Where is the line between fitting and actual patterns? - page 17

 
Avals: For example the stock indices of prosperous countries have a constant tendency to grow.

There is no longer a non-stationarity here. There is a global pattern -

Avals : "let the profits grow and cut the losses"

)))

paukas : It depends on where you work. ;)

Sure )))) It depends on where, how, when and most importantly "for how much?" ))))

 
Reshetov:
Once again for the especially gifted: run it on OOS and by the results choose the best one. Another thing is that in MT this procedure is problematic even for 1000 tests. At least, you need a program that parses the optimization results and launches tests in the terminal through the command line, substituting parameters into *.set files, then parses the test results and puts them into files. Even after such automation it's all terribly slow and the sound should be turned off, otherwise the beeping will make your ears fade.

1) Thank you for the high praise.

........................................................................

2) Already implemented. I could not find any stable patterns in the selection methods.

A lot has changed since those posts, for the better)), but still, the issue is not closed.

.......................................................................

3) And most importantly. We all speak different languages here. We operate with notions, but everyone puts their own meaning into this notion.

Consequently, in this state of affairs it is in principle very difficult to agree on anything or convince an opponent.

An example?

The discussion is already on page 16 and I've just realized that OOS is "different".

HOWEVER! (As V.V. writes) At first, after reading your post, I wagged my finger at my temple, but then I understood that we are talking about different periods.


My optimization period is one and indivisible. ( joo and Avals have already written about the harm of such a division in financial markets.)

But at the same time, the post-processing analysis functionality, allows me to set a condition, for example,

-- I propose to choose all optimization passes where a certain final segment (e.g., 1/5 marked with a grid) meets certain values of SP, PDF or other parameters.

This approach seems to me more correct, and sort of includes and does not reject your variant.

 
LeoV:

There is no longer a non-stationarity here. There is a global pattern -

)))

Non-stationarity does not preclude the existence of regularities. Even almost constant ones :)

Non-stationarity is purely a mathematical concept - like the distribution changes with time, or its parameters

 
lasso:

My optimization period is one and indivisible. ( joo and Avals have already written about the harm of such a division in financial markets).

But at the same time the post-processing analysis functionality, allows me to set a condition, for example,

-- select all optimization passes in which some final segment (e.g. 1/5 in Fig. marked with a grid) satisfies certain values of PF, FS or other parameters.

This approach seems to me more correct, and as it includes and does not reject your variant.

When something seems, it is necessary to be baptized (c) Popular proverb

And in order not to seem and not to glitch, forward tests are applied. They do not guarantee, but they separate flies from cutlets.

PF and FS may be drawn in the tester whatever you want. But you can't put it on bread and put it in your pocket.

Two friends of mine did not manage to discover any gold nuggets this summer, although they searched all the mountains with a very expensive high-tech metal detector. They brought back two backpacks of pyrite, which they foolishly mistook for gold. If these fellows had studied physics, they would have been able to distinguish pyrite from gold by its density, because gold is heavier than lead with the same volume and they could hardly drag two knapsacks of nuggets.

But this does not mean that the business of prospecting is a waste of time. An experienced miner never takes a metal detector into the mountains - it is absolutely useless.

 
Avals:
for example the stock indices of prosperous countries have a constant upward trend. And the dumb rule of 'let profits rise and cut losses' is the use of non-stationarity with positive mo for longs. True that's also why dips happen much faster than periods of growth))) Profits are made up of some very unsteady trades :)


Been watching your posts with interest for a year or more). At this moment there is no one on this forum who has a better trading efficiency than you do ;)

 
storm:


I've been following your posts with interest for a year or more. At this moment there is no one on this forum who has a better trading efficiency than you do ;)


Thanks a lot)))) When you write you start to understand :)
 
Reshetov:

When something seems, you should be baptised (c) Folk proverb

And to keep it from seeming and glitching, ...............

Thank you. A very informative answer.

I have no more questions for you.

.......................................

Interesting, to hear from those who understand what it is all about.

 
lasso:

-- select all optimization passes in which some final segment (e.g. 1/5 in Fig. marked with a grid) satisfies certain values of PF, FS or other parameters.

This approach seems to me more correct and includes and does not reject your version.

.......................................

It is interesting to hear from those who understand what we are talking about.


I'm afraid I don't get it either, and the point of looking at PF, PV and other things on the training period? What specific values are we talking about?
 
I wonder if anyone checks the OOS for repeatability of BP with a learning period? Could it be that the OOS is exactly the same as .... shorter learning curve. Put it on real - bummer.
 
Figar0:

I'm afraid I don't understand it either, but what is the point of looking at PF, FS and so on during the training period? What particular values are we talking about?


The meaning is exactly the same as Reshetov's

Reshetov:
Run it on OOS and choose the best one based on the results.


But the negative effect that joo wrote about on page 3 disappears.

joo:

That's right. There's simply no better way to check how well optimised the TS is (not fit).

There remains one problem. The relevance of such optimization lags behind exactly at the value of OOS. Well, yes, one may say, all that remains is to optimize TC up to now.

Booga-ha! And there's no test OOS period for comparison.

One can't help but think that we must chose the OOS size as small as possible to increase the optimization relevance. But herein lies the problem - the smaller OOS, the greater the probability that the real-time TS will fail.

It is a stick with two ends, so to say. Or even three ends.


Honestly, I do not even know how to explain it better. I kind of made up a picture.

Ask specific questions and I will try to be specific in my answers.