Subsystem "Asset Management" - page 2

 
thecore, efficiency is easy to check-put on the minutes. In any case the news will deflect the price from the statistics. Looks nice though. What is it called?
 
sayfuji >>:
thecore, эффективность легко проверить-поставить на минутки. В любом случае новости отклонят цену от статистики. Хотя смотрится приятно. Как называется?

The statistics are taken over a very long period, e.g. 5000 bars, so the prediction also takes into account movements on possible news.

Statistically the angle of slope of the MA is estimated. That is, I refuse to predict price movement. I predict the movement of the MA and then

and then I search for the interval of price movement relative to the predicted MA.

And it is not the precise angle of slope of the MA that is analyzed, but some "cloud" of closely spaced slope angles.

The "cloud" width is calculated dynamically and is based on the statistical sample size (number of elements found).

It has not been named yet. It is an indicator I am still working on.

 
An interesting undertaking. Difficult, but the result should be worth it. I only have a similar type of task in my plans so far. In this kind of thing (indicator prediction), neuro-brains (networks I mean) are very good to help.
 
sayfuji >> :
An interesting undertaking. Difficult, but the result should be worth it. A similar type of task is only in my plans so far. For such things (indicator prediction) neurobrains are very good help.

Neural networks (in this context) are a very slow and very crude solution to a problem with many unknowns.

Try sometime to teach the network on a history of 5000 bars, or better 20.000.

In this indicator even taking into account the adaptive fuzzy logic everything happens in Real Time.

Besides at the points of extremum where MA[1]~MA[2] any predictions for one indicator are meaningless.

What is required is the involvement of slower trend-tracking MAs (essentially older periods).

But what is useful is the cloud of nearest expected values.

Useful when you are about to close a trade and you are not sure if you should wait some more or

close it right now.

In addition, there is no binding binding to any particular indicator.

Rather, it is a statistical research methodology for the purpose of predicting the near future.

Or if you like near-term trends.

 

Yes, I agree with you about learning on a large number of bars. However, predicting MAs over large periods does not accurately predict price. The longer the period of the MA, the more the price deviation from the MA is tolerated. Take yesterday's Friday on the major currency pairs. The price deviation from the moving average was about one hundred points. 1 or 2 lots is bearable. 10 is too much.

In defense of neural networks I would say that there are special software products that provide results of data processing in real time. You can connect them to Mt4 using API.

Theoretically it is possible to combine several neural networks in one TS. It's just difficult and long. It's easier in this sense with fuzzy

 
thecore >> :

Neural networks (in this context) are a very slow and very crude solution to a problem with many unknowns.

Try sometime to teach the network on a history of 5000 bars, or better 20.000.

In this indicator even taking into account the adaptive fuzzy logic everything happens in Real Time.

In addition at the points of extremum where MA[1]~MA[2] any predictions for one indicator lose their meaning.

What is required is the involvement of slower trend-following MAs (essentially older periods).

But what is useful is the cloud of nearest expected values.

Useful when you are about to close a trade and you are not sure if you should wait some more or

close it right now.

In addition, there is no binding binding to any particular indicator.

This is more of a statistical research methodology with the aim of predicting the near future

Or if you want near term trends.

Learning a network for 5000 or 20,000 bars is like measuring the average temperature in a hospital. Neural networks have other disadvantages, you haven't mentioned them here.

 

Учить сеть на 5000 или 20000 баров все равно что измерять среднюю температуру по больнице

Choomazik, when using SC quotes, reference to it is mandatory)))

 
grasn писал(а) >>

to Prival

Prival, you must have forgotten your old opponent :o) Oh, well, I'll make it through diplomacy:

Your attentive look should not have escaped the fact that I'm trying to master Markov chains and linear programming for slightly different purposes, namely asset management, i.e. search and selection of optimal trading decisions, rather than forecasting itself. What I've proposed is a kind of study of the theory within the framework of optional development. And I define pivot points in a completely different way, as I wrote above and demonstrated.

And as for "everything else will fit" - you are extremely mistaken. Just refer to the wisdom of the ancient Chinese and my observations :o): You will never find a black cat in a dark room, especially at times when it is not there. Rely on me for this expert judgement completely - I have a black house cat and believe me, I know what I'm talking about. :о)))

Well, to answer an "old" opponent.

There are a lot of beautiful words. But there is no essence behind them, no depth of understanding of the processes involved in these mathematical apparatuses.

Цепи Маркова (https://ru.wikipedia.org/wiki/%D0%A6%D0%B5%D0%BF%D1%8C_%D0%9C%D0%B0%D1%80%D0%BA%D0%BE%D0%B2%D0%B0) - The main postulate is that it is not important what happened yesterday, it is important what is happening now. Let me explain, let us say the exchange rate of one currency is 1.2345, the other 2.3451, etc. Let us multiply it by the transition matrix, i.e. just by a number. Let's say in our example 1.2345*0.99991119999= what is there.

