Where is the line between fitting and actual patterns? - page 4
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
The answer is wrong. Just when training NS I take the period of OOS less than the period of the training sample. Because BP is non-stationary and if you do the opposite, you just get a short sample fit and a very questionable result on the OOS.
And I didn't say that the OOS should be greater than or equal to the training sample.
I'm saying that as the OOS size decreases, i.e. as the optimization relevance increases, the representativeness of the OOS itself also decreases. That is, very quickly you reach that saddle, which you rightly mentioned, there is a very unpleasant effect of optimization with too small OOS, in which TC is optimized on OOS, but not on training sample - "learning in reverse".
As always, the golden mean is somewhere in the middle. :) And this middle ground for each particular TS is in different places.
In short, there is no and can not be a clear recommendation on what size OOS should be. You can only rely on your gut and experience.
fitting of the second instance...
;)
fitting of the second instance...
;)
:) It's a confidence booster, though!
The trouble with this one is it keeps getting higher and higher...
Not age-appropriate.
;)
Where is the line between fitting and real patterns?
Looking at the market we see that possibly existing patterns cannot be parametrically constant. Every system has a level of fit and a level of regularity of one or more events.
And the preponderance towards the second level is responsible for the rationality of the trading idea itself.
Thinking abstractly. The thoughts of others would be interesting.
It depends on many factors, and the main dependence is on the system itself. For example:
1.Large sample - well that's always a good thing, if at 6000 in a row it works consistently - why not work some more?
2. The correlation of some expected characteristics - for example, the influence of an event on the market is expected and confirmed by history - then you can take into account a not very large sample, say, about 100 events, or even less.
3. Matching of parameters to some expected parameters. In principle the same as in item 2, but from the other side - for example for trend systems the %% of successful transactions and the ratio of average profit to the loss are approximately clear.
And so on.
Most importantly, there is no 100% working method. Although, from an engineering point of view, there is one - it's called "diversification". :)
Same place - where the pendulum is...
:o)... The expected laugh... But really, the statistics are cooler than a lot of other people. It's just the stereotypes imposed by the system that prevent many from getting out of the circle they walk in. It's a series of TA textbooks, where everything starts with moving averages... The question is, why bother studying them if the whole direction with averaging data is rubbish. They show the present at best. I'm talking about ALL indicators embedded in MT :o). Except maybe a zigzag, which is like milk. And where to go? As we say in Deribasovskaya... That's what we do :o).
.... Namely 5/95% not for the better....
Tell us please, where did you get these statistics?
And this is a compilation.I taught for two years at a TA exchange academy... Roughly 60 weeks for 10 - 15 people - about 700 people, of which in a couple of years I see only 20 people, and it does not mean that all of them earn. I am the only one who earns :o))
:o)...expected laugh...
And this is a collection of stuff.I taught at a TA stock academy for two years... Roughly 60 weeks of 10 - 15 people - about 700 people, of which after a couple of years I see only 20 people.And that does not mean that all of them earn. I am the only one who earns :o))
Nibora! You?
;)
And this is a compilation. I taught at a TA academy for two years. Roughly 60 weeks for 10 - 15 people - about 700 people, of which in a couple of years I see only 20 people. And it does not mean that all of them earn. I am the only one who earns :o))
The point is that such statements without specifying the time period and how they were obtained are completely meaningless. even for teachers.
And the real statistics for example rann posted. But it is also not on clients, but on accounts.
And I didn't say that the OOS should be larger or equal to the training sample.
I'm saying that with decreasing the size of OOS, i.e. with increasing the relevance of optimization, the representativeness of OOS itself decreases. That is, very quickly you reach that very saddle you rightly mentioned, there is a very unpleasant effect of optimization when OOS is too small, when TC is optimized on OOS, but not on the training sample - "reverse learning".
As always, the golden mean is somewhere in the middle. :) And this middle ground for each particular TS is in different places.
In short, there is no and can not be a clear recommendation about how big should be the OOS. You can only rely on your gut and experience.
There's no need for any flair here. Sample period and OOS at using of neural network packets are selected empirically once for specific inputs and further on they are not changed. I.e. if inputs of NS are adequate, then everything else is a matter of technique, but not intuition.
As for the MT tester, everything is more complicated because, as already mentioned, there is no way to separate flies from cutlets, i.e. optimized sample from forward one and it's almost impossible to catch the moment when optimization transforms into fitting. To be exact it is possible to manually interrupt optimization and run forwards gradually increasing number of passes to catch the moment but taking into account that optimization time can be quite long and forwards you need to change date each time, then interest in such approach goes way down the peg.
Nibora! You?
;)
Didn't get it...
I taught for two years in a TA academy. Roughly 60 weeks at 10 - 15 people - about 700 people, of which a couple of years I see only people 20. I am the only one who earns :o))
The point is that such statements without specifying the time period and how they were obtained are completely meaningless. even for teachers.
And the real statistics for example rann posted. But it is also not about clients, but about accounts.
I know the real statistics. And you also know it, if you do something in the market. Especially since the period and method are clearly written at the top.