How can I tell the difference between a FOREX chart and a PRNG? - page 25

New comment
 
serferrer:

And this is the real deal.


Read https://www.mql5.com/ru/forum/143224/page21#754529


So what? It's a common situation - the real one is being poured by the EA, which was adjusted in the tester.

Read the standard situation - the real one is being fitted by the EA in the tester.

 
serferrer:

The councillor is the same.


What was the lot size you were testing? 1 or 0.1?
 
No, the problem is that someone came up with the idea of testing by ticks on candles over 1 minute, hence all the problems - at least a review should read how the tester works
 
AlexEro:

https://en.wikipedia.org/wiki/Autocorrelation


And here is Karl himself about his brainchild:

That is, both correlation and autocorrelation and ANC for fitting and regression of any random variables can be used, but Karl and his friend Yule cannot give a complete guarantee for the case if they are not Gaussian distributed.

The accuracy of these methods is then simply unknown.



Great. And here's where we come in.

1. The formulas for calculating the ACF are completely the same.

2. There is no other formula, i.e. no matter how much we discuss calculate autocorrelation in another way (no one can).

3. But the third point is interesting "The accuracy of these methods is then simply unknown." Let me explain.

More than 50 years were beating about the solution of the Stratanovich equation, while Berg brought the formula, which showed that the exact solution can not be obtained (computing resources tend to infinity with breakneck progression), and the scientists left this equation behind. But the time came andGennady Petrovich Slukin showed the solution, just think of the beauty of the phrase, he found the solution"with sufficient accuracy for practice...".

Let me explain it with a more understandable example. You can try to forecast the movement of EURUSD accurate to 0.001 points and never find a solution. And it is possible to make accurate predictions with the accuracy of +-5 points and it will appear to be enough for earning.

The purpose of building ACF is to see and understand the process, which formula can describe it (a model). And with this model one can predict the rate with accuracy enough for practice. In the figure, the blue curve is a model which almost exactly repeats the ACF of a real process.

 
Prival:

So I understand that accuracy is the point and that by manipulating the accuracy, it is possible to judge whether or not what is there will last by the amount of accuracy.
 
Prival:


Excellent. And what have we come to.

(in Vitsin's voice)

Sorry, not "we", but "you".

Privalov, you'd better correct your ACF and explanations to it, you'd brush it up, figuratively speaking. Of course, even such an ACF in codebase is a hundredfold better than no ACF at all. But still, this is a new thing, people will want to understand. And to avoid repeated bloody cuts, like the one involving nihilist hrenfx

"zero sample correlation does not mean there is no linear (usually forget the word linear too) relationship in that sample."

https://www.mql5.com/ru/forum/128968

in which 5-6 knowledgeable mathematicians on this forum never understood each other - so for that you should put beacons: where is the theoretical auto- and simple correlation and where is the sample correlation, how do the sample sizes and auto-correlation plot relate, how are the multipliers normalized. Otherwise it turns out that your periodic pure sine function has a damped autocorrelation


https://www.mql5.com/ru/forum/128968/page15

The formula for autocorrelation mentioned in the wiki above is clear and suitable for pogramming, while yours is not yet.

This is not a criticism, it's for the sake of clarification.

Regarding everything else, there are several robust and non-parametric ways to calculate autocorrelation, but here we ("no, now it's you"(c)) are entering the thin ice of the connection between theorwave and DSP, and personally I don't want and can't go further.

"2. There is no other formula, i.e. no matter how much we discuss calculating autocorrelation differently (no one does). "

Privalov, be careful with negative statements. Firstly, negative statements themselves are difficult to prove - in mathematics, mathematical logic, philosophy, orthodox theology and even Moses and Aaron's Judaism.

Secondly, if someone claims that chocolate cakes do NOT revolve around Saturn, how can this be proven and verified?

Thirdly, how do you know there is no other formula? You don't have to answer, it's a rhetorical question.

"But the time came andSlukin Gennady Petrovich showed the solution, just think of the beauty of this phrase he found the solution"With sufficient accuracy for practice..." "

Well, he did SHOW that accuracy, in percentages. That is, you know in advance approximately how accurate his method will be. But the author himself and his friend Yul can't show any precision in the correlation formula. Because it will go anywhere - depending on the shape of the distribution. No one makes instruments with "technical accuracy", since all instruments (and all mathematical methods) have a pre-determined PERCENTIALITY (residual term, magnitude of smallness, etc.) to work with. This is what they do in both engineering and science.

I have shown and told you quite enough on the subject, so hereby I bid you farewell.

 

If we remember Gödel, before discussing ACF we have to go back to the initial conditions of the problem, which are not provable and are unformalizable within the methods used. These conditions are considerations about the quotient as such.

Cotier is fundamentally non-stationary process, i.e. a process with variable, not necessarily linear mo, and if we subtract this variable mo from the process, the resulting residue will be a variable variance with a lot of intricacies. But this is not enough. There are events in the market after which the found regularities (variable mo and variable variance) can change radically. And we can learn about it after getting the necessary amount of history. This leads to a big trouble - a pattern found in history must be extrapolated to the future very carefully.

DSP assumes that there is a signal in a random process which is noisy. And it is very important that the noise (as it seems to me) is always Gaussian. So everything written in books on correlation, regression and other analyses always applies to noise in DSP. But the methods from these books cannot be applied directly to the quotient due to non-stationarity of the quotient. By the way, it means that DSP methods cannot be applied to the market. But mathematics and equipment within DSP created and worked out on real and model data will always work in the future - the TV set works as it is.