It is important to be able to calculate this number 0.99991119999 and it depends on time, it is one thing to forecast what rate will be in a minute and another thing what it will be in one day. The forecast is possible only if there is a movement pattern.

Теперь про управление(https://ru.wikipedia.org/wiki/%D0%A2%D0%B5%D0%BE%D1%80%D0%B8%D1%8F_%D1%83%D0%BF%D1%80%D0%B0%D0%B2%D0%BB%D0%B5%D0%BD%D0%B8%D1%8F)-

What matters is the destination, where to manage, where to move. And what to manage with 1 ruble or 1 million dollars is of secondary importance.

Here's an example, let's say that the rate will be as shown in the figure, and it's 100%, ironclad, etc. And we are at point 1. We can buy at point 1 and sell at point 2 or 3. Optimal? No. Optimally (that means best, there is no better !!!) sell at point 1 then buy at point 2 and close the trade at point 3. We do not know what's next, we have no forecast, so we cannot do anything.

Z.U.

That's why we have to base our operations on forecasts and the accuracy of these forecasts. If it is accurate and correct, I will go in with all my assets, I will go in up to my tomatoes )). But mathematics is always applicable, you just need to know where to apply it.) And there are a lot of beautiful words and theories....

It's hard to look for a black cat in a dark room, especially when it's not there.

 

to Aleku

I will write to you when I think about it.

to thecore

Это статистическое предсказание на N баров вперед движения MA (плавные штрих-пунктирные линии справа от последнего бара) и статистическая величина границ, в которых наиболее вероятно будет двигаться цена (вертикальные штрих-пунктирные линии возле каждого предсказанного бара для быстрой MA)

Thanks for the insight. I don't think I would be much mistaken to say that probably everyone, or almost everyone, has MA predictions. I'm glad that you are so far advanced in your research. It's not clear what exactly the phrase "... statistical prediction..." means in this context. But anyway, if you want to - you tell me, but not - I'm not offended, commercial confidentiality is sacred to me and I understand you perfectly.

I will tell you about my researches in this field, which I periodically return to. I even developed a secret zigzag for one of my models. Now the "secrecy" about this project is mostly gone :o). I have developed it myself, but taking into account its simplicity I am not claiming the copyright, it is quite possible, that someone has done the same thing. Its construction is as follows:

(1) For a fixed sliding window of the original process B(n), the MA(n) and the standard deviation SKO(n) are calculated. The standard deviation plotted upwards k* SKO(n) and downwards -k* SKO(n) from MA(t) in its sense defines the boundaries of, shall we say, the "main process", i.e. some zone where more samples are located. With small sliding window lengths, the frequencies of the B(n)-MA(n) difference magnitudes have an approximately normal distribution, a fact which ensures the "legacy" of some of its laws.

(2) The upper and lower bounds of k* SKO(n) uniquely define consecutive series of samples lying above and below this zone.

(3) In each zone, depending on the type of series (is it above or below SKO), there was a corresponding minimum or maximum and this zigzag was obtained:

This zigzag has an interesting quality, because it is in fact completely determined by statistical parameters of the initial series, and these parameters can be used for the forecast itself as well. The model was very simple:

Step 1: The described zigzag is built

Step 2: The MA forecast is performed. I made forecasts using different methods, including AR, ARIMA, and tried more exotic methods. Of course, I haven't forgotten about the set of functions calculated by ANC, i.e. the sum of the form C(1)*F1(x)+ C(2)*F2(x)+ ...+C(n)*Fn(x) (quite an interesting direction)

Step 3: Knowing properties of MA (for example, phase lag), statistical features of the zigzag, extreme points of the forecast zigzag and, finally, the formula of MA calculation itself, we could calculate the future zigzag precisely enough which would provide "statistical justice" of the forecast MA. And to make an estimate of the size of the bars within the forecast area is quite simple.

By the way, sayfuji, minutes, hours, days, etc. are essential for these methods, as the amount of variance in the time series (process) strongly influences the identification itself. And this very identification is a global problem, and its solution is really a trade secret and the essence of the secret is that it is almost impossible to "guess" model parameters all the time, i.e. statistically steadily. :о)

Neural networks (in this context) are a very slow and very crude solution to a problem with many unknowns...

