|
PureBytes Links
Trading Reference Links
|
comments below,
----- Original Message -----
From: "Michael Barrow" <profitabull@xxxxxxxxx>
To: <systems-only@xxxxxxxxxxxxx>; "Omega List" <omega-list@xxxxxxxxxx>
Sent: Friday, May 11, 2001 5:29 PM
Subject: Re: SO_Modeling The Market , study proposal... ( the unholy grail )
> Bilo:
>
> Interesting post. I'm not sure about exact terminology or
> classifications, but here's what I've been working on.
>
> I'm a end of day stock swing trader (not that it matters
> for the sake of this larger discussion). I'm paraphrasing,
> but I believe it's been said on this list before that
> "simple systems based on simple indicators suck". I
> believe this premise. Therefore, I have developed 2
> indicators of my own.
>
> One predicts the expected volatility of a stock over the
> next couple of days. The other predicts the expected price
> direction of a stock over the next couple of days. I have
> systematically combined together many indicators to create
> these "composite indicators". None of my selected
> component indicators are highly correlated with each other,
> and each of them by itself does a decent job of predicting.
> Put them together and they enhance each other's
> performance. My stock price predictor indicator is looking
> not just at the stock's price history but also at its
> industry group, as well as the market averages overall, as
> well as many commonly-used indexes such as interest rates,
> commodity prices, market indexes and their relationships to
> one-another, etc.
good post. what you are describing is what most of wall street
have been doing for a while. namely next day trend + volatility
predictions aka arima + garch approach. i think it somewhat
worth mentioning. arima tells what the trend is for the next day
and garch tells you what range to lock in... simple approach
not so simple solution. on a scale of 1-10 i would rate it about 2.
if you are the expert in arima or garch i would let you do it if you want
for the list
but very short.
>
> I boil all of this down and scale my indicators to a 1.0 to
> 10.0 scale. Through testing I have found that I want my
> volatility predictor to be at least a 7.0. I have also
> found that if my price predictor composite indicator is 7.5
> or greater, go long. If 3.5 or below, go short. These
> have given fairly consistent, decent signals over the last
> 10 years. I feel good about them because, believe it or
> not, they have performed best since 04/2000 when the market
> crashed and changed. I'm not a statistician, but I'm
> confident enough in my indicators that I am using them with
> my hard-earned money.
wait a second. looks like you might have a different solution
after all. what is this composite indicator that you are working with?
>
> Now, here's my question that I think might have some
> relevance to this thread. In addition to these predictor
> indicators, I have done a lot of research into
> pattern-based stuff like Japanese Candlestick patterns and
> most of the other typical Western patterns like inside
> bars, gaps, breakout patterns, etc.
ok. that's fine. the good old pattern probability studies.
and what did you find? any reliable patterns that
can do better than 50/50?
>
> On the one hand I have indicators that return numeric
> values, and I have combined these intelligently and scaled
> them to a 1.0 to 10.0 scale. On the other hand, I have
> boolean-type indicators that return a true/false value
> based on whether the stock's recent price history meets the
> criteria of the pattern definition.
>
> The question is, how to incorporate this boolean,
> pattern-based information into the numeric predictor
> indicators to enhance their information and performance? I
> am calculating a couple of numerical statistics for these
> boolean pattern indicators:
> 1) Number of bars that are true/false
> 2) Average Future percentage change (FPC) over the next
> couple of bars (let's use 5 as an example)
well you can do NN or you can do Fuzzy NN or you can
do just Fuzzy
>
> So say I have a pattern indicator and I test it on 100,000
> bars of daily stock data (for several hundred stocks all at
> once going back several years each). The pattern indicator
> returns true on 10,000 bars, which means it is firing 10%
> of the time. Let's also say that when it fires true, the
> average FPC 5 bars into the future is +1.25% and when it
> fires false, the FPC5 is +0.15%. Given this and nothing
> more, it would seem to me that this indicator is a fairly
> good predictor of price direction over the next 5 bars. If
> the pattern is tested and fires true, that gives me a
> probabilistic edge. OK, great. Now what? How can I
> combine the results from testing, say, 500 of these
> different patterns into my numerical predictor variable to
> enhance its performance?
you can use you 500 pattern probabilities as inputs to a NN maybe?
to tell you the truth i am no fan of either of PP or NN or
similar approaches. which means nothing. just i don't get turned on
by never endless tweaking.
>
> The other thing that complicates this is correlation
> between these pattern indicators. I might have Pattern A
> which is a bae pattern that gives decent results. I might
> also have Pattern B that is a superset of Pattern A plus
> one or two additional criteria that makes it fire less
> frequently, but when it does it gives much stronger signals
> (i.e. the FPC5 discrepancy between true and flase is
> wider). Since these are obviously going to be correlated
> results, I would want to test B first and if it fires true,
> incorporate something into my predictor. But then I
> wouldn't want to test Pattern A because it would be
> redundant and would skew the results toward what I was
> testing, not toward what was true.
>
> Has anyone played with these concepts? I remember vaguely
> from my college days 20 years ago that you can get into
> statistical trouble combining numerical, ordinal and
> boolean scales. But this pattern-based stuff is useful
> information and I don't want to ignore it.
Mike, i'd say your best bet would be NN and FNN
cuz you can mix any type inputs in there.
i am sure you are doing just that.
but... what is your success rate after all?
anything spectacular?
my interest in purely professional.
bilo.
>
> Michael Barrow
>
>
|