Futures options
 Call holders: have the right to enter a long position futures with the deliver price to
be FT and receive cash FT − K if FT ≥ K, where FT is the latest settlement futures
price before T (usually FT is the closing price on the date T ) and K is the strike price
in futures options.

 Put holders: have the right to enter a short position of futures with the deliver price
to be FT and receive cash K − FT if FT ≤ K.

∗ Since a futures contract is worth zero when it is first created (in the above cases, it is
at T ), it can be concluded that the payoff of futures option is similar to that of exercising
plain vanilla options, and the only difference is to replace ST with FT . This observation
implies that we can develop the pricing formula for futures options by following the same
way to develop the option formula for plain vanilla options.

4-16
  Black-Scholes formula for futures options:
dF
Suppose the futures price follows the geometric Brownian motion: F = µdt + σdZ
c(F, t) is the call on futures, and according to the Itô’s lemma
1 ∂2c 2 2
⇒ dc = ( ∂c
∂t +
∂c
∂F µF + ∂c
2 ∂F 2 σ F )dt + ∂F σF dZ
∂c
Construct portfolio π = −c + ( ∂F )F = −c (the initial value of futures is 0)
∂c 1 ∂2c 2 2
dπ = −dc + ( ∂F )dF = −( ∂c
∂t + 2 ∂F 2 σ F )dt
= rπdt = r(−c)dt
2
∂c 1 ∂ c 2 2
⇒ ∂t + 2 ∂F 2 σ F = rc.
∂c
(Comparing with the PDE for stock options, these is no such term ∂F (r − q)F .)
Thus, one can set q = r and S0 = F0 in the Black-Scholes formula to derive the formula
for futures options.)
2
ln(F0 /K)+ σ 2·T √
c = e−rT [F0 N (d1 ) − KN (d2 )], where d1 = √ and d2 = d1 − σ T .
σ T

 Binomial Tree
e(r−q)∆t −d 1−d
Because q = r, p = u−d = u−d (but we still use r as the discount rate)

Another way to derive p
Figure 4-11
payoffT ( )





  ( F0u  F0 )  cu




  Futures (at F0 ) cu  cd
 

“=”   =
 1  c

F0u  F0 d
 value( )  c


  ( F0 d  F0 )  cd




⇒ −c = [∆(F0 u − F0 ) − cu ]e−rT (the initial value of the futures is 0)


⇒ c = e−rT (p · cu + (1 − p) · cd ), where p = u−d
1−d

• The Black-Scholes model as well as the binomial tree model are versatile models:
Treat stock index, currency, and futures like a share of stock paying a dividend yield q.
For stock index options: q = average dividend yield on the index over the option life.
For currency options: q = rf .
For futures options: q = r.

4-17
V. Introduction of Combinatorial Method

• Combinatorics (組合數學) is a branch of pure mathematics concerning the study of dis-

crete (and usually finite) objects. Aspects of combinatorics include “counting” the objects
satisfying certain criteria, deciding when the criteria can be met, or finding “largest”,
“smallest”, or “optimal” objects.

• Combinatorial method for European options

Based on the binomial tree framework, applying the combinatorial method is far faster
than the backward induction procedure. In fact, it is not necessary to build the binomial
tree in the combinatorial method, which is another advantage of the combinatorial method
because is can save memory space for computer programming.

Figure 4-12

S0  u n  d 0
S0  u n 1  d
M
M

S0 S0  u n  j  d j
M
M K
M
M
S0  u1  d n 1
S0  u 0  d n

T is partitioned
into n time steps

n
   
−rT
P n n−j j n−j j n
European option value = e p (1−p) max(S0 u d −K, 0), where ,
j=0 j j
√ √
also denoted as Cjn , is the combination of j from n, u = eσ ∆t
, d = e−σ ∆t
, and p =
e(r−q)∆t −d
u−d
.

∗ For the binomial tree model, its complexity is O(n2 ), whereas for the above combina-
torial method, the complexity is O(n). The difference of required computational time is
substantial for a large number of n, e.g., n > 5000.

4-18
Appendix A. Binominal Tree Model for Jump-Diffusion Processes

• The content in this appendix belongs to the advanced content.

• Binomial tree model for the jump-diffusion process in Amin (1993):

 Consider Merton’s (1976) lognormal jump-diffusion process as follows.
dS
S
= (r − q − λKY )dt + σdZ + (Y − 1)dq,
where KY = E[Y − 1] and ln Y ∼ N D(µJ , σJ2 ). By the Itô’s lemma,
d ln S = (r − q − 21 σ 2 − λKY )dt + σdZ + ln Y dq = αdt + σdZ + ln Y dq.

 The lattice model proposed by Amin (1993):

Figure 4-13

ln S t  ln S0  t  2 t ln S i t  ln S 0   i t  2  t
ln St  ln S0  t   t ln S it  ln S 0   i t    t
ln S0 ln St  ln S0  t  ln S it  ln S 0   i t 
ln St  ln S0  t   t ln S it  ln S 0   i t    t
ln S t  ln S0  t  2 t ln S it  ln S 0  i  t  2  t

0 t  i t  nt  T

4-19
 Backward induction:
Figure 4-14

OQ 6 W  ' W OQ 6 W  D ' W  0 - V ' W

‫ڭ‬

OQ 6 W  'W OQ 6 W  D 'W  V 'W

OQ 6 W  'W OQ 6 W  D'W  V 'W
OQ 6 W OQ 6W  'W OQ 6W  D'W
OQ 6 W  'W OQ 6 W  D'W  V 'W
OQ 6 W  'W OQ 6 W  D'W  V 'W

‫ڭ‬

OQ 6 W  ' W OQ 6 W  D ' W  0 - V ' W

∗ Define the branching probabilities as follows.

