[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Significance of Statistics in System Testing



PureBytes Links

Trading Reference Links

Just one comment.
There is a statistical test called, I think, Student's t test.
What this test does is tell you if you have enough data and indicates how
much more data you need.
For example, you run 9 tests.  The t test will tell you if you have enough
data for the level of confidence you require, or how many more pieces of
data (samples) you need for this level of confidence.  Its not infallible,
and has been abused by poll takers

Lionel Issen
lissen@xxxxxxxxx
----- Original Message -----
From: Guy Tann <grt@xxxxxxxxxxxx>
To: <metastock@xxxxxxxxxxxxx>
Sent: Monday, April 17, 2000 5:41 PM
Subject: RE: Significance of Statistics in System Testing


> Rudolf
>
> I hope you don't mind, but I've added a few comments to your post.  Again,
> these are just my opinions.
>
> Regards,
>
> Guy
> Fax (630) 604-1589
>
> -----Original Message-----
> From: owner-metastock@xxxxxxxxxxxxx
[mailto:owner-metastock@xxxxxxxxxxxxx]On
> Behalf Of rudolf stricker
> Sent: Monday, April 17, 2000 8:42 AM
> To: metastock@xxxxxxxxxxxxx
> Subject: Re: Significance of Statistics in System Testing
>
> On Mon, 17 Apr 2000 07:26:45 +0530, you wrote:
>
> >In a system testing it is said that more the events (statistics)
> > the better would be the result.
>
> This may not be correct, because market characteristics can definitely
> change over time. If so, it does not make much sense to fit the system
> to "old behavior". Moreover, it might make sense to emphasize the most
> recent market behavior, e.g. by weighting over time.
>
> Very astute!  I'm impressed.  This is something my dad has said for years
> and something most traders don't agree with.  It's not what worked well 5
> years ago, but what's working in the last six months.  It has taken me a
> long time to accept this premise, as I want everything to work the same
way,
> forever.  Unfortunately, the market is dynamic, not static.  Emotion,
> weather, famine, war, euphoria and the gamut of human psychology all
impact
> the markets at any given time.  I'll save more on this for your last
> remarks.
>
> >This also tells that a system
> >having lesser number of events is doubtful?
>
> This in general seems to be true, because more events lead to more
> (statistically) reliable test samples.
>
> >If you have data
> >of a longer period than getting a large number of statistics is
> > not a problem but if you have a data of lesser number of
> >periods than what to do? Does it mean, that one cannot apply
> >the system to a stock which has a less trading history
> > (say 2-4 years).But a question arises how many events a system
> > should produce to draw a reasonable conclusion?
>
> You may test this e.g. by constructing (probability) distributions of
> price changes for different time periods. If the distributions do not
> change "too much" e.g. in terms of tail-iness, you may use the shorter
> time period.
>
> >If a system looks into one single pattern it may produce lesser
> >events. So can we say that the pattern is doubtful because of
> >lesser statistics? What is the solution?
>
> Imo, in most situations it does not make much sense to use only one
> single system, because in general any system works like a filter,
> which is looking only for (a) special pattern(s). This way, all other
> trading strategies (or profitable patterns) are not considered by this
> system.
>
> A solution might be to use _a bunch of systems_ directed to different
> profitable patterns. This kind of compound system may smoothly
> maintained and optimized by a genetic algorithm, which can handle
> "generations" of systems.
>
> Again I'm impressed.  This is exactly what we have done, and why it has
been
> so hard for me to put it into words.  We have a bunch of unique systems
that
> we have developed, that all work fairly well by themselves.  By combining
> these various systems (I call them indicators, but reflecting on your
words,
> systems would be a better choice) with weighting factors, we are able to
> 'statistically' recognize different markets.  If we have discovered
nothing
> over the last 50 years, it's that we're attempting to develop
methodologies
> to 'forecast' a constantly changing environment.  Since we started
building
> these systems before there were hand calculators, computers, etc., we were
> forced to develop a certain 'feel' for numbers and the marketplace.  This
> 'feel' is something that only comes from experience.  There is absolutely
no
> way that we would be able to develop our existing system and techniques
> using the tools available today.  I'm not saying that's good or bad, just
> that what we do could never be developed using Metastock, for example.
> Still we use Metastock to test new ideas and techniques, but I have to
> admit, most of our new stuff is developed in Clipper.
> Anyway, just some additional comments I thought I'd throw into the mill.
>
> mfg rudolf stricker
> | Disclaimer: The views of this user are strictly his own.
>