Machine learning in trading: theory, models, practice and algo-trading - page 316

 
I would like to add to the above. In order to engage in MO you need to thoroughly know the subject area. In our case, trading. Many people just look at the katir on a non-stationary time series, forgetting that this is the market and it has certain rules and nuances. When you try to automate any process, you need to know the technology of the process to the smallest detail. I'm telling you as a mechanic with many years of experience.
 
20170413163153.mp4 -- Радикал-Видео
20170413163153.mp4 -- Радикал-Видео
  • radikal.ru
Радикал-Фото - сервис публикации изображений для форумов и блогов: Видео
 

Interesting video. What's the point?
 
Mihail Marchukajtes:

What's the point?

In clustering.


Machine learning can very roughly be called clustering. There is a certain hyperspace of predictors, and the need to divide it into several subspaces, where the point belonging to a particular subspace (class) means for forex the action of trading - in the case of 3 classes - "buy", "sell", "exit".
The video clearly demonstrates teacherless learning on two predictors (X and Y), how automatic clustering changes the boundaries of subspaces depending on the amount of data. With respect to forex, it metaphorically shows the duration of the backtest, and how its (duration) change affects the clustering result. A model trained on one week of data sees and knows much less than a model trained on two, three, etc. weeks.

The second part of the video shows how the expert assesses the clustering results and makes adjustments to the model. The expert sees that the obtained 3 classes are not enough, and by eye you can see at least 6 classes, then according to his experience, he corrects the model parameters so that it is able to accurately perceive these 6 classes.
This is the general idea. In my opinion, this step is impossible in Forex, because there are dozens of predictors and it is difficult to perceive more than three dimensions. As I understand it, this step includes the experience of automatic rather than manual adjustments, when changes in model parameters are accompanied by simulated trading, where the criterion for well-chosen parameters is good trading.

 
Dr.Trader:

Machine learning can be very roughly called clustering. There is a certain hyperspace of predictors, and the need to divide it into several subspaces....

It's not all about trading, not about risk. Trading is primarily psychology, not mathematics, you're digging in the wrong direction dear, study women better this will help trade more than teaching computers.
 

And who checks the quality of the code of these networks?

https://www.mql5.com/ru/forum/190948

Alglib MLP (нейронная сеть) портирование неправильно?
Alglib MLP (нейронная сеть) портирование неправильно?
  • www.mql5.com
Библиотека Alglib уже давно является частью MQL5. Нейронная сеть из этой библиотеки пока единственная из официально доступных...
 
pantural:

Yes, it's either this or that, but not at the same time.

You need at least three people - manager, trader and software developer, a cohesive team, the manager must practice team-building, trader to average, and developer to use design peters everywhere and even if one practices pair programming, then it will be good.

An archivist is also needed. And on the side...
 

After my experiments with eurusd with different patterns it seems to me that the price is seriously regulated in order to bring more profit to brokers and banks.

A typical situation - we teach the model on a couple of weeks of data, and then on new data we obtain only 50% of profitable trades (random, in fact) and slow spread loss.
But if we experiment with models, try to look for patterns, then we get a bit different situation - some patterns are profitable for some weeks, then suddenly fall to 50% of success, i.e. random. But after a month or two they work again, but you have to trade against their prediction. And after a couple of weeks their prediction is again at 50% random. And somewhere in the future they will be profitable again. And so on.

From all this I have the following conclusion - the banks set the prices according to their programs, algorithms. These programs they periodically change, use different combinations of them, change the prices in the opposite direction than their program offers, etc., all just to create a new situation on the market. Otherwise their algorithms would have been figured out and used against them.
And at the same time, people with thehanalysis or machine learning are trying to look for patterns, which have existed for a long time. And the patterns change at the snap of someone's finger or contradict themselves, no wonder trading is so difficult.

The working models must take it all into account - and the fact that the patterns work only in certain time segments, sometimes in the opposite direction, and be able to understand from the current situation which set of patterns to use.

All is futile?

 

I can see that here, too, an understanding is gradually coming -- albeit slowly and with difficulty -- that

The market is a controlled dynamic system.


But realizing this fact makes us reconsider the approaches to its consideration and description.

Then comes the understanding of the fact that statistical methods are not adequate methods by which an adequate model of the market can be built, and are suitable only for "talking" about tails. Thinner tails or thicker tails.

;)

 
Dr.Trader:

After my experiments with eurusd with different patterns it seems to me that the price is seriously regulated in order to bring more profit to brokers and banks.

A typical situation - we teach the model on a couple of weeks of data, and then on new data we obtain only 50% of profitable trades (random, in fact) and slow spread loss.
But if we experiment with models, try to look for patterns, then we get a bit different situation - some patterns are profitable for some weeks, then suddenly fall to 50% of success, i.e. random. But after a month or two they work again, but you have to trade against their prediction. And after a couple of weeks their prediction is again at 50% random. And somewhere in the future they will be profitable again. And so on.

From all this I have the following conclusion - the banks set the prices according to their programs, algorithms. These programs they periodically change, use different combinations of them, change the prices in the opposite direction than their program offers, etc., all just to create a new situation on the market. Otherwise their algorithms would have been figured out and used against them.
And at the same time, people with thehanalysis or machine learning are trying to look for patterns, which have existed for a long time. And the patterns change at the snap of someone's finger or contradict themselves, no wonder trading is so difficult.

The working models must take it all into account - and the fact that the patterns work only in certain time segments, sometimes in the opposite direction, and be able to understand from the current situation which set of patterns to use.

All is futile?


I didn't expect such thoughts from a reasonable person :)

Occam's Razor: "Don't make unnecessary entities".