 √
 prob(ln St+∆t − ln St = α∆t + σ √∆t) = (1 − λ∆t)pu + λ∆t∆N (k)
p(k) = prob(ln St+∆t − ln St = α∆t − σ √ ∆t) = (1 − λ∆t)pd + λ∆t∆N (k) ,
prob(ln St+∆t − ln St = α∆t + kσ ∆t) = λ∆tN (k), for − MJ ≤ k ≤ MJ

√ √
where ∆N (k) = N ((α∆t + (k + 12 )σ ∆t − µJ )/σJ ) − N ((α∆t + (k − 21 )σ ∆t − µJ )/σJ )
√
if k 6= ±MJ , MJ ≡ int( σ3σ
√ J ), ∆N (MJ ) = 1 − N ((α∆t + (MJ − 1 )σ ∆t − µJ )/σJ ), and
2
∆t √
∆N (−MJ ) = N ((α∆t − (MJ − 12 )σ ∆t − µJ )σJ ).
(Note that using λ∆t∆N (k) in the above way can fully capture the effect of the jump
component ln Y dq in the d ln S process.)

4-20
∗ Determine pu and pd by matching the mean and variance of the remaining diffusion process,
d ln S = αdt + σdZ
α∆t

 pu St u + pd St d = St e

pu St2 u2 + pd St2 d2 − (pu St u + pd St d)2 = St2 σ 2 ∆t,

 p +p =1
u d
√ √
where u = eα∆t+σ ∆t
and d = eα∆t−σ ∆t
. Analogous with the CRR binomial tree model,
we can infer
eα∆t −d
pu = u−d
.

∗ However if we consider the probabilities of (1 − λ∆t)pu and (1 − λ∆t)pd for the upward and
downward branches, respectively, when the jump component is introduced. The mean and
variance generated by the probabilities of (1 − λ∆t)pu and (1 − λ∆t)pd are no more St eα∆t
and St2 σ 2 ∆t.
α∆t

 (1 − λ∆t)pu St u + (1 − λ∆t)pd St d = (1 − λ∆t)St e

(1 − λ∆t)pu St2 u2 + (1 − λ∆t)pd St2 d2 − [(1 − λ∆t)St eα∆t ]2 ≈ (1 − λ∆t)St2 σ 2 ∆t,
 (1 − λ∆t)p + (1 − λ∆t)p = (1 − λ∆t)

u d

For the variance equation, it can be rewritten as (1 − λ∆t)[pu St2 u2 + pd St2 d2 − (1−
λ∆t)(St eα∆t )2 ] ≈ (1 − λ∆t) · St2 σ 2 ∆t, p √ that (1 − λ∆t) ≈ 1.
if ∆t is small such
Thus, the standard deviation become (1 − λ∆t)St σ ∆t. By comparing with the mean,
(1 − λ∆t)St eα∆t , one can find that the deviation of the mean from its true value p is more
serious than the deviation of the standard deviation from its true value because (1 − λ∆t)
is closer to 1 than (1 − λ∆t) when ∆t is small.
∗ To offset the above undesirable effects, we should solve pu and pd in the following system of
equations.
 α∆t

 pu St u + pd St d = S1−λ∆t
te

St2 σ 2 ∆t
pu St2 u2 + pd St2 d2 − (pu St u + pd St d)2 = 1−λ∆t

 pu + pd = 1


∗ In Amin (1993), only the adjustment for the mean is considered, and thus the probability
pu can be derived as
eα∆t
1−λ∆t
−d
pu = u−d
.

4-21
∗ Option value of a node can be expressed as
Option value of ln St
P √
= p(k) · (option value for ln St+∆t = ln St + α∆t + kσ ∆t)
−MJ ≤k≤MJ

 If the adjustment for the variance is also considered, one can adjust the grid size
√ of the
∗ ∗ α∆t+σ ∗ ∆t
σ
multinomial tree to achieve this goal. By defining σ = 1−λ∆t , u = e
√ , d∗ =
∗
√
eα∆t−σ ∆t , one can derive the following tree structure and the corresponding branching
probabilities.
Figure 4-15

ln S t  t  ln S t  t  M J  *  t

ln S t  t  ln S t   t  2 * t
ln S t  t  ln S t    t   *  t
ln S t ln St  t  ln St  t
ln S t  t  ln S t   t   *  t
ln S t  t  ln S t   t  2 *  t

ln S t  t  ln S t   t  M J  *  t

√
 prob(ln St+∆t − ln St = α∆t + σ ∗ √∆t) = (1 − λ∆t)p∗u + λ∆t∆N (k)


p∗ (k) = prob(ln St+∆t − ln St = α∆t − σ ∗ √ ∆t) = (1 − λ∆t)p∗d + λ∆t∆N (k) ,

∗
prob(ln St+∆t − ln St = α∆t + kσ ∆t) = λ∆tN (k), for − MJ ≤ k ≤ MJ

√ √
where ∆N (k) = N ((α∆t+(k+ 12 )σ ∗ ∆t−µJ )/σJ )−N ((α∆t+(k− 21 )σ ∗ ∆t−µJ )/σJ ) if k 6=
√
±MJ , MJ ≡ int( σ∗3σ√J∆t ), ∆N (MJ ) = 1−N ((α∆t+(MJ − 12 )σ ∗ ∆t−µJ )/σJ ), ∆N (−MJ ) =
√ √e
α∆t
−d∗
N ((α∆t − (MJ − 21 )σ ∗ ∆t − µJ )/σJ ), and p∗u = 1−λ∆t
u∗ −d∗
and p∗d = 1 − p∗u .

4-22