Where is the line between fitting and actual patterns? - page 25
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
If anyone finds inaccuracies in the scheme itself, please let me know...
Put the dimensions true, but still a muddy image. In the separate window - clear.
if instead of "test runs" make "test run" then both schemes are identical:
but you can keep the "runs" only they should be an order of magnitude less than the optimization runs. The fewer the better (less chance of fitting). Better still 1 run. That is why there is a difference in the number of OOS runs. In the second scheme it is only one. This is better from the viewpoint of fit, but somewhat worse from the viewpoint of only one set being checked (FN). If the difference between the sets is only a Mach period, then of course there's no reason to test several sets with OOS. But if the difference is more global - for example the way the position is supported, then it might make sense. Then you can take the extra risk of adjustments (compared to a single OOS run). Some options can be used in such a way that with their different values there will be fundamentally different systems and it makes sense to run them separately on OOS, rather than choosing one FN
P.S. is the "G-point" where I thought it would be? :) Then the scheme takes on a new meaning))))
3)What will you see in ten consecutive OOS that you cannot see at 2H? Reminder, this is the central question....
So, what will I see miraculous that I won't be able to see at the moment of 2H?
Without dividing the 8 months into parts, at the moment of 2H with good results there may be sets that give growth in the first 6 months and just "not plummeting" in the next two, or ALREADY plummeting a little. How will you filter them out? And by the way, why doesn't it say anything about the 2H moment? "Analysis of results and criterion selection of sets"? Looking at the balance/equity curve? Hmmm. Long.
By taking a 6-month run out of 8 months, it is possible to select sets that give stable growth with the right characteristics. The next OOS run for 2 months will sift out of them, those that do not give the growth of desired characteristics in these 2 months (periods taken from the scheme, for different strategies, the amount of time needed differently). Accordingly, I exaggerate of course, but 10 consecutive going OOS will give us check of needed characteristics of quality of growth at 10 periods, which will bring us closer to the searched goal - search and confirmation of stable system at all discrete periods, instead of at huge time interval.
The advantage is the search time. And the quality of the search.
So, what will I see marvellous that I can't see at the 2H moment?
Without dividing the 8 months into parts, at the time of 2H with good results there may be sets that give growth in the first 6 months and just "don't drain" in the next two, or ALREADY drain a little. How will you filter them out? And by the way, why doesn't it say anything about the 2H moment? "Analysis of results and criterion selection of sets"? Looking at the balance/equity curve? Hmmm. Long.
By taking a 6-month run out of 8 months, it is possible to select sets that give stable growth with the right characteristics. The next OOS run for 2 months will sift out of them, those that do not give the growth of desired characteristics in these 2 months (periods taken from the scheme, for different strategies, the amount of time needed differently). Accordingly, I exaggerate of course, but 10 consecutive going OOS will give us check of needed characteristics of quality of growth at 10 periods, which will bring us closer to the searched goal - search and confirmation of stable system at all discrete periods, instead of at huge time interval.
1) Not about quality yet, but about the search time advantage - genius.
You suggest stopping 10 times, analysing sets 10 times, etc. etc.
At the end of the road you will already forget where you were going! ))
.......................................
2) All you have described "charms" I just as (even simpler) can easily see at 2H while not making 10!!! stops on this path. Plus some more possibilities appear. )
............
So your answer -- NO credit.
If "test runs" is replaced by "test run" then both schemes are identical:
but it is possible to keep the "runs" only they should be an order of magnitude smaller than the optimisation runs. The fewer the better (less chance of fitting). Better still 1 run. That is why there is a difference in the number of OOS runs. In the second scheme it is only one. This is better from the viewpoint of fit, but somewhat worse from the viewpoint of only one set being checked (FN). If the difference between the sets is only a Mach period, then of course there's no reason to test several sets with OOS. But if the difference is more global - for example the way the position is supported, then it might make sense. Then you can take the extra risk of adjustments (compared to a single OOS run). Some options can be used in such a way that with their different values there will be fundamentally different systems and it makes sense to run them separately on OOS, rather than choosing one FN
P.S. is the "G-point" where I thought it would be? :) Then the scheme takes on a new meaning)))
1) Vyacheslav, as usual, you can't argue with you. I totally agree.
