(n)

Put Zk

(n)

= ln St

(n)

1 (n)
(n)
(n)
Zk = Zk1 + n hk1 n +
2

(n)
((k )21)

(n)
(n)
hk1 nk

(n)

where
(n)

(n)

= 0

(n) (n)

(n) (n)

(n)

+ 1 hk1 + 2 hk1 (k

(n)

c(n))2

(n)

(n)

(n)
n + 1 hk1n
n
n
(n)
(n) (n) (n)

(n) (n)
+ 2 2 hk1Yk
n + 3 hk1k
n. (3)
n
n

(n)

hk1 =

(2)

This is a slightly different and extended specification compared to

last time. Theres a plethora of GARCHes. This is called NGARCH;
a large negative return raises volatility more than a large positive.

Note that
(n)

Var(Yk

(n)

E(((k )2 1)2 )
2
(n) 4
(n)
E((k ) + 1 2(k )2)
3+12
=
=
=1
2
2

) =

With our knowledge of the Euler-scheme, its not hard to conjecture

the limit of (1).
2

GARCH can approximate 2-D diffusion.

and write

(1)
hk

hk

(n)

Now name to coefficients (left to right) 0 , ..., 3 , put Yk

, and consider a specification

(2) is harder, but the the trick is to write

(n)

Literature: Nelson (1990) is standard reference; Duan (1997) has

many more examples; this follows Frey (1997).
Set-up: [0; T ] is split op into n intervals [tk1; tk [ of length n = T /n.
Consider a GARCH-model living on these points. What happens when
n ?
Problem: No explicit time-step length dependence in our GARCHdefinition. We have to introduce that ourselves.

(k

(n)

(n)

(n)

+ (c(n))2 + 1)

(n)

We have that E((k )2 1)) = E(k ) = 0, and by normality (just

symmetry will do, actually) further that
(n)

(n)

(n)

(n)

E(k ((k )2 1)) = E((k )3 ) E(k ) = 0,

so we have a decomposition into uncorrelated variables.
(n)

Now plug this into equation (2) & subtract hk1 on both sides to get
(n)

hk

For each n we let {k } be a sequence of independent N (0, 1)distributed variables,

(n)

c(n))2 = ((k )2 1) + (2c(n)k

(n)

(n)

hk1 = 0

(n)

+ (1

(n) (n)

(n)

(n)

+ 2 ((c(n))2 + 1) 1)hk1
(n)

(n)

+2 hk1((k )2 1) + (2c(n))k

(n)

Note that (1) and (3) is almost the Euler-scheme for (4)-(5). Yk
has the right mean (0) the right variance (1) and the right covariance
with  (0). It only it were Gaussian ...
Fortunately, it can be shown that convergence (appropriately uniformly) of local first and second order moments suffices if the jumps

of the discrete processes vanish in the limit and here the -factor
takes care of that. (Standard reference is the somewhat incomprehensible Ethier & Kurz; I suspect results from Kloeden & Platen would
do too. )

If 2 = 0 you get a model where theres only one source of noise.

If c = 0 then theres 0 covariation between stock and volatility.
In the non-degenerate cases we have

(n)
(n)

1
1 + 2 n(c2 n + 1)
= 1
n
n
(n)

If this is to go to 0, then 1
(n)

But in a nutshell: (4)-(5) is the limit.

must be of the form

= 1 + 1 n 2 n

What do the convergence conditions on the s mean?

Assume that as n 0 we have

(n)

(n)

(n)

(n)

The condition 0 n 0 is OK if

(n)

0 1 1 2 2 3 3 .
n
n
n
n

(n)

Let W1 and W2 be independent Brownian motions. Look at the

solution to the SDE
p
1
(4)
dZ(t) = ( ht)dt + htdW1 (t)
2

dh(t) = (0 + 1ht)dt + 3htdW1 (t) + 22htdW2(t)

(5)

= n 0 .

Unless
(n)

= 2 n + o( n) and c(n) = c n + o( n)

then theres no convergence.

If 2 = 0 and c = 0 the we get a degenerate limit model; h isnt
stochastic. You can perfectly legally assume this. Its just a question
(n)
of how rapidly 2 and c(n) tend to 0.
7

So:

(1/252) 0.06847 = 2 2 = 1.087 ( 1.537 in front of dW2)

And: (1/252) 0 = 0.000221 0 = 0.05575432

And for 1

0.92523 = 1 + 1n 2 n

so 1 = 252 (0.92523 + 0.068478612 1) = 1.585

The long term level for h is
(n)

= 0.0351,
0=
(n)
(n)
1
1 2 2

and its is 0.188. Finally

dh(t) = 1.585(0.0351 h(t))dt + 1.537h(t)dW2 (t)

10

Estimation example
Crude application where we just detrend and forget the 1/2-fromIto-term and set c = 0. Daily observations so = 1/252. Recall
from (1) that in this case
Zt Zt

N (0, h(t))

> library(tseries,survival)
> SP500<-na.remove(get.hist.quote(instrument = "^spc", start = "1971-01-01", quote = "Close",origin="1960-01-01"))
> tst<-garch((returnSP500-mean(returnSP500))/sqrt(dt))
> tst$coef
a0 (beta_0^n) a1 (beta_2^n) b1 (beta_2^n)
0.0002212473 0.0684786126 0.9252386066
9