Machine designed strategies. Do they work?

Discussion in 'Strategy Building' started by v75z52, Mar 20, 2012.

  1. Yeah, I get it. Good luck with that.

    No need to trouble yourself with a "Machine designed strategies" thread :confused:
     
    #31     Mar 25, 2012
  2. tim888

    tim888

    There are several. The most important is that you do not know how they do it. I have an example of a SPY search with daily data I downloaded from Yahoo. I checked the data and they are clean. I used a typical profit target and stop-loss of 4%. The in sample is from beginning of ETF trading to 12/31/2008. The OOS from 01/02/2009 to 03/23/2012. The search does not see the OOS. Here is the result of the search:

    [​IMG]

    It found 11 long and 1 short. This is just one example.
     
    • spy.jpg
      File size:
      245.2 KB
      Views:
      947
    #32     Mar 25, 2012
  3. tim888

    tim888

    Now here is the OOS test:

    [​IMG]

    Four patterns have a profit factor less than 1 out of the 11. The combined profit I calculated to be equal to $105.0261 per share. The buy and hold gain in that period was $27.33. This is a significant result. You can get these results and evaluate the program only with the demo. The only thing you cannot see is the code of the patterns.

    Apparently they know something you don't. I have gotten many more results like this one. The good thing about it is that you can figure out the best targets and stop-loss levels to use.

    Edit: I will add some filters to these patterns like a triangular MA sort of thing and the profit potential will explode towards the sky.
     
    #33     Mar 25, 2012
  4. No, it's not.

    The appropriate null consists of random longs with 4% target and 4% stop, not the buy and hold return. During the [carefully chosen OOS] period such random longs hit at a nearly 75% rate, for a nearly 3 pf in the null case.

    Also, keep in mind that with such a wide target/stop band, the trades will take a while to play out. They will overlap so that without a bankroll a large multiple of trade size, you will not be able to play them all.
     
    #34     Mar 25, 2012
  5. Good work Tim. Are you using the demo or the full version of the software?

    Yes, we will take your word for it...Give me a break...:)

    Why only long? If you decide to use only long you introduce data snooping bias. I see that Tim's system included a short pattern which was a loser. The null run is not what you decide after looking at the data and the original system.

    Essentially you say that all random long traders during the OOS made a pf = 3 which defies common sense, or that maybe some monkeys are driving around in a RR.

    I see now... You do not have to play multiple of size. You can move your stops and targets around if trades overlap (new trade signals come after one is already in place) or even use margin in case of concurrent signals. Increasing size doubles the risk but increasing target increases the profit at the same risk. This may result in some reduction of overall pf but the risk is also reduced and drawdown is less.
     
    #35     Mar 26, 2012
  6. This is all bullshit.

    It's real easy for a vendor to do an in-sample test, then say it was an out-of-sample test.

    "Oh look, it performs OOS !!"

    You guys are wasting your time with this mumbo jumbo approach to strategy design.

    ...

    Ok, now go on an claim to be billionaire algo traders.

    Then pitch your system again, rational economic agents.

    Bah.
     
    #36     Mar 26, 2012
  7. jcl

    jcl

    That's absolutely correct - but as not anyone seems to have understood this post, I'll repeat with other words.

    Tim888: I don't want to talk your test result away, but there are two obvious problems with it. First problem is the strong asymmetry in long and short. This falsifies the result completely when you're testing it with trending price data, because random trades with the same long/short asymmetry then give you the same positive result. The second, even worse problem is that you're testing a 3-bar pattern signal with 4% targets. This holds the trade far longer than the pattern duration, and makes the result completely uncorrelated to the pattern. So this test only shows that SPY was going up during that period, and tells us nothing of your price patterns.

    But you can modify your test and then get a significant result. First, use either detrended prices or set up PAL so that it generates the same number of patterns in short or long. Or if PAL does not support detrending, use an asset that has about the same price at the begin and end of the test period. You can find this in Forex for most periods. And of course you must not use stop or profit targets here, but you must exit the trade after a period that does not exceed the prediction horizon of the pattern - in this case, certainly not much more than about 3 bars. This could then reveal if a 3-bar price pattern has any significance.
     
    #37     Mar 26, 2012
    shuraver likes this.
  8. You haven't justified why this is a problem. You just talk out of your as* as always. Most successful trading strategies for equity indices are long only because of the longer-term bias and the fact that corrections are very fast and hard to capture.

    Again, you’re talking out of your as*. You have provided no proof of your statement. Testing with random longs also involves data snooping bias. You must include the same proportion of shorts as in the PAL case, at least.

    You again talking out of your as*. Who told you that those patterns have only 3-bars? They do not. I know PAL and most of its patterns are 4 and 5 bars long. Some have 6 bars.

    Then, any talk about a trend is not appropriate. If you are trading (I think you are not and possible you have never ) SPY moved sideways for all of 2011, a period long enough to kill most of systems, especially trend-following ones.

    I mean this is laughable. Have you looked at a longer -term chart of SPY? Since 1995 it looks like a sine wave. Why are you making statements like this before looking at charts?

    This is laughable. You are really talking out of your as*.
     
    #38     Mar 26, 2012
    shuraver likes this.
  9. Tim, do not feed the trolls

    [​IMG]
     
    #39     Mar 26, 2012
  10. jcl

    jcl

    Goodgoing: Might you be someone who has really paid $10K for that software? This of course explains your upset :). Spending $10K is not a good start into trading - better first read beginner's books, they come a lot cheaper. But please stay out of such discussions until you've read them. This saves forum space and bandwidth. Thank you.
     
    #40     Mar 26, 2012