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

RE: [amibroker] Re: Profit Prediction Was: Auto-optimization AFL uploaded



PureBytes Links

Trading Reference Links




<SPAN 
class=523274613-30102003>thanks for jumping in Phsst, and no, you haven't 
overstayed your welcome, not by a long shot (:-). if I poke and prod in my 
return, it's exploration, not disagreement.
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>it's not exactly that I haven't developed any 
strategies that show a profit. that I've done, repeatedly. it's that nothing 
I've done has given me the confidence that it was arrived at by a process that 
consistently identifies profitable systems. <FONT 
face="Courier New" color=#0000ff size=2>that's 
really what auto-optimization is about, defining an objective process for 
evaluating systems that predicts profitability reasonably well. 

<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>to put it differently, if I test a system over one 
range of time and it does well, do I know what I have and haven't proved? what 
does it mean if someone says (not just here, btw) that they've backtested 
CCI(25) crossing zero on some index since 1988 and it works well enough that 
it's possibly a good general market timing signal? suppose I wanted to 
trade a stock universe better represented by a different index, and wanted to 
investigate a good period to use. I can optimize over all data to date, look for 
ranges of profitable settings, and pick one in the middle, as b and herman have 
suggested. but would I have gotten the same result if I'd done that in 2000, or 
'95? if not, why should I buy any one of those results, or what that other 
person said, or the conventional wisdom? most importantly, why should I have 
confidence in my process should I want to evaluate another system? 

