Upload
tu-shirota
View
221
Download
0
Embed Size (px)
Citation preview
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 1/9
3. The Wiener Process and Rare Events in
Financial Markets
The Wiener process works as a mathematical
model for continuous time stock price.
To make the process even more realistic oc-casional rare events causing jumps into the
sample paths are added by modeling with
some point processes.
Consider first the modeling the ”ordinary”events with Wiener process (Brownian mo-
tion) (see here for a heuristic construction of
the Brownian motion).
1
Definition 3.1: A Wiener process, W t, rela-
tive to a family of information sets {I t} (fil-
tration), is a stochastic process such that
1. W 0 = 0 (with probability one).
2. W t is continuous.
3. W t is adapted to the filtration {I t}.
4. For s ≤ t, W t − W s is independent of I s,with E[W t − W s] = 0 and Var[W t − W s] =
E(W t − W s)2 = t − s.
2
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 2/9
Equivalently we can define:
Definition 3.2: A random process Bt, t ∈[0, T ] is a (standard) Brownian motion if 1. B0 = 0.2. Bt is continuous.3. Increments of Bt are independent. I.e., if
0 ≤ t0 < t1 < · · · < tn, then Bt1 −Bt0, . . . , Btn −Btn−1 are independent.4. If 0 ≤ s ≤ t, Bt − Bs ∼ N (0, t − s).
-20
-10
0
10
20
30
80 82 84 86 88 90 92 94 96 98 00
Time
W ( t )
+std
-std
Figure 3.1: Three sample paths of a standard
Brownian motion.3
Properties:
(a) W t is an I t-martingale.
(b) W t has independent increments.
(c) E[W t] = 0 for all t.
(d) Var[W t] = t.
(e) The law of iterated logarithms
lim supt→∞
W t√ 2t log log t
= 1 (with probability one)
(f) W t/t → 0 w.p.1 as t → ∞.
(g) W 2t − t is an I t-martingale.
(h) exp{σW t − (σ2/2)t} is an I t-martingale.
4
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 3/9
Proof : (a) Let s ≤
t, then E[W t−
W s|I
s] =
E[W t − W s] = 0. So E[W t|I s] = W s.
(b) Let 0 ≤ s < t < u then W t − W s is I t-
measurable (W s, W t ∈ I t ⊂ I u) and W u −W t is independent of I t, which implies that
W u − W t is independent of W t − W s.
(c) 0 = E[W t − W 0] = E[W t].(d) Var[W t] = Var[W t − W 0] = t − 0 = t.
(e) Will not be proven, but means that if b >
1, for sufficiently large t, W t < b√
2t log log t.
(f) Follows from the law of iterated loga-
rithms.
(g) and (f) Left as exercises.
5
-40
-30
-20
-10
0
10
20
30
40
80 82 84 86 88 90 92 94 96 98 00
W_
t
Time
std
log-log
Figure 3.2: The law of iterated logarithms.
6
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 4/9
Example 3.1:
S t = S 0e(µ−12
σ2)t+σW t ,where S 0 > 0, and µ and σ > 0 are constants. This is an exampleof generalized Brownian motion, called also a Geometric Brow-nian motion, and serves as a basic model for stock prices. Thedistribution of log S t is normal with mean
log S 0 +
µ − 1
2σ2
t
and variance σ2
t. The continuously compounded return to thestock, per unit of time (e.g. if annual, then annualized cumula-tive returns), over time interval [t, t + u], u > 0, is
Ru = 1
u (log S t+u − log S t) = µ − 1
2σ2 +
1
uσ (W t+u − W t).
Because for fixed t, W t+u − W t is a standard Brownian motionwith time index u, property (f) above implies that
Ru → µ −1
2σ2
as u → ∞.
Note. E [S t] = S 0eµt.
7
0
400
800
1200
1600
2000
80 82 84 86 88 90 92 94 96 98 00
S ( t )
Time
E[S(t)]
Figure 3.3: Sample paths of monthly stock price processes S t =
S 0e(µ−12
σ2)t+σW t with S 0 = 100, and annual rate of return µ = 10%
and volatility σ = 25%.
8
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 5/9
-30
-20
-10
0
10
20
30
80 82 84 86 88 90 92 94 96 98 00
M o n t h l y r e t u r n ( % )
Month
Figure 3.4: Monthly log-returns Rt = 100(log S t − log S t−s) for
the sample period 1980:1 to 2001:12.