In general - I agree. And this solution often creates an illusion of the very possibility of the solution.

to Prival

A lot of beautiful words

Thank you for praising such a modest contribution to world literature. :о) For my part I was a bit surprised with "academic" style of your post. I searched for myself and found explanation. It's very simple and it's in the most prominent place - it's numbers, just below your avatar. Such a high activity on the forum of an active teacher (I hope everything remains the same), combined with other, not less important concerns, except of course conquering Forex - does not leave you much time for thoughtful reading of what is being asked and what is being written about.

I think it's important to explain in more detail. In most cases just reading is not enough, especially when you don't have a serious understanding of the subject, you need to read thoughtfully.

But there's no essence behind them, no depth of understanding of the processes involved in these matrices

No, no, no - that will not do . You will crush your students with such a powerful intellect, and with me let us be simple - I am not one of them, I have a broad mind, my soul is wide - I can risk my reputation and just like that, just say something, by the way - I have already written something as an example of demonstration of capabilities. On what basis do you draw such conclusions? No, you dear fellow, let's go in order, thoroughly explain the way you think. If you need Freud's writings for that, I'm familiar with some of them.

And what do you mean by "these matapparats"? I haven't written anything at all about LPs yet, I was just about to - and I don't understand anything.

Цепи Маркова (https://ru.wikipedia.org/wiki/%D0%A6%D0%B5%D0%BF%D1%8C_%D0%9C%D0%B0%D1%80%D0%BA%D0%BE%D0%B2%D0%B0) - The basic tenet is that it doesn't matter what happened yesterday, it's what's important now.

And in general I'm glad you learned something about Markov chains.

Let's say the exchange rate of one currency is 1.2345, the other 2.3451, etc. Let's multiply it by the transition matrix, i.e. just by a number. Let's say in our example 1.2345*0.99991119999= what is there. It is important to be able to calculate this number 0.9999111999999 and it depends on time, this is a forecast one thing is what the rate will be in a minute and another thing is what it will be in one day. The forecast is possible only if there is a movement pattern.

You multiply the predicted rate (one number) by its probability and what do you get? Can you elaborate, just take your time, I'm writing it down.

Here's an example, let's say the rate will be as in the figure and it's 100%, ironclad, etc. And we are at point 1. We can buy at point 1 and sell at point 2 or 3. Optimal? No. Optimally (that means best, there is no better !!!) sell at point 1 then buy at point 2 and close the trade at point 3. After that we do not know, there is no prediction, so we cannot do anything...

Dear "optimizer", maybe you should yourself sit down and figure it out? Or is it the demonstration of the essence and depth of understanding of the processes?

 
grasn >> :

to thecore

Thanks for broadening the mind. I don't think I would be far wrong in saying that everyone, or almost everyone, is predicting MA. I'm glad you are so far advanced in your research. It's not clear what exactly the phrase "... statistical prediction ..." means in this context. Anyway, if you want to tell me, but if not, I won't be offended, commercial confidentiality is sacred to me and I understand you perfectly.

A classic MA prediction (as I understand it) is something like this:

- take MA[2] and MA[1] and use the data to calculate the difference of MA[2]-MA[1] or angle.

- then move further to the left and find the same angle on the history

- from this found point take as many bars FORWARD as we want

- write the average value of all found values into the array

- pass as many BACK bars of the history as we want, but it is desirable that the trends change several times during this time

- the result is an array of averaged prediction points

I have made some improvements to this method based on the following observations:

1. If you only track a certain slope angle, you get a very small sample,

e.g. for some slope angles it can be <0.1% of the number of bars.

2. MA prediction by itself is not very informative - it is desirable to have a cloud of probable values

of prices themselves.

Improvements are as follows:

- it is not the angle of slope of MA that is tracked, but a cloud of slope angles is used. The criterion of cloud width

is % of the number of bars. It means that the range of values of MA close to the initial one is widened

until the amount of collected data is greater than the specified % of the investigated number of bars.

- Having got the prediction point of the MA, the distances from MA to High and Low are read and then these values are

as in the case of MA are averaged. As the result we have some limits where the price moved when the slope angle of

MA was approximately the same as it is now.

Why do I think the work is not over?

- if we take timeframes larger than D1, everything turns out badly, because there is little historical data and

the market changed a lot

- If we take timeframes less than H4, everything is not good either, because we have to consider the activity of the currency.

The slope angle of MA during night time (at non-market time) cannot be predicted the same way as during the day time because of

obviously different volatility.

- Prediction of inflection points for one MA loses its meaning because depending on the trend direction they may be

are diametrically opposed.

- and most importantly. The statistical study I conducted earlier shows

that the collected data does not have a normal distribution, and therefore it is not correct to obtain an average

is not correct. The data shows that such averages are not one, as on a Gaussian curve, but three or more,

so it is not enough to get a 3-sigma to establish a confidence interval

probability interval. Alas it is not.

So there you have it colleagues.