TSR - resuscitating trading systems - page 10

 
voltair:
I do. Give me a scoop on how to do it (if anything, you can do it in person), pls!
You don't have to do it in person - you just have to make the process iterative, which is technically a bit more complicated, but it's a hell of a lot more revealing.
 
TheXpert:
You don't need to do it personally - you just have to make the process iterative, which is a bit more technical, but it's a hell of a lot more representative.
I don't get it. What's in an iteration? Optimize-test-trade? Doesn't work - do it all over again?
 
voltair:
Didn't catch it. What's in the iteration? Optimisation-test-trade? It didn't work - all over again?

In other words, to turn almost the entire test period into an OOS.

joo:

But the amusing thing is that just yesterday (before the activation in several branches Reshetov), I voiced a specific method (somewhat different from Reshetov), which allows you to uniquely establish the presence of the found patterns TC, and there is a possibility of identifying multiple patterns simultaneously (the number is limited only by the hardware capabilities and the amount of available history for analysis). I announced the local respected forum member in a personal conversation (if he sees fit, confirm the existence of the method).

Yeah, a productive conversation, for sure.

 
TheXpert:
You can do it without personal data - you just have to make the process iterative, it's a bit more complicated technically, but it's a hell of a lot more indicative.

Is it possible that joo has indeed found some way to forge a pattern? But that very method is probably more revealing than the demo example from the first post of this thread, but hardly more revealing than my workhorse:


1. joo reports a large amount of history to be loaded and high comp performance requirements. I use R-Net, on the same amount of 2 months Sample and 1 month OOS, performance requirements are minimal - celron with 500 MB RAM - could not find less performance, because EA on mql only runs a test on 2 months and saves the results as signals from indicators preceding a trade and profit obtained after its closure in the file in CSV format. An external program takes this data, breaks it down into two parts and formalizes it in mql code and outputs it as a ready-made EA with a single filtering mode. The only thing left is to test this EA on OOS to make sure everything is OK and we can trade. Plus R-Net makes it so that not a single trade at Sample is losing, i.e. initial filtration at TC level is 100%, and therefore still at this stage all false and true trade signals are clearly separated, i.e. even if there is any error it is only at the last stage and it is minimal.

2. joo says that his method requires constant additional optimisation. My profits on OOS are steadily increasing within 3 to 5 months. So there is no need to systematically and continuously rake up the computer.


So kids, you can take with you to your grave the terrible secret of joo's top-secret confidential method for forging patterns, as it is of no interest at all, at least to me, for sure.

 
Reshetov:

Is it possible that joo has indeed found some way to forge a pattern? But that very method is probably more revealing than the demo example from the first post of this thread, but hardly more revealing than my workhorse:


1. joo reports a large amount of history to be loaded and high comp performance requirements. I use R-Net, on the same amount of 2 months Sample and 1 month OOS, performance requirements are minimal - celron with 500 MB RAM - could not find less performance, because EA on mql only runs a test on 2 months and saves the results as signals from indicators preceding a trade and profit obtained after its closure in the file in CSV format. An external program takes this data, breaks it down into two parts and formalizes it in mql code and outputs it as a ready-made EA with a single filtering mode. The only thing left is to test this EA on OOS to make sure everything is OK and we can trade. Plus R-Net makes it so that not a single trade at Sample is losing, i.e. initial filtration at TC level is 100%, and therefore still at this stage all false and true trade signals are clearly separated, i.e. even if there is an error it is only at the second stage and it is minimal.

2. joo says that his method requires constant additional optimisation. My profits on OOS are steadily increasing within 3 to 5 months. I.e. there is no need to systematically and continuously race the comp.


So kiddies, you can take the terrible secret of joo's top secret confidencial method of forging regularities with you to the grave, as it's of no interest at all, at least to me, for sure.

Reshetov, you have somewhat misunderstood my posts.

1. Indeed, you don't. Large volumes of history will be needed, if you want (for example for the sake of sporting interest) to find as many regularities as possible. For a profitable trade, 1-2 regularities will be quite enough, and the large date is no longer necessary.

2. Additional optimization is desirable, but not obligatory for all the TC in general, not only for mine in particular, and this "desirability" results precisely from the changeability of patterns over time. I don't need to race the computer all the time, i.e. to engage in compophilia.


I sincerely wish you all the best!

 
Joo, is what you discussed with TheXpert related to the approach I mentioned? If I understand correctly, you mean looking through the non-stationarity prism? In that case, indeed, quite a lot of computation is needed.
 
joo:


And to you, sincerely, I wish you all the best!

OK. So the question of "where is the line between fitting and real patterns?" can be considered resolved, as it is completely useless at the moment:


1. There are various methods for identifying patterns from fitting and methods for identifying and eliminating false signals

2. Fitting is not evil, but a source of information about patterns or false signals

 
joo:

But the funny thing is that just yesterday (before Reshetov's activation in several threads) I voiced a specific method (somewhat different from Reshetov's), which allows to uniquely identify the presence of the patterns found by TC, and there is a possibility to identify multiple patterns simultaneously (the number is limited only by the hardware and the amount of history available for analysis).

By the way, it is twofold funny, I had also solved this very dilemma on that very day. Either the stars are so aligned, or "smart" people walk in groups, or Nibiru opens our chakras, or maybe telepathy?) Miracles. But the fact is there. Now finished a very comprehensive testing of the methodology, at least 80-90% of my NS training results are able to work successfully. I can't believe it myself...
 
Figar0:
By the way, it is doubly amusing, that very day I have also solved the same dilemma for myself. Either the stars are so aligned, or "smart" people walk in groups, or Nibiru opens our chakras, or maybe telepathy?) Miracles. But the fact is there. Now finished a very comprehensive testing of the methodology, at least 80-90% of my NS training results are able to work successfully. I can't believe it myself...

Only I had it all much earlier. To find out the date, check out the last thread about the fringe. I tried to explain something about noise on that day, but everyone was making noise and no one was listening. Paukas also made a good joke about noise being only in the head. It was clear to me, that it is useless to explain, because (I will not specify a parable about some beasts and beads). What I needed was a concrete example. In the evening I finally formulated the principle of the filter. And the next day the first Expert Advisor was ready and it yielded positive results on forwards.


Let's say it took me more time to select TSs for the demo examples because many of them were more or less successful in forward testing after at least one fitting. And we needed both of them to pass OOS without filter.


Not to mention the fact that I've already written an entire article in the meantime, which later turned out to be unsuccessful for the chosen topic. And the demo was matched to this very article.


Such are the pies.


So the stars did not align simultaneously. Although what the fuck does it matter, the most important thing is that the results of almost everyone (except for some stubborn whiners who have no use in proving or demonstrating anything) have converged in profit and now the question of (un)promising TA can be put a full stop.

 

Hmm. That's all fine, really happy for everyone. I've had ~70-90% of selected systems showing some sort of survivability time, too, and for a long time (a year and a half). But here's wishing not empirical methods, but mathematical ones.

By the way, I already mentioned one (of) similar to Reshetov's proposed idea. It is in general additional filtering. The filter idea is simple, but not very easy to implement. So, I arbitrarily took some TS, placed it on another symbol and timeframe, applied an additional filter to Sample ( no parameters of the TS are changed) and looked at the OOS - it works. :)