Tutorial on Itˆ o’s Formula
Zheng-Liang Lu
Department of Computer Science & Information Engineering National Taiwan University
November 18, 2019
Agenda
• Wiener process.
• Generalized Wiener process (Itˆo process).
• Itˆo integral.
• Martingale.
• Quadratic variation of Wiener process.
• Itˆo’s formula.
• Calculation examples.
• Black-Scholes option pricing theory.
Wiener Process
• A process W is called the Wiener process if the following conditions hold:
(1) W (0) = 0.
(2) The process W has independent increments: if r < s ≤ t < u, then W (u) − W (t) ⊥ W (s) − W (r ).
(3) For s < t, W (t) − W (s) ∼ N(0, t − s).
(4) W has continuous trajectories.
• Note that W has a nowhere-differentiable trajectory (see the next page).
Itˆ o Process
• A stochastic process X (t) is given by X (t) = a +
Z t 0
µ(s, X (s))ds + Z t
0
σ(s, X (s))dW (s), (1) where a is the initial condition of X (0), µ(t, X (t)) and
σ(t, X (t)) are two adapted1 processes, and W (t) is a Wiener process.
• The third item at the right-hand side of Equation (1) is to be defined.
1Let X and Y be stochastic processes. Y is adapted to FtX-filtration if Y is FtX-measurable.
Itˆ o Integral
• Let g be a process satisfying the following conditions:
• g is square-integrable, that is, Z b
a
Eg2(s) ds < ∞.
• g is adapted to the FtW-filtration.
• We define the Itˆo integral as follows:
Z b a
g (s)dW (s) , lim
∆t→0 n−1
X
k=0
g (tk) [W (tk+1) − W (tk)] .
• Why using theforward increments?
• Because we cannot foresee the future.
• Then the following relations hold:
E
Z b a
g (s)dW (s)
= 0, (2)
E
"
Z b a
g (s)dW (s)
2#
= Z b
a
Eg2(s) ds, (3)
andRb
a g (s)dW (s) is FbW-measurable.2
2We could say that the integral is deterministic at time b.
Sketch of Proof for Equation (2)
E
Z b a
g (s)dW (s)
≈ E
"n−1 X
k=0
g (tk)∆W (tk)
#
=
n−1
X
k=0
E [g (tk)] E [∆W (tk)] (∵ W (tk) ⊥ ∆W (tk))
= 0.(∵ E[∆W (tk)] = 0)
Sketch of Proof for Equation (3)
• For all i , j with i 6= j , we first calculate
E [∆W (ti)∆W (tj)] = E [∆W (ti)] E [∆W (tj)] = 0.
• Note that the first equality results from the property of independent increments.
• Then Equation (3) is proved as follows.
E
"
Z b a
g (s)dW (s)
2#
≈ E
n−1
X
k=0
g (tk)∆W (tk)
!2
=
n−1
X
k=0
Eg2(tk) E ∆W2(tk) + X
i
X
j
E [g (ti)g (tj)] E [∆W (ti)∆W (tj)]
=
n−1
X
k=0
Eg2(tk) E ∆W2(tk)
=
n−1
X
k=0
Eg2(tk)
(tk+1− tk)
→ Z b
a
Eg2(s) ds.
Martingale
3• A stochastic process X is called an Ft-martingaleif the following condition hold.
• For all t, E [ |X (t)| ] < ∞.
• X is adapted to the filtration {Ft}t≥0.
• For all s and t with s ≤ t, E [ X (t) | Fs] = X (s).
• Now let X (t) =Rt
0 g (s)dW (s) with 0 ≤ t0 < t.
• Then we have E
h
X (t) FtW0
i
= X (t0) +E
Z t
t0
g (s)dW (s) FtW0
= X (t0).
• By Equation (2), every stochastic integral is a martingale.
3It is a notion of fair games.
Digression: Is the Market a Martingale?
• For stock markets, the stock prices are not martingales.
• Consider that you deposit S (0) in the bank with r ≥ 0.
• Then S (t) = S (0)ert, which is riskless.
• Or we rewrite the equation above like
E[S (t) | F0] = ertS (0) ≥ S (0).
• This implies that the riskless asset is asubmartingale.
• Because ofrisk aversion, one should expect a higher return for taking higher risk, that is,
E[S0(t) | F0] > ertS0(0), where S0(t) is a process of one risky asset.
