|
PureBytes Links
Trading Reference Links
|
<SPAN
class=269361804-19112003>thanks yuki, simple, clean solution for random
subsets.
<SPAN
class=269361804-19112003>
<SPAN
class=269361804-19112003>number of positions vs size of universe is interesting,
different for different systems I've found. if a system finds relatively few
opportunities, more stocks is usually good, since it should raise the exposure
and therefor the return. if it finds lots of entries, some of which are much
better than others, then the quality of your PositionScore matters more and more
as the unverse gets larger. rotational systems are very sensitive to the
numbe of stocks, in a difficult to scale way. the more stocks there are,
the more turnover, sometimes cutting positions short before they get a chance to
develop as intended.
<SPAN
class=269361804-19112003>
<SPAN
class=269361804-19112003>dave
<SPAN
class=269361804-19112003>
<BLOCKQUOTE
>
No
worries, Dave.
<FONT face=Arial color=#0000ff
size=2>
It's
quite easy, thanks to AB. I just ran an exploration and assigned
(addtocolumn) a random number to each stock. I then sorted the
stocks by that column and picked off however many stocks I wanted in my
list.
<FONT face=Arial color=#0000ff
size=2>
Yes,
the results correlate quite nicely, unless the system isn't very
"robust". Of course, some systems simply need a lot of stocks in
order to perform well. I guess I'm saying the the number of
positions (PosQty as most of us use) should be adjusted depending on the size
of the database. If I expect PosQty to be about 15 for the entire
universe, then I'll set it to about seven for half the universe and so
on.
<BLOCKQUOTE
>
<SPAN
class=104414103-19112003>Chuck, not to be lazy, but how did you go about
creating your random subsets? have you found that results from them
correlate well to All Stocks?
<SPAN
class=104414103-19112003>
<SPAN
class=104414103-19112003>dave
<SPAN
class=104414103-19112003>
<BLOCKQUOTE
>
<FONT face=Arial color=#0000ff
size=2>You and I are certainly singing from the same hymnal.
Thanks for saying it so succinctly!
<FONT face=Arial color=#0000ff
size=2>
<FONT
color=#0000ff>IMO, we should always be backtesting against
all stocks with all of the filtering going on in the AFL. Not
that there's anything wrong with preliminary testing on a (random)
subset. My stock database (that I use with AB) now contains
17,000 stocks including about 9,000 de-listed stocks.<SPAN
class=104414103-19112003> I
have random subsets of that database comprising 2,000, 5,000 and 11,000
stocks that I use when I'm in a hurry to see some results.<SPAN
class=104414103-19112003><FONT
face="Courier New">
<BLOCKQUOTE
>
<FONT face="Courier New" color=#0000ff
size=2>not chande, don't play him on tv,
but here's my take fwiw. IMO, the selection of the initial universe to
test is the result of another set of filters with another set of
parameters that aren't being talked about at the moment, and are just as
subject to (over)optimization as anything else.
<SPAN
class=579214201-19112003>
<SPAN
class=579214201-19112003>most fixed universes smaller than All Stocks
are really hard to backtest reliably without bias, so I don't when I'm
testing for real. for example, testing in '98 using the '03 N100 is
looking into the future and selecting only stocks that ended up big 5
years later -- a major long bias. even if you have historical N100
membership data, for example, it's hard to apply it on a continuous
basis over the life of a test.
<SPAN
class=579214201-19112003>
<SPAN
class=579214201-19112003>so, I often test ideas out on small subsets for
speed, but nearly always plan on running against all stocks eventually,
and build in the criteria needed to narrow the stocks considered as
required. if you do anything else, I'd suggest it needs some careful
thought.
<SPAN
class=579214201-19112003>
<SPAN
class=579214201-19112003>dave
<BLOCKQUOTE
>Dead
silence on a question of seemingly high importance means you
haveasked a dumb (really dumb) question. Maybe like letting
a loud fart inchurch.I'll go back to Chande and other
sources and try to determine why myquestion below was such a dumb
one.Thanks for taking pity on me and not saying
anything.Ken-----Original Message-----From: Ken
Close [mailto:closeks@xxxxxxxx] Sent: Tuesday, November 18, 2003
11:03 AMTo: AmiBroker ListSubject: [amibroker] Optimizing
Selection Criteria vs Timing SignalsHere is a (perhaps) semi
philosophical question:We are familiar with the various
pitfalls and prohibitions in optimizingand "curve-fitting" timing
systems.Do the same pitfalls and prohibitions exist for
keeping a constanttiming signal (whatever it is) yet optimizing
variables that alterbasket selection criteria? Can you
"curve fit" a set of selectioncriteria?It "feels" like the
problems would be different (and perhaps minimized)because of the
dynamic changes that occur, ESPECIALLY if stop conditionsare
used. OTOH, I suppose it would not be different using the
samecriteria on the same basket of stocks over the same time
period as thiswould yield the same baskets of stocks each
time. What I mean by thatis that for a given combination of
criteria, the same stocks would beselected.Hard to
visualize.Any comments on this
aspect?Ken
Yahoo! Groups Sponsor
ADVERTISEMENT
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 the Yahoo! Terms of Service.
|