|
PureBytes Links
Trading Reference Links
|
I guess step one would be a program that figures out how auto-correlated
returns are in a time series, by running a regression for instance.
Step two would examine the error term between the predicted one-day forward
return generated from the above regression.
So if Y is today's daily return, and X is yesterday's, you can first create
a simple linear model which solves Y=mX+c by the least-squares method or
something. Then apply the same coefficients to predict tomorrow's (unknown)
change, call it Z, from today's (known) one, by Z=mY+c [same constants].
That's not rocket science by any means. And is probably incompatible with
random walk hypothesis.
I think GARCH then looks at the distribution of E, the error between
predicted Z and actual Z outcome. Don't ask me what happens next - I have no
idea. I know a man that would be happy to explain it all you you all at
long, painful and totally incomprehensible length - he works for CSFP, and I
think he wrote a thesis on it. email address available on request. As far as
I'm concerned, if it can only be expressed as strings of Greek letters, it's
a little over my head.
Rus
-----Original Message-----
From: Gary Funck <gary@xxxxxxxxxxxx>
To: omega-list@xxxxxxxxxx <omega-list@xxxxxxxxxx>
Date: 12 February 1999 15:38
Subject: Re: GARCH
>
>On Feb 12, 9:22am, Rus Newton wrote:
>> ...
>> The 64kb limit on EL programs means you'd probably have to nest dozens of
>> the little chaps. Better to code something like that in C++ I imagine.
>> Perhaps one of the OSPs can help you with an addin? But I doubt that'll
come
>> cheap.
>>
>> Of course like you I'd be interested if someone has coded GARCH.
>
>I'd be interested if someone here could describe GARCH in terms that
>it can be programmed (in any language).
>
>
>--
>| Gary Funck, Intrepid Technology, gary@xxxxxxxxxxxx, (650) 964-8135
>
>
|