|
PureBytes Links
Trading Reference Links
|
some robustness issues that have been rattling around in my head over the
weekend...
- the lite version of Robustitude #1 seems biased to the long side,
especially when testing older time frames. by definition, the S&P 100 and
NASDAQ 100 are stocks that grew big. if you're testing in '95, you're
peeking into the future just by selecting this collection of stocks.
I wonder what more representative cross-section we could use. maybe top
200-500 by volume? since liquidity is a primary screen for tradability
anyway in my book, this might not be a bad idea. using Andrew's N-th ranked
composite code (still haven't had a chance to work through it), we should be
able to test these stocks at any point in history that we've got data for,
or trade them in real time.
- if one of our first tests is 3 2-year periods, that's 6 years of in-sample
testing. I'm concerned that we're running short of out-of-sample years. the
QuotesPlus data I use goes back to 1990 theoretically, a total of almost 14
years. I'd be somewhat concerned that over 40% of that was in-sample, but
it's actually much worse than that, since nowhere near all current stocks
have been active that long. for example, of the 192 S&P 100 + NASDAQ 100
issues traded last friday, only 21 of them traded on 1/2/90 according to
QuotesPlus, 127 of them on 1/2/92, 160 on 1/3/95, and 189 on 1/3/00.
even more striking, out of the entire current QuotesPlus database, only 23
stocks traded over $1 on 1/2/90. or at least, only that many have QuotesPlus
data going back that far.
this also reinforces the last point: the older our tests get, the fewer
lite-version stocks are actually included, and the ones that are, are ones
that have had exceptional longevity.
- I know we're aiming for robustness, meaning wide applicability to future
results. even given that goal, how similar do we think 1990 or 1995's market
dynamics really are to today's, especially when you get to the level of time
constants or volume levels? how applicable to today's market is what we
learn from these older tests? how would be figure that out? and again, the
more we restrict our testing to recent years, the less out-of-sample time
there is.
- if one system does better in bull years and another in bear, the one that
does better in reality will depend on the proportion of bull and bear years
that actually occur. when we weight bull, bear and sideways markets equally,
are we matching their proportions in real life? what time frame would we
want to base that judgment on?
dave
------------------------ Yahoo! Groups Sponsor ---------------------~-->
Rent DVDs from home.
Over 14,500 titles. Free Shipping
& No Late Fees. Try Netflix for FREE!
http://us.click.yahoo.com/I3w.vC/hP.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/
|