Digression: Risk-Neutral Valuation & Martingale
• Under a physical measure P, it is known that EPS0(t) F0 > ertS0(0).
• Let Y (t) = e−rtS0(t).
• Under the risk-neutralmeasure Q, the discounted asset price is a martingale because
EQ[Y (t) F0] = Y (0).
• This result is used to price derivatives as follows:
p = EQh
e−rTΠ(S0(T )) F0i , where p is the derivative price and Π is a stochastic contingent claim for S0 with the time to maturity T .
Quadratic Variation of Wiener Process
• Define ∆t = t − s and ∆W = W (t) − W (s) with s < t.
• By definition, we have
• E[∆W ] = 0,(∵ ∆W ∼ N(0, ∆t))
• Var[(∆W )] = ∆t.
• Now we are interested in the quadratic variation (∆W )2, which has:
• E[(∆W )2] = ∆t,
• Var[(∆W )2] = 2(∆t)2. (Why?)
• This is because the trajectory of W isrough!
• In differential form, it reads
(dW )2= dt.
• This identity will be used in the Itˆo’s formula.
Itˆ o Formula
• For convenience, notations are simplified unless necessary.
• For example, X (t) and µ(t, X (t)) are replaced by X and µ, respectively.
• In a differential form, Equation (1) is equivalent to
dX = µdt + σdW . (4)
• Let f be a C2-function.4
• Define the process Z by Z = f (t, X ).
• Then Z has a stochastic differential given by df = ∂f
∂t + µ∂f
∂x +1 2σ2∂2f
∂x2
dt + σ∂f
∂xdW . (5)
4The function f is said to be of (differentiability) class Ckif the derivatives f0, f00, . . . , f(k)exist and are continuous.
Sketch of Proof for Itˆ o Formula
• It is known that the second-order Taylor expansionfor f is
df = ∂f
∂tdt+∂f
∂xdX +1 2
∂2f
∂x2(dX )2 + ∂2f
∂t∂xdtdX + 1 2
∂2f
∂t2(dt)2 .
• We then calculate (dX )2 with the identity (dW )2= dt so that (dX )2= µ2(dt)2+ 2µσdtdW + σ2(dW )2
∼ σ2dt.
• Note that · · · is negligible compared to the dt-term.
• As a result,
df = ∂f
∂tdt +∂f
∂xdX +1 2
∂2f
∂x2(dX )2 +1 2
∂2f
∂t2(dt)2+ ∂2f
∂t∂xdtdX
= ∂f
∂tdt +∂f
∂xdX +1 2
∂2f
∂x2σ2dt
= ∂f
∂tdt +∂f
∂x(µdt + σdW )+1 2σ2∂2f
∂x2dt
= ∂f
∂t + µ∂f
∂x +1 2σ2∂2f
∂x2
dt + σ∂f
∂xdW .
• Hence the proof is complete.
• Note that · · · is used as the second form of Itˆo’s formula.
Example 1: E [W
4(t)] =?
• Define Z by Z (t) = W4(t).
• Then we have
∂Z
∂W = 4W3, and
∂2Z
∂W2 = 6W2.
• By the Itˆo formula,
dZ = 6W2dt + 4W3dW with Z (0) = 0.
• Written in integral form, this reads Z (t) = 0 + 6
Z t 0
W2(s)ds + 4 Z t
0
W3(s)dW (s).
• Taking the expected value on the equation above, the stochastic-integral term will vanish.
• So we have
E [W4(t)] = 6 Z t
0
E [W2(s)]ds = 6 Z t
0
sds = 3t2.
• Note that the exchange between doing an integration and taking an expected value works in most cases of financial math.5
• This result could be used to prove Var[(∆W )2] = 2(∆t)2.
5See Fubini’s theorem.
Example 2: E [e
αW (t)] =?
• Define Z by Z (t) = eαW (t) with Z (0) = 1.
• The Itˆo formula gives us dZ (t) = 1
2α2eαW (t)dt + αeαW (t)dW
= 1
2α2Z (t)dt + αZ (t)dW (t).
• In integral form, this reads Z (t) = 1 +1
2α2 Z t
0
Z (s)ds + α Z t
0
Z (s)dW (s).
