Machine learning in trading: theory, models, practice and algo-trading - page 1211

 
http://www.machinelearning.ru/wiki/index.php?title=%D0%9C%D0%BE
Here's a course of lectures by Voronov...but I wonder if he has a full-fledged textbook...
 
Martin Cheguevara:
My trading robot is always on the upside. I think it always gets better, I don't know when it's possible to trade with one order at a time. So I need to know when probably the robot will work worse than usual... because the risks are limited, my profit is purely a question of time... and sometimes you have to wait for a week... And in a week if there were no such drawdowns, you could get a lot more...
In fact, there is a dependence of e.g. a flat-trend state of the market on the trading results (e.g. the cache for the last deals of the day or for the last 30 deals)... It's not a problem to determine the trend-flat state... there is a mechanism that does it.
The problem is in a neural network...
I used to get links on lectures... But lectures it's clear... But is there a tutorial or self-study on the "do it yourself" principle?)
Why did I take the cattle? Because the "loosening" of trading in plus or minus means the growth of risks in any case...

Martin and drawdown are two inseparable friends

And no matter how you twist the trend / flat, it will always be that way.

PS

Can you give me a link to the book, please?

 
Martin Cheguevara:
Esteemed forum members, could you please tell me, because it is too lazy to read 1200 pages, has anyone here tried to implement machine learning based on the results of trading on the closed morders of the Expert Advisor?

https://www.mql5.com/ru/code/22710

BestInterval
BestInterval
  • www.mql5.com
Рыночные закономерности зависят от интервалов внутри суток или недели. По этой причине разумно ограничивать торговлю ТС по времени. Например, есть скальперские ТС, хорошо торгующие кроссы на азиатско-американской торговой сессии. Или же практикуется выключение ТС в период высокой волатильности. Соответственно, встает задача, как найти наилучшее...
 
Renat Akhtyamov:

Martin and drawdown are two inseparable friends

And no matter how you twist the trend / flat, it will always be so.

PS

Do you want to give me the link to the book?

I'm asking for a link myself:) I can only throw a link to a course of lectures...
 
Renat Akhtyamov:

Martin and drawdown are two inseparable friends

And no matter how you twist the trend / flat, it will always be so.

PS

Can you please send me the link to the book?

Well... if you read carefully, then you can see that I meant that I only work one order (opening and closing) and only with a minimum lot. How is it possible that this is (in your opinion) martin?) This is not a martin this is not a trend-follower, it's all together. Only with one order at a time and so that trades a trend and a flat at the same time.
 
If you trade only by trend or by flat - you will definitely lose, at least on major ones, I've already figured it out. Because the number of pullbacks in the history for the last 15 years and the number of trends for the same 15 years is the same, and the mutual exclusion of trends by a flat and vice versa is almost ideal.
Unfortunately, only a few can realize this fact, as I understand it... Of course, it's easier to blindly believe in luck...
 
Aleksey Vyazmikin:

The preliminary results (since I haven't made all predictors yet) on creating a model that determines profitable models (1) were not so bad, here is the breakdown by y - profit on the independent sample, and by x - 1 - TP+FP, and 0 - TN+FN.

The target was profit of 2000, well it hasn't been achieved so far, but only 3 models out of 960 have entered the loss area, which is not a bad achievement.

The table of conjugacy



The average unclassified financial result is 1318.83, after classification 1 - 2221.04 and 0 - 1188.66, so the average financial result of the models has increased by 68%, which is not bad.

However, it is not clear yet whether this model can work with models built on other data.

Logloss training - surprisingly, the test sample (on which the model is automatically selected - not the training sample) and the independent (examination) Logloss_e converge almost perfectly.

So does Recall.

And the indicator Precision surprised me, since by default I usually use it to select the model, I had no training because it immediately equaled 1 on the first tree.

And the different metrics on the test and exam - the result surprises me a lot - a very small delta.

From the graphs of course it is clear that the model is retrained and could have stopped training at 3500 trees, or even earlier, but I didn't adjust the model and the data is actually with default settings.

Mistake somewhere, there is no such thing as an even test and track. Well, or grail, then share :D
 
Maxim Dmitrievsky:
Mistake somewhere, there is no such thing as an even test and track. Or grail, then share :D

It's not a grail, I've got another 100k models and I've got not very good results with them - all 70% of losing models, but profitable ones too.


I think it's the effect of a closed system, ie some kind of stationarity obtained, because the models are similar to each other, I just managed to identify their features well, so there is such a small discrepancy between the results.

I'm finishing the planned predictors, and here's what thought occurs to me - maybe I immediately remove the models that I would not choose myself (large drawdowns, strong imbalance of buys and sales, very small probability distribution, etc.), then the information about obviously bad models will decrease, but there will be more emphasis on choosing a model from hypothetically good (of course, the good model on the test sample may have bad results in the exam). So I don't know whether to cut the sample or not, what do you think?

Well, I will also give up the bare profit as a target - I will select models by a number of criteria - alas, this will reduce the target "1", but maybe there will be deeper relationships, which will allow to evaluate the model on the test results.

 
Martin Cheguevara:
Please advise, I am too lazy to read 1200 pages, has anyone tried to implement machine learning based on results of trading on closed orders of Expert Advisor?

There's no need to read this topic, believe me, you'll litter your mind, try to do immediately as hereWORKING FORESTS PREVIOUSLY TRENDS This is an excellent introductory course on the application of MO in algorithotrading, and in general MO is a very broad subject, in fact MO is an extension of classical statistics, mostly with heuristics and engineering tricks, so it's not a science but a technogenic shamanism, which on the one hand is interesting, but on the other hand is fraught with speculation and abuses. If I've learned about the Indicator Generation it's more probable that the trader has forgotten what he originally started to do, and the MO is a bottomless hole, you can't come back to it, besides you should have a good mathematical background, at least a Bachelor's degree of technical specialty, in order to really deal with indicators instead of boring the parameters of libraries and packages.

Случайные леса предсказывают тренды
Случайные леса предсказывают тренды
  • www.mql5.com
Изначально целью построения торговой системы является предсказание поведения некоторого рыночного инструмента, например, валютной пары. Цели предсказания могут быть разными, мы же ограничимся предсказанием трендов, а точнее предсказанием роста («лонгов») или падения («шортов») значений котировки валютной пары. Обычно, для решения проблемы...
 
Aleksey Vyazmikin:

Not a grail, there are 100k more models and the result is not very good for them - yes completely unprofitable cuts well - only 2% hit, but profitable models slaughtered too many.


I think it's the effect of a closed system, ie some kind of stationarity obtained, because the models are similar to each other, I just managed to identify their features well, so there is such a small discrepancy between the results.

I'm finishing the planned predictors, and here's what thought occurs to me - maybe I immediately remove the models that I would not choose myself (large drawdowns, strong imbalance of buys and sales, very small probability distribution, etc.), then the information about obviously bad models will decrease, but there will be more emphasis on choosing a model from hypothetically good (of course, the good model on the test sample may have bad results in the exam). So I don't know whether to cut the sample or not, what do you think?

Well, I will also give up the bare profit as a target - I will select models by a number of criteria - alas, this will reduce the target "1", but maybe there will be deeper relationships, which will allow to evaluate the model on the test results.

well, of course, if there is an obvious futility, you can remove