Machine learning in trading: theory, models, practice and algo-trading - page 990

 
forexman77:

If I understood correctly, instead of a price series to use a series of fractals (I mean an indicator that builds on fractals)?

What are multifractals?

I do not know how to automate it

If you do not know how to use multifractals, then you should use them by eye. There is a book by Almazov, it is lying on the net somewhere

http://www.al24.ru/pdf_kniga_15078.html

I killed a lot of time with a friend in order to automate the search for fractals. I got bad results using correlation (I looked for the coincidence of the chart with the f-axis and made a forecast where the f-axis went further). I have lots of bad predictions because of correlation (correlation is high in trend areas but accuracy is poor in fact)

The multifractal example - self-similar fractal cycles that follow each other. In forex, this also occurs from time to time.

The last successful example is bitcoin. Of course, all is not so beautiful in real life, but the analogy is complete.


 
Imho, if you do not have your own strategy, no MO will help. At least learn how to do it.)
IO will help, but it won't work for you.
 

Yeah...

Well, okay - while I, knocked to the very bottom of the Hilbert space by an extremely unsuccessful EURUSD trade, am slowly climbing out of it, there is silence here as well...

However!

Where are the heroes of machine learning? Where are Max, Teacher, Koldun, Aliosha, Fa, Doc and their friends?

Sleeping... Okay, I won't wake them up.

 
Maxim Dmitrievsky:

By the way, Mandelbrot's legacy is economophysics.

They have their own formulas and methods there, but I haven't studied them. It is postulated as a replacement for the obsolete theory of the efficient market.

The roots of economophysics lie in the works of the classics.Benoit Mandelbrot discovered in 1965 that the dynamics of financial series (price fluctuations in the stock exchange) are absolutely the same in small and large time scales: it's almost impossible to determine from the graph of such series, whether it represents price fluctuations during an hour, a day or a month. Mandelbrot called this propertyself-similarity, and the objects possessing it -fractals. Investigations of processes with such properties are very actively conducted in physics and analysis methods developed there often (unfortunately, not always) help to notice anomalies in behavior of financial series - the harbingers of sharp price collapses or rallies. The French mathematicianLouis Bachelier at the very beginning of the 20th century in his"Theory of Speculation" tried to describe the dynamics of financial series by analogy with Brownian motion - chaotic motion of molecules in a liquid or gas. Modern models which generalize such approach generate fractal processes which are statistically very similar to real financial series. Many of these models are based on the theory of chaotic dynamical systems - equations generating complexdynamics, sometimes almost indistinguishable from a random process - created in the 1970s-1990s. Modern econophysics also uses other powerful toolsof theoretical physics- for example, thecontinuum integral, the most important toolof quantum mechanics andquantum field theory. But perhaps the most fashionable direction today isevolutionary games, directly imitating the activities of countlessinvestors following certain preferences and principles.


This is the only right direction when creating your own TS. Its logical conclusion can and should be the NS, which has as its target function the similarity of entropy or nonentropy as a measure of chaos/self-organization.

The only thing this theory is not suitable for is event times (the arrival of tick quotes as an analogue of moments of collision of particles in Brownian motion).

If in Brownian motion the time between collisions deltaT -->0 and a simple discrete event stream is formed with p=0.99 (probability of next collision) and q=0.01 (probability of no collision), then going to the Erlang flow of order 30, we obtain a Wiener process with drift with uniform readout time = 30 sec, which is what Perrin used in his experiments.

This is not the case in Forex. The probabilities of events (arrival/absence of a new quote) are on average = 0.5, and when passing to Erlang flows of order 30 and higher we have the so called Laplace motion.

I'm going to study it and make you happy, my friends.

Don't worry, Uncle Sasha will do it.

 
Alexander_K2:

This is the only correct direction in creating a TC of its own. Its logical conclusion can and should be the NS, which has as its target function the similarity of entropy or nonentropy as a measure of chaos/self-organization.

The only thing this theory is not suitable for is event times (the arrival of tick quotes as an analogue of moments of collision of particles in Brownian motion).

If in Brownian motion the time between collisions deltaT -->0 and a simple discrete event stream is formed with p=0.99 (probability of next collision) and q=0.01 (probability of no collision), then going to the Erlang flow of order 30, we obtain a Wiener process with drift with uniform readout time = 30 sec, which is what Perrin used in his experiments.

This is not the case in Forex. The probabilities of events (arrival/absence of a new quote) are on average = 0.5, and when passing to Erlang flows of order 30 and higher we have the so called Laplace motion.

I'm going to study it and make you happy, my friends.

Don't worry - Uncle Sasha will do it.

Of course, I expect a converted BP, and the neuronet will do the rest :)

 
Maxim Dmitrievsky:

Of course, I'm waiting for your transformed BP, and the neural network will do the rest :)

I remember my promise. Only it won't be a logarithmic stream of events (quotes), on which I'm burning, but the Erlang stream of a certain order, so that it would be Laplace motion for sure. I wonder how the NS will work out this process... I'm sitting on order 18 this week.

Actually, I thought this problem was easier to solve. Alas, we'll have to wait a bit...

 
Alexander_K2:

I remember my promise. Only it won't be a logarithmic stream of events (quotes), which I've burned on, but an Erlang stream of a certain order, so that it would be Laplace motion for sure. I wonder how the NS will work out this process... I'm sitting on order 18 this week.

Actually, I thought this task was easier to solve. Alas - we have to wait a little longer...

One more nuance - for HP superprecision of BP transforms is not necessary, you can even throw away not very good transforms. At the expense of statistical advantage can be taken out, the main thing that at least some regularities would be

 
Maxim Dmitrievsky:

Another nuance - for NS super-precision BP is not needed, you can even discount not very good conversion. At the expense of the statistical advantage can be taken out, the main thing that at least some regularities would be

Okay. On Saturday I'll post the 18th order - let's see.

Aleshenka said something about exponential thinning of BP for a reason. Oh, not for nothing... He's smart beyond his years. So we'll go on the beaten path - to take away the coveted Grail from Koldun and Alexey.

 
Alexander_K2:

Okay. On Saturday, I will post the 18th order, and we will see.

There's a reason why Alyoshenka said something about exponential thinning of BP. Oh, not for nothing... He's smart beyond his years. So we'll go down the well-trodden path - to take away the coveted Grail from Koldun with Alyosha.

The last month or two the forum has become a kind of madhouse. Adequate, on the fingers you can count.
Too bad, it was a good forum...
 
Yuriy Asaulenko:
The last month or two the forum has become a kind of madhouse. Adequate, on the fingers can be counted.
Too bad, it wasn't a bad forum...

What category would you put yourself in?