Machine learning in trading: theory, models, practice and algo-trading - page 3228

New comment
 
fxsaber #:

They show OOS on 20 sets taken next to different peaks of the target function. This means that if there are 19 false (fit) peaks and one positive (pattern) peak, we will see it immediately. And we won't care about all other results.

Answered this question here.

No offence, but I just don't understand the attempt of a purely theoretical person to influence the decisions of an expert practitioner. Even at the zero stage of selecting initial data (quotes) I disagree with you fundamentally.


MO researchers, as a rule, use a hypothesis that there is a pattern in the initial series, which can be traded in plus. This is a hypothesis, not confirmed by anything.

And then I give out a series that has a pattern 99.9% of the time. I expect that the most advanced generation methods should not break it at all.

If you can create such a generation with GARCH, honour and praise.

Don't generalise your ignorance of confirmation.

This is exactly why you are not responding substantively: the IO is looking for patterns that predict the future, not just patterns.

 

Theoretical question.


Condition.

  1. Someone has invented a very simple TS.
  2. Realised what should be done with quotes, so that the best passages on any Sample after optimization of the TS would show themselves perfectly on OOS.
  3. I generated a random sequence of quotes and added the result of point 2 to it.

As a result, there is a certain history, which definitely has a regularity. This history can be of any required length.


Question.

Will a super-advanced (someday in the future) MO methodology with infinite computational capabilities find the TS (data from p.3) as in p.1. or with properties as in p.2?

 
Mine finds even where it is not 😀 but it is not designed for ticks

I'm speeding up some f-i's, but so far I can only dream of ticks.

For example, can yours show oos on bars? Mine does. It's not clear what to compare with what, it depends on the approach.

I'll try generative grid on ticks later, I've forgotten Pythorch, I need to refresh my memory. Everything should be found, I just tried fast methods of sampling from distributions. They are generally Markovian and are not designed to reproduce serial patterns. I've contrived and added new dimensions just to generate series in chunks at once, instead of one sample at a time. But something doesn't work.
 
fxsaber #:

Theoretical question.


Condition.

  1. Someone has invented a very simple TS.
  2. Realised what should be done with quotes, so that on any Sample after optimisation of the TS the best passes would show themselves perfectly on OOS.
  3. I generated a random sequence of quotes and added the result of point 2 to it.

As a result, there is a certain history, which definitely has a regularity. This history can be of any required length.


Question.

Will a super-advanced (someday in the future) MO technique with infinite computational power find a TC (data from item 3) as in item 1. or with properties as in item 2?

It will not find your strategy specifically. 100 fics can be divided in millions of different ways. Depends on the formula used to select the best split. Here, I think, someone generated a bunch of different trees and selected from them by OOS.
Random Forest also generates many trees (each tree is given several random fiches) and then averages the result of all trees. If there are patterns in most of the fiches, the result will be good and stable. If there is one common pattern in all fiches.
If there are only 1-3 good fiches, then they will be averaged with trees with only noisy fiches, and the result will be noisy. In trading, all chips can be considered noise.

If you have only 1000 trades out of 6 million ticks (activated by some condition of yours), that is 0.017% of all data. MO will never find such a thing, it will find something common and try to trade on 50% of ticks, you can tighten the conditions and trade 1% (only on the best leaves), but you have even less.
Basically, each leaf is a separate strategy like yours. But a tree can be divided into 100 leaves or 1000 or 10000.... And it trades all these 10000 strategies at the same time (or rather the ones you choose, you can for example only trade leaves with 90% or even 99% probability of winning on Traine, but on OOS such a clean (maybe retrained) leaf does not guarantee anything).
 
Forester #:
If you have only 1000 trades out of 6 million ticks (activated by some condition of yours), it is 0.017% of all data. MO will never find such a thing, it will find something common and try to trade on 50% of ticks, you can tighten the conditions and trade 1% (only on the best leaves), but you have even less.

1000 trades in six months is very active trading. If it's not enough, then what are you looking for on MO when you feed in many years? Duration distribution cited.

Honestly, very rarely seen more trades. And not significant - times two only. But there on the edge of stability.

 
fxsaber #:

Question.

Will a super-advanced (someday in the future created) MO technique with limitless computational capabilities find TCs (data from p.3) as in p.1. or with properties as in p.2?

The question is rather strange....
If you have infinite computational power, you can do anything, even now
 
mytarmailS #:
That's a very strange question...
If you have infinite computing power, you can do anything, even now.

And if you think about it logically?

If there are infinite computational possibilities, then all the money will flow into these computational wallets.

Who's going to let them if other computational powers want to take it away from the first.

Think about it at your leisure. It's good to use logic.))

 

Forum on trading, automated trading systems and testing trading strategies

Machine learning in trading: theory, models, practice and algorithmic trading

Renat Fatkhullin , 2023.09.10 10:44

We plan to launch another championship aimed at promoting neural networks:
1) we will provide a single MQL5 robot template with downloadable model.onnx
2) within 5 months participants will upload their modules as model.onnx
3) daily they will be automatically run on history from 2023.01.01 to the current day at 4 main exchange rates
4) the daily rating of participants will be published
5) at the end of the preliminary period of accumulation of participants within 5 months, the real trading period will begin within 1 month
6) based on the results of work within 1 month, the winners will be determined
7) the prize fund from our company 30,000 dollars will be divided into three winners as 15,000, 10,000 and 5,000 dollars
8) we guarantee that at the end of the championship all model files will be deleted to preserve the intellectual property of the developers

The purpose of the championship is solely to stimulate the development of machine learning in trading. Programs only in the form of a single unchangeable MQL5 template + model.onnx

 
fxsaber #:

1000 trades in six months is a very active trade. If that is not enough, then what are you looking for on MO when you feed in many years? The distribution of duration was given.

To be honest, very rarely have I seen more trades. And not significant - times two only. But there on the edge of stability.

Well, manually I did 100 trades a day on 8 instruments, probably out of boredom and adrenaline)). On 1 instrument I probably did 10 a day on average - it will turn out to be about the same as yours. I averaged and of course I lost. That's why I switched to algo-trading, and then to MO. More precisely algo-testing, because no model interested me to put money on it.


I just make a forecast for each bar, now it's M5 (288 bars per day) and if the forecast is good - you can trade. If we take the probability of success 0.5, then on average 144 trades per day. If 0.9, then less, if 0.99, it can stand idle for a year and then actively trade for a week (White Swan Catcher).
By analogy with bars, you can forecast every tick, and there are 6 million of them - that's why I said "only 1000". On ticks you should probably not predict every line/tick, but filter them in some way. You did it with your manually detected algorithm and got 6-7 trades per day.

 
Uladzimir Izerski #:

And if you think about it logically?

If there is limitless computing power, then all the money will flow into these computing wallets.

Who will let them, if other computational capabilities want to take it away from the first.

Think about it at your leisure. It's useful to apply logic.))

Don't talk about logic, you have no idea what it is.....
Get a dictionary, read the meaning of the word infinite/infinite.

Then answer the question whether there are supercomputers with infinite deductive powers on the planet....

Then answer the question whether supercomputer owners need to make money if they print it.


He's applying logic.
What do you know about logic and reasoning?
1...322132223223322432253226322732283229323032313232323332343235...3574
New comment