• Why bother?6
6One can rewrite the stochastic process in form of · · · dt + · · · dW via the Itˆo formula. Starting from this form, it is easier to derive the expected values associated with the stochastic process. For most time, you cannot derive
• Now define m(t) = E [Z (t)] and differentiate the resulting equation as follows:
dm(t) = 1
2α2m(t)dt.
• Using the ODE technique7, we have
m(t) = E [eαW (t)] = e12α2t.
• Note that E [eαW (t)] is the moment-generating function (MGF)8 of W (t) so that you may follow the definition of MGF to produce the same result.
7To be more specific, you need the identity dxx = d ln x .
8See https://en.wikipedia.org/wiki/Moment-generating_function.
Example 3: R
t0
W (s)dW (s) =?
• Define Z by Z (t) = W2(t).
• By the Itˆo formula,
dZ (t) = dt + 2W (t)dW (t).
• In integral form this reads
Z (t) = W2(t) = t + 2 Z t
0
W (s)dW (s).
• So we have
Z t
0
W (s)dW (s) = W2(t) 2 −t
2.
• The second term in the RHS differs from the ordinary calculus!
Example 4: Geometric Brownian Motion (GBM)
• Let µ and σ be constant, and W be under the P measure.
• A GBM is given by
dS = µSdt + σSdW .
• Now take X = ln S with X (0) = ln S0.
• It is easy to see that
∂X
∂S = ∂(ln S )
∂S = 1 S, and
∂2X
∂S2 = ∂2(ln S )
∂S2 = − 1 S2.
• By the Itˆo’s formula, dX = ∂(ln S )
∂t dt+∂(ln S )
∂S dS + 1 2
∂2(ln S )
∂S2 (dS )2
= 1
SdS+1 2(−1
S2)S2σ2dt
= 1
S(µSdt + σSdW )−1 2σ2dt.
= (µ −1
2σ2)dt + σdW .
• In integral form, this reads ln S = ln S0+
Z t 0
(µ −1
2σ2)dt + Z t
0
σdW
=ln S0+(µ −1
2σ2)t+σW (t).
• This gives us
ln S (t) ∼ N
ln S0+(µ −1
2σ2)t,σ2t
.
• Note that the price volatility of one asset is σ√ t.
• In the end, we have
S (t) = S0e(µ−12σ2)t+σW (t),
which follows a so-called lognormaldistribution with E [S (t)] = S0eµt,
Var [S (t)] = S02
e(2µ+σ2)t− e2µt
.(Why?)
Exercise: Futures Price
• Assume that S (t) follows a GBM.
• It is known that the futures price F (t) is given by F (t) = S (t)er (T −t).
• By the Itˆo’s formula,
dF = (µ − r )Fdt + σFdW .
• If we shift to the Q measure (i.e., µ is replaced by r ), then dF = σFdW
with E [F (t)] = F0, which is a martingale.
Exercise: Product of GBM Processes
• Let Y and Z be two GBM processes:
dY
Y = a dt + b dWY, dZ
Z = f dt + g dWZ, where dWY and dWZ has correlation ρ.
• Consider the product of two GBM processes, U = YZ .
• By the Itˆo’s formula, dU = Z dY + Y dZ + dY dZ
= YZ (a dt + b dWY) + YZ (f dt + g dWZ)+
YZ (a dt+b dWY)(f dt+g dWZ)
= U [(a + f +bg ρ)dt+ b dWY + g dWZ] .
• Rewrite the above equation as below:
dU
U = (a + f + bg ρ)dt + b dWY + g dWZ.
• We show that the product of correlated GBM processes thus remains a GBM.
• In particular, we can also show that Sn is also a GBM process for n ∈ N.
Exercise: Quotients of GBM Processes
• Consider the quotient of two GBM processes, U = Y
Z,
where Y and Z are drawn from Example 6.
• By the Itˆo formula, dU = 1
ZdY − Y
Z2dZ − 1
Z2dYdZ + Y Z3(dZ )2 ...
= U(a − f +g2− bg ρ)dt + b dWY − g dWZ .
• This example reminds us to collect alldt-terms.
Example 5: Vasicek Model
9• X is a Vasicek process, defined by
dX = κ(θ − X )dt + σdW , with θ, κ, σ > 0.
• Let Y = eκtX .
