|
PureBytes Links
Trading Reference Links
|
Not picky, I enjoy the questions and peer review because it forces me
to rethink things I've taken for granted in enough detail to (try) to
answer coherently. That can only be good, for me at least.
I think there's some confusion because this started out as an
explanation of *my personal* robustness criteria and bled over into
other topics such as issue selection, not that that's bad. And oh by
the way, some of the robustness tools are great for *part* of issue
selection as I see it but not all. I think we're lost in the trees,
so let's step back and look at the forest.
You said you're trying to understand my thinking. OK, but remember
you asked for it! Let's take robustness and issue selection one at
time.
1. Robustness. You've seen criteria 1-5. Criteria 1 and 2 use the
same 6 years of bull, bear and sideways data. Criteria 3-5 use *all*
data that I have for those same stocks and this varies mostly from
early 80s to sometimes much less, especially w/small caps. So you
could say I'm doing some sort of OOS during robustness criteria 3-5 by
looking at performance graphs and simulated metrics based on all
data, for stocks that were initially selected by their performance
over only 6 years of data. But with that said, I'm actually trying to
use robustness to simplify and get away from all the OOS stuff once
and for all. With the simple idea being: find systems that work
most of the time on most stocks and then be concerned with the only
OOS that really matters anyway, the future. Keep things simple, use
blunt tools like graphs and the bootstrap and try to minimize
unhelpful human influence. Want to know what one of my "proprietary"
end dates is for one of my 2 year periods? My grandfather's 100th
birthday. Why? I believe picking a date that way is more robust than
thinking too much about it and somehow screwing up the range.
2. Issue selection. Purpose is to find the best stocks to trade with
the robust systems from the previous (robustness) phase. I personally
divide issue selection into 2 steps, nomination and confirmation.
Nomination is the creative part and Steve K could give you much better
input on this than me, such as rank ordering stocks by %profit/bar
(anyone know how to quickly do this for a large list)? But I am
researching this and trying to do better. Now once nominated by
whatever metric (perhaps only a single number) I like to confirm each
candidate by running them through full criteria 3-5. That gives me a
really good feel for potential future performance.
Now to try to answer your question about the anchoring in 1971. Note
that I had a caveat about more data being better. This is always true
for data from some processes but not always true for market derived
data. And recall that here we're concerned with the interaction
between a system and a nonstationary time series. To make a long
story short there's a tradeoff between having enough datapoints in the
basket to provide robust estimators and having so many that they're
biased toward a distant past that won't re-occur. I am however
personally more biased toward adding more recent data because I
believe they're more likely to be more relevant. The best place to
draw the line is literally a billion dollar question but my approach
has been a robust one... to simply go with what I have in my database
which, so far, has proven sufficient.
Now aren't you glad you asked?
A few questions for you... now that you've seen the robustness
criteria, what do you think overall and what would you say is the
concept's greatest strength? Weakness?
--- In amibroker@xxxxxxxxxxxxxxx, "Fred" <fctonetti@xxxx> wrote:
> I'm not sure I understand the difference in why one would want to
add
> data to the confirmation phase and not the selection phase or is
that
> your flavor of OOS testing ? That aside for a moment if as you say
> criteria 3-5 work better with more data then wouldn't one want to
> anchor some beginning point in time long ago like 1971 or whatever
> and use all data from then up through current POSSIBLY leaving some
> segment out for OOS testing ? and then as time goes along include
> more recent data as in sample for re-evaluation of the systems
> robustness and issue selection and confirmation ?
>
> Note: I'm not trying to be picky here, I'm only trying to
understand
> your thinking which in at least SOME ways appears to be getting
> closer to my own.
>
> > When you say issue selection, are you talking about the
nomination
> or
> > the confirmation part? If confirmation, it would be better to
*add*
> > the new data in as they become available because criteria 3-5
(when
> > used for robustness or confirmation) inherently work better
(within
> > limits) with more data. If nomination (for example using an
> > algorithm)maybe but I doubt it because I've never been able to
wring
> > much advantage out of OOS techniques (with data that exists) and
> that
> > sounds like a twist on OOS if I understand you correctly. But
still
> > worth looking into. I've been surprised before.
> >
------------------------ Yahoo! Groups Sponsor ---------------------~-->
Rent DVDs Online - Over 14,500 titles.
No Late Fees & Free Shipping.
Try Netflix for FREE!
http://us.click.yahoo.com/xlw.sC/XP.FAA/3jkFAA/GHeqlB/TM
---------------------------------------------------------------------~->
Send BUG REPORTS to bugs@xxxxxxxxxxxxx
Send SUGGESTIONS to suggest@xxxxxxxxxxxxx
-----------------------------------------
Post AmiQuote-related messages ONLY to: amiquote@xxxxxxxxxxxxxxx
(Web page: http://groups.yahoo.com/group/amiquote/messages/)
--------------------------------------------
Check group FAQ at: http://groups.yahoo.com/group/amibroker/files/groupfaq.html
Your use of Yahoo! Groups is subject to http://docs.yahoo.com/info/terms/
|