ACF is like litmus paper. You calculate it, look at it, and decide what to do next. No more than that. Since we always remember that quotients are non-stationary, we must first make sure that the series is stationary and only then apply the entire correlation analysis tool.

For example.

When we take a program from kodobase or matcad and do an ANC fitting, we should always understand that the obtained coefficients are not always what they are calculated to be. You have to look at the error in calculating these coefficients, and if the ISC calculation program does not provide this information, it should not be used at all. And this is not all the information that is needed to make a decision on whether to use the derived coefficients. That is why even for the simplest calculations of ANC type one should use specialized packages (R, EVIEWS ....). Using specialized packages we will immediately see how rigidly the treatment of stationary and non-stationary series differs.

That's why I take the whole conversation about ACF to be purely theoretical and irrelevant to quotes.

But the topic of the topic is interesting. Above I even linked a book on the subject - how to distinguish deterministic trends from stochastic trends. From the mentioned book it follows that in the general case these trends cannot be distinguished, so it turns out that a quotient cannot be distinguished from a specially processed (e.g. artificially creating thick tails) PRNG.

 

faa1947:

That is why I take the whole conversation about ACF to be purely theoretical, which has nothing to do with the quote.

But the topic of this thread is interesting. Above I even linked a book on the subject - how to distinguish deterministic trends from stochastic trends. From the mentioned book it follows that in the general case these trends cannot be distinguished, so it turns out that a cotier cannot be distinguished from a specially processed (e.g. artificially creating thick tails) PRNG.

Mm-hmm. Some depressing conclusion comes out. To make it more upbeat, I'll repeat it for you: George Marzaglia made some remarks that put everything in its place, and show where the bridge is. (of course not directly, it takes a lot of thought, good knowledge of DSP and a hell of a lot of programming.) You can do it without Marsaglia, but it will be a much longer way.

Grandpa Marsaglia was sort of a nihilist, and as far as I understand he even had a conflict with NIST - the monster of American scientific bureaucracy, which stupidly parasitized his works, and (attention) it seems they have flaws in standard encryption-decryption-hashing methods there.

For dessert (or as an appetizer for the kino-picture above) here's a simplest, but high quality RNG by Marsaglia (though they should be initiated by another RNG):

// Another example has k = 257,  period about 2 ^ 8222.
// Uses a static array Q[256]  and an initial carry 'c',
// the Q array filled with 256 random  32 - bit integers
// in the calling program and an initial carry c < 809430660
// for the multiply - with - carry operation.
// It is very fast and seems to pass all tests.

static unsigned long Q[256], c = 362436;  /* choose random initial c<809430660
and */
/* 256 random 32-bit integers for Q[]
*/

unsigned long MWC256 (void)
{
        unsigned long long t, a = 809430660 LL;
        static unsigned char i = 255;
        t = a * Q[++i] + c;
        c = (t >> 32);
        return (Q[i] = t);
}

// Here is a complimentary-multiply-with-carry RNG
// with k=4097 and a near-record period, more than
// 10^33000 times as long as that of the Twister.
// (2^131104 vs. 2^19937)

static unsigned long Q[4096], c = 362436; /* choose random initial
c<809430660 and */
/* 4096
random 32-bit integers for Q[]       */
unsigned long CMWC4096 (void)
{
        unsigned long long t, a = 18782 LL;
        static unsigned long i = 4095;
        unsigned long x, r = 0xfffffffe;
        i = (i + 1) & 4095;
        t = a * Q[i] + c;
        c = (t >> 32);
        x = t + c;
        if (x < c)
        {
                x++;
                c++;
        }
        return (Q[i] = r - x);
}

It couldn't be easier. Marsaglia's grandfather was good at maths, theorizing and statistics.

 
faa1947:
AlexEro:

Mm-hmm. Some depressing conclusion comes out. To make it more cheerful, I'll repeat it for you: George Marsaglia made some remarks, that put everything in its place, and show where the bridge is. (of course not directly, it takes a lot of thought, good knowledge of DSP and a hell of a lot of programming.) You can do it without Marsaglia, but it will be a much longer way.

Grandpa Marsaglia was sort of a nihilist, and as far as I understand, he even had a conflict with NIST - the monster of American scientific bureaucracy, which stupidly parasitized his works, and (attention) it seems they have flaws in standard encryption-decryption-hashing methods there.

For dessert (or as an appetizer for the kino-picture above) here's another simplest, but high quality RNG by Marsaglia (though they should be initiated by another RNG):

It couldn't be easier. Marsaglia's grandfather was good at maths, theory and statistics.


And here's the topic of the topic of the thread has interest. Above I even linked a book on the subject - how to distinguish deterministic trends from stochastic trends. From the mentioned book it follows that in the general case these trends cannot be distinguished, so it turns out that a cotier cannot be distinguished from a specially processed (e.g. artificially creating thick tails) PRNG.



AlexEro:

Mm-hmm. That's a depressing conclusion.

what's the depression? You do not need to be able to identify PRNG data to make money on the real series
 
AlexEro:

For several pages I try to follow this verbiage - I can't understand anything.

Prival says that there is no other way to calculate autocorrelation, no matter how much one criticizes the existing one, and in response the reference to Moses and Aaron.

The thesis that in reality it is almost impossible to distinguish real quotes from gpsch, and in response the remark that George Marsaglia showed several bridges. And in addition another gpsch is given. why?

That is a lot of bukouffe, a lot! Please keep it short and to the point, otherwise people will get scared.

1...18192021222324252627282930
New comment