-60
-40
-20
0
20
40
60
80 82 84 86 88 90 92 94 96 98 00
A n n u a l i z e d c u m u l a t i v e r e t u r n ( % )
Month
Figure 3.5: Annualized cumulative log-returns RAt = 100 12
t (log S t−
log S 0) for the period 1980:1 to 2001:12 (monthly observations).
9
Remark 3.1: In the above example the geometric Brow-
nian motion is usually given in the differential form
dS t
S t= µ dt + σ dW t,
(”model of the return”) or
dS t = µ S t dt + σ S t dW t.
10
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 6/9
Rare Events and Poisson Process
In stock markets we occasionally observe sud-
den jumps in the prices due to some impor-
tant or extreme event that affects the price.
To model these a reasonable assumption couldbe that the jumps—being consequences of
extreme shocks—are independent of the usual
information driving innovations dW t.
11
Following Merton∗ we can introduce a model
where the asset price has jumps superim-
posed upon a geometric Brownian motion
such that
dS t
S t= (µ − λ ν )dt + σ dW t + dJ t,(1)
where µ the asset’s expected return, λ isthe jump intensity (e.g. average number of
jumps per year), ν is the average size of the
jump as a percentage of the asset price, σ
is the return volatility, and J t is related to
the (independent) Poisson process generat-
ing the jumps.
Once the jump occurs we can superimpose it
into the process by assuming that it is gen-
erated from a normal distribution with mean
ν and standard deviation equaling the jump
volatility.
∗Merton, R.C. (1976). Option pricing when underlying stock
returns are discontinuous. Journal of Financial Economics , No.
3, 125–144.
12
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 7/9
Example 3.2: Consider a situation where the underly-
ing process has a jump frequency one per year (λ = 1/2)(i.e., on average once every two year). The averagepercentage jump size ν is assumed for simplicity to beequal to zero and the volatility of the jump is 20%.
Suppose the stock price now is 100, the drift is 8%,σ = 15%. Below are four sample baths from the dif-fusion process
dS t
S t= µ dt + σ dW t + dJ t.
We model dJ t as dJ t = 0.20 Z U , where Z is a standardnormal variable and U is a random variable such thatP (U = 1) = λdt and P (U = 0) = 1 − λdt. In additionZ and U are independent.
A discrete approximation of the above difference equa-tion is
S t = S t−h
1 + 0.08h + 0.15
√ h Z t + dJ t
,
13
0
100
200
300
400
500
600
700
90 92 94 96 98 00 02 04 06
S1 S2 S3 S4
Sample paths of a jump process:S(t) = S(t-dt)*[1 + 0.08dt + 0.15 sqrt(dt) Z(t) + dJ(t)],dJ(t) = 0.2 Z U(t), with P(U(t) = 1) = 0.5 dt
Month
S(t)
14
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 8/9
We can rewrite the diffusion process in this example
asdS t
S t= (µ − λν )dt + σ dW t
if the Poisson event does not occur, and
dS t
S t= (µ − λν )dt + σ dW t + 0.20Z
if the Poisson event occurs.
It can be shown that the process then is
S t = S 0 exp
µ − 1
2σ2 − λν
t + σW t
J (N t),
where
J (N t) = exp N t
i=1
J i
with J i = 0.20Z i, and Z i are independent N (0, 1) ran-
dom variables. N t has the Poisson distribution, dis-
cussed more closely below.
15
In order to model the jump intensity, let N tdenote the total number of extreme (unordi-
nary) shocks until time t (counting process).
Then once the event occurs, N t changes by
one unit and remains otherwise unchanged.
Conceptually we can model this within an in-
finitesimal interval dt by
dN t =
1 with probability λ dt0 with probability 1 − λ dt
(2)
This causes a discrete jump (once it occurs),because the size does not depend on dt (the
size can be a random variable as in the pre-
vious example).
16
7/27/2019 mfdc3d
http://slidepdf.com/reader/full/mfdc3d 9/9
Poisson process has the following properties:
(1) During a small interval h, at most one
event occurs with probability ≈ 1.
(2) The information up to time t does not
help to predict the occurrence of the event
in the next instant.
(3) Events occur at a constant rate λ.
Remark 3.2: Within any time interval [t, t + h], ∆N t+h =N t+h − N t has the Poisson distribution
P (∆N t+h = k) = (λh)k
k! eλh, k = 0, 1, . . .(3)
E[∆N t+h] = λh = Var[∆N t+h].(4)However, if h is small
P (∆N t+h = 1) = (λh)e−λh ≈ λh(5)
as stated above. We denote ∆N t+h ∼ Po(λh). Thus
because N 0 = 0, we have in the previous example
N t ∼ Po(λt).
17