• By the Itˆo’s formula, we then have dY = κeκtXdt + eκtdX
= κeκtXdt + eκt(κ(θ − X )dt + σdW )
= κθeκtdt + σeκtdW .
9Vasicek (1977). It is one of extension of the Ornstein-Uhlenbeck process, proposed by Ornstein and Uhlenbeck in 1930. Now the Vasicek model is out-of-date. The main focus aims at the LIBOR market model (LMM).
• So it reads
eκtX = X0+ Z t
0
κθeκsds + Z t
0
σeκsdW .
• Moreover, we could calculate
E [ X ] = X0e−κt+ θ(1 − e−κt), Var [ X ] = σ2
2κ 1 − e−2κt .
• As t → 0, it is easy to see that E [ X ] = X0 and Var [ X ] = 0.
• As t → ∞, E [ X ] = θ and Var [ X ] = σ2κ2, which is finite due to the mean-reverting property!
• Note that X is a process following a normal distribution.
(Why?)
Black-Scholes Option Pricing Theory
• Assume that the stock price St follows a GBM (see p. 22).
• For this stock, we now consider to sell a European call option which expires in time T and has the payoff function
Φ(ST) = (ST − K )+.
• Insert a figure as an illustration of options.
• Define the call price Ct= f (t, St).
• By the Itˆo’s formula, df = ∂f
∂tdt + ∂f
∂St
dSt+1 2
∂2f
∂St2(dSt)2
= ∂f
∂tdt + ∂f
∂St
(µStdt + σStdW ) + 1 2
∂2f
∂St2σ2St2dt
= ∂f
∂t + µSt
∂f
∂St +σ2St2 2
∂2f
∂St2
dt + ∂f
∂StσStdW .
• What is the fair priceof this call option?
• The no-arbitrage principle comes into play.
• Construct a riskless portfolio as follows: buy ∆ = ∂S∂f
t shares of the stock and sell one European call.
• The portfolio value V is V = ∆ × St− f .
• For a small variation of St,
dV =∆ × dSt− df . (6)
• If the market is free of arbitrage, then the risk-free asset must earn the risk-free rate, denoted by r > 0.
• This gives us
dV = rVdt =r (∆ × St− f )dt. (7)
• Now equate (6) and (7):
r (∆ × St− f )dt=− ∂f
∂t +σ2St2 2
∂2f
∂St2
dt.
• Hence we derive the famous Black-Scholes PDE as follows:
∂f
∂t + rSt
∂f
∂St +σ2St2 2
∂2f
∂St2 = rf . (8)
• Define ∆ = ∂S∂f
t, Θ = ∂f∂t, and Γ = ∂S∂2f2 t.
• Then we have another representation of BS-PDE as follows:
Θ + rSt∆ +σ2St2
2 Γ = rf .
• If one considers the delta neutral (∆ = 0), then the previous equation becomes
Θ + σ2St2
2 Γ = rf .
Feynman-Kac
10Theorem
• This discovery bridges two research domains (PDE andSDE)!
• If f (t, x ) with t ∈ [0, T ] is a solution to
∂f
∂t + µ(x )∂f
∂x +1
2σ2(x )∂2f
∂x2 = rf , f (T , x ) = Φ(x ), then f (t, x ) has a representation
f (t, X ) = e−r (T −t)EQ[ Φ(XT) | Xt = x ], where X follows a Itˆo process.
10Mark Kac (1914–1984), a Polish American mathematician.
• Now replace X by S .
• Then the call price is
C = f (0, S0) = e−rTEQ[ (ST − X )+].
• This is calledrisk-neutral valuation.
• The price of European call options is
C = S0N(d1) − Ke−rTN(d2), (9) where N(·) is a cdf of a standard normal distribution,
d1 = log(SK0) + (r − σ22)T σ√
T ,
and d2 = d1− σ√ T .
References
• John Hull, Options, Futures, and Other Derivatives, 10/e, 2018.
• Tomas Bj¨ork, Arbitrage Theory in Continuous Time, 2009.
• Steven Shreve, Stochastic Calculus for Finance II:
Continuous-Time Models, 2010.
• Y.-D. Lyuu, lecture slides of Principles of Financial Computing, 2019.11
11See https://www.csie.ntu.edu.tw/~lyuu/finance1.html.