<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>*that's* why I built the auto-optimizer originally, not 
to trade constantly re-optimized systems, though of course I wondered how well 
that would work too. but the failure of auto-optimization in its current form 
confirmed my feeling that I didn't really know how to predict the profit a 
system would make going forward. that means I shouldn't be taking my results in 
the system selection biz very seriously until I get my act together. make 
sense?
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>I don't think of myself as believing much of the 
"recommendations" I hear without investigation. if I was easier about that, I'd 
be trading this stuff sooner, instead of still testing and thinking (:-). also, 
I didn't "assume" that the test date range mattered a lot, I found that it did 
experimentally, at least the way I was testing. can you give me an example of a 
trading system or way of testing that's more stable over time? privately if you 
want, discretion assured.
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>it's interesting that you propose "complicating the 
issue by adding further pattern filters". most talk I've heard about systems 
that are truly robust in the long-term says the opposite, go simple. 
<SPAN 
class=523274613-30102003>more than once, I've tried taking only trades 
recommended simultaneously by several moderately profitable indicators, and 
haven't been that impressed.
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>anyway, how can you confirm that the 8500% return you 
backtest on a double-mojo-bear-star-winky-thing coincident with DEMA(29.5) 
of the VIX crossing 73.1 on alternate tuesdays when the moon is full isn't a 
curve fit?
<SPAN 
class=523274613-30102003> 
<SPAN 
class=523274613-30102003>dave
<BLOCKQUOTE 
>
  <...the general vibe 
  was that this strategy and settings were and hadbeen somehow golden for 
  forever. (of course I'm exaggerating; theseare smart, unpretentious 
  people, but the underlying logic still seemedlike that to 
  me.)>Dave,Before I start here I'll preface my comments with 
  the disclaimer thatI am relying on memory which is not a particularly 
  reliable thingwhere I am concerned...You're a bright guy. But 
  don't I recall that you posted as recently astwo or three weeks ago that 
  you have not yet developed a profitablemechanical trading system using AB? 
  Hey, I could be wrong and if I amthen I apologize.I should 
  probably stop here and wait for a response, but time is noton any of our 
  sides, so I'll press on <g>...I would speculate that your 
  original premise was correct in that a fewpeople here do in fact have 
  strategies and settings that have beenreliable over a *very* long period 
  of time and that there is noexaggreation in that assumption.< 
  but I kept hearing things like "CCI(something)", and "MA(this) 
  crossMA(that)", with some single test and/or optimization period to back 
  itup.>You cannot judge anything based upon what you think you 
  have heardhere. If nothing else, you can count on the fact that those who 
  havegood profitable trading systems here are NOT going to part with 
  themeasily.<I concluded that what you'd  recommend 
  depended to a huge degree onwhat dates you looked at. that's one of the 
  main reasons why I wrotethis, to automate optimization over a bunch of 
  time periods, and seeif things stayed constant and/or profitable.  
  neither. sowhattheheck.?My advice is to quit taking the 
  recommendations of others, and quitassuming that anything depends to a 
  'huge degree' upon what dates arelooked at.I'll go even further in 
  a positive direction:Look at conventional wisdom related to 
  indicators, but NEVER STOPTHERE... be sure to go the extra distance and 
  complicate the issue byadding EXTRA pattern filters to your 
  backtests.By combining conventional wisdom with pattern filtering, you 
  mightdiscover that there are in fact long term strategies that 
  makepredictable profits under known mkt conditions. As usual, I 
  always take too long to get to my point...You have abandoned the basic 
  search for 'root' strategies in favor ofdeveloping a complex 'auto 
  optimization' piece of code that seeks toferret profits out of improbable 
  strategies.Here is my advice to you...Focus upon backtesting 
  basic technical strategies combined withpattern matching. Then, only after 
  you find promising tradingopportunities should you seek to optimize the 
  various aspects of thestradegy.I might be wrong, but I suspect 
  that all of your effort related tothis 'auto optimization' code is not 
  going to yield you any resultsuntil you have done the more basic task of 
  developing 'core'strategies that show promise.Again, as usual I've 
  talked too much and worn out my welcome.Best 
  regards,Phsst--- In amibroker@xxxxxxxxxxxxxxx, "Dave 
  Merrill" <dmerrill@xxxx> wrote:>   Isn't there some 
  aspect of past performance (results) that can"persist"> for some 
  time.  That is to say, that a stock or mutual fund that is> 
  profitable for awhile (esp from a backtest or ranking perspective)does 
  not> and probably will not remain profitable "forever".  However, 
  it canand does> remain a good performer for some period.> 
  > you'd think so, wouldn't you? what's the point of 
  backtestinganything ever,> if the future can't be at least kind of, 
  sort of, approximately,> semi-fake-estimated from what happened in the 
  past?> > but I kept hearing things like "CCI(something)", and 
  "MA(this) cross> MA(that)", with some single test and/or optimization 
  period to backit up.> people didn't usually talk about optimizing 
  over one time period, then> testing with those optimal settings for a 
  while *after* that, then> optimizing again later, etc.. the general 
  vibe was that thisstrategy and> settings were and had been somehow 
  golden for forever. (of course I'm> exaggerating; these are smart, 
  unpretentious people, but the underlying> logic still seemed like that 
  to me.) when I optimized some of those same> things myself, over 
  different time periods, I concluded that what you'd> recommend depended 
  to a huge degree on what dates you looked at.that's one> of the 
  main reasons why I wrote this, to automate optimization overa 
  bunch> of time periods, and see if things stayed constant and/or 
  profitable.> neither.> > sowhattheheck.> 
  >   What parameter(s) measure this?  It probably changes 
  over time but it> seems a combination of return along with absence of 
  large volatilityis a> key measure.  The UPI or UI or even 
  CAR/Mdd seems like it capturesmore than> just raw return.  
  While this seems more true for mutual funds (actively> managed ones at 
  that), it can also be true for certain stocks in certain> 
  industries.> > > >   Just rambling out 
  loud..I have not had time to try yoursystem----can it> optimize 
  against CAR/Mdd? for example?> > > >   
  Ken> > as shipped, the framework comes with two metrics to 
  choose from,profit per> bar, and net bars on the right side of the 
  market per bar, meaningbasically> this:>   ((bars 
  long and rising + bars short and falling) - (bars long andfalling> 
  + bars short and rising)) / number of bars> > both metrics can 
  penalize the score by a selectable percentage of max> drawdown. for 
  instance, at zero penalty, MDD doesn't effect thescore; at> 50%, 
  20% MDD reduces the score by 10%; at 100%, by 20%.> > you can 
  also add your own metrics, AND OBVIOUSLY YOU SHOULD, 
  becauseneither> of these are that great, or none of the trading 
  rules tested are, or> something. another one I've thought of is some 
  kind of measure of the> smoothness of the equity curve, which is MDD is 
  a crude look at.many of the> standard metrics are something of a 
  pain to code on the results of the> Equity function. hopefully, at some 
  point tomasz will give us AB's full> suite of report metrics, as 
  functions that work off Equity.if youhaven't> check the framework 
  out yet, I'd suggest waiting for thenew-and-improved> v1.1, due out 
  soon.> > dave






Yahoo! Groups Sponsor












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.