Only -- all this is seen from point 2H. You must agree that from the present one can see the past clearly, though in different ways ;-))
........................................
2) ))) About point G -- a little later, when there is an answer to the question posed....
lasso:
All the "delights" you describe I can just as easily (even more easily) see at point 2H without having to make 10!!.stops along the way. Plus some more possibilities appear. )
Why don't we talk about quality? And the timing depends on how you in your 2H point estimate the desired sample. Looking at all balance curves? That's longer than criterion filtering. You're prying the wonders of approaches from others, while you're talking in riddles... Instead of wasting a lot of time on your scheme, you could describe failure of "traditional approach" (why is it traditional, by the way?) and all the charms that you can easily see in your approach. That is to say, as of.
lasso:
HOWEVER! (as V.V. writes) At first, after reading your post, I wagged my finger at my temple, then I realized that we are talking about different periods.
My optimization period is one and indivisible.
-- select all optimization passes, in which some final segment (e.g. 1/5 in fig. marked with a grid) meets certain values of PF, FS or other parameters.
How are you going to filter them out FF and others at section 1/5? Are they counted separately?
>>At the end of the road you'll already forget where you were going! ))
Debatable. I have written, that, it will approach us to the searched purpose - search and acknowledgement of stable system on all discrete sites, instead of on an indivisible piece of optimization as at you.
Yes, I wanted to ask the same question: why do you think, Vitaly, that what you have described is a traditional scheme? Can you give me a link?
If not, then "traditional" may be conditionally called what is described by Pardo. But it's more serious there, by the way.
We talk about wasps-schmos, samples-memples, but no one said a word, and probably no one even thought, but do such approaches (division into Sample, OOS, etc.) apply to all TCs without exception? If not applicable to all, which ones are not applicable? I posed a question to Reshetov, leading up to this topic, but he did not see fit to answer.
Let's try to figure it out.
In terms of time spent in the market of each single trade (each single trade, because TS may conduct several parallel trades simultaneously) TSs may be divided into two types:
1. TS with unknown time of each trade. This type includes all TS, in which it is not known in advance when there will be the next entry signal into the trade (or whether it will be at all), and in some systems it is not known in advance when there will be an exit signal.
2. TS with the maximum trading time known in advance. This type includes systems in which there is a signal to enter strictly periodically, for example on every bar or on a certain day of the week at a certain time. There is always a signal for the exit, because the maximum lifetime of each trade is predefined. The condition of exiting a trade can be the end of time or reaching of stops.
What thoughts arise if we divide the TS into such types? What follows from this, within the framework of the topic? I will listen to opinions, and then I will express myself.
Is OOS always needed if forward testing has shown good results?
That is, we select the best in-sample optimum and go to battle.
Is OOS always needed if forward testing has shown good results?
That is, we select the best in-sample optimum and go to battle.
Yes, I wanted to ask the same question: why do you think, Vitaly, that what you have described is a traditional scheme? Can you give me a link?
If not, then "traditional" may be conditionally called what is described in Pardo. But it's more serious there, by the way.
1) This is just a picked up term to separate concepts. Picked up here in the second post from the bottom.
2) I am ready to argue, discuss, learn from Pardo, Reshetov, etc., from anyone.
But to do this, the concepts need to be clearly defined (introduce a standard, our ISO if you like...).
Maybe create a branch like "Glossary. Definitions of basic concepts in this forum."? Discuss first, then put the approved definitions in the top.
.........................
Something I haven't come across a discussion about the profit factor. Only its usefulness, its application, etc. are discussed. Why?
Because there is a generally accepted definition.
Everyone uses the phrase "OOS" as they please. I've even mentioned "ten consecutive OOC's".
Plus the confusion with the place on the timeline: is it the real future, or the virtual(section of history)
And everyone is right.
...........................
Here joo has set an interesting task.
But there will be no use in discussions as long as everyone has their "waspshmos".
imho.