Machine learning in trading: theory, models, practice and algo-trading - page 3227

 
fxsaber #:

Unfortunately, these are all hypotheses that require realisation and testing.

@Maxim Dmitrievsky is trying his options, I am trying mine.

Yes, of course, every approach needs testing.....

Here is another Python method for generating time series from a sample.

The topic is interesting, but I can't devote enough time to it yet.

 
Aleksey Vyazmikin #:

The topic is interesting, but, so far, no opportunity to devote sufficient time to it.

I wonder if Kaggle would do it.....

 
We are planning to launch another championship aimed at promoting neural networks:
1) we will provide a single MQL5 robot template with downloadable model.onnx
2) during 5 months participants will upload their modules in the form of model.onnx
3) every day they will be automatically run on the history from 2023.01.01 to the current day on 4 major exchange rates
4) daily ranking 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 start within 1 month
6) according to the results of work during 1 month 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 erased to preserve the intellectual property of the developers

The purpose of the championship is solely to stimulate the development of machine learning in trading. Programmes only as a single unchangeable MQL5 template + model.onnx
 
Awesome, that's great!
 
Renat Fatkhullin #:
6) winners will be determined based on the results within 1 month

Is randomness being beaten out for prizes? Why not use the Kaggle method when there is a closed LONG sample? Then backtest on it and OnTester shows the winner immediately.

 
Renat Fatkhullin #:
3) daily they will be automatically run on history from 2023.01.01 to the current day at the 4 major exchange rates

Since we are far from algo-trading, I inform you that MQ-Demo quotes have a very low earning potential. Roughly speaking, if I know the future, in MQ-demo with perfect execution I would earn 100 u.e., and on XXX-demo 1000 u.e.. This suggests that many existing (traded on the real not in the kitchens) patterns you simply can not try out.

Therefore, if you want, for example, to attract scalpers, who have a very high statistical significance of the results, you need to change something in the source of quoting MQ-demo. Now it is not suitable for many studies.

 
Renat Fatkhullin #:
3) daily they will be automatically run on history from 2023.01.01 to the current day at 4 major exchange rates
Look in the same Signals and see on which symbols the leaders are mostly trading.
 
fxsaber #:


It seems like an optimisation graph can show how hard the search process is going. So here we go.


What are these graphs for?

Well, suppose you find some patterns in HISTORY and learn how to generate similar ones.

Why?

We need such repetitive pieces of graphs, AFTER which a quite definite section follows LEGALLY. Exactly AFTER, in the FUTURE. All economic science can be divided into two parts: analysis and prediction. But prediction does NOT follow from analysis, because all financial data is NOT stationary.

Hence MO models.

Any MO model is adept at finding patterns, similar plots on historical data. What is fundamental in MO is that these patterns - PATTERNS are put in accordance with the FUTURE value of the teacher. Only in MOE is such an approach possible - FUTURE.

By the way, in GARCH one looks for a mathematical pattern and predicts the future, hoping that the pattern found will not change.

 
Renat Fatkhullin #:
3) daily they will be automatically run on history from 2023.01.01 to the current day at the 4 major exchange rates

One model is unlikely to work well on different currencies, rather you need a different model for each currency.

 
СанСаныч Фоменко #:

What are these graphs for?

They show the 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 the other results.

Well, suppose you find some patterns on HISTORY and learn how to generate similar ones.

Why?

Answered that question here.

We need such repetitive pieces of graphs, AFTER which a quite definite section follows LEGALLY. Exactly AFTER, in the FUTURE. All economic science can be divided into two parts: analysis and prediction. But prediction does NOT follow from analysis because all financial data is NOT stationary.

No offence, but I just don't understand the attempt of a purely theoretical person to influence the decisions of a skilled practitioner. Even at the zero stage of choosing the 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, which is not confirmed by anything.

And then I give out a series in which 99.9% there is a pattern. 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.