Machine learning in trading: theory, models, practice and algo-trading - page 108
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
In continuation of my small publication https://www.mql5.com/ru/forum/86386/page98 about spectral analysis and adaptation to the real market parameters.
I made a small experiment, just to reinforce the theory with practice, the essence of the experiment is to check whether the indicator will be more effective if you change the period every time to the one that is objectively present in the market
the indicator took the "RSI" (just out of luck), trading rules are elementary more than 70% sell, buy less than 30%, stupid trading reversals, no stops
First I took the usual RSI indicator with a period of 14 (this period is the most common in all books and articles) just to compare it with something
the indicator didn't lose, i'm honestly surprised ....
Now the adaptive RSI
Conclusion: the adaptive approach is much more affective than the usual
1) But, in fact, there is no such thing as "training" and "coaching. All sorts of crossvalidation and OOS checks do not and cannot give the effect expected of them. The point is that such tricks are nothing but a search and then selection of those values, which approximate work satisfactorily both in the training and in the testing area, i.e. this set of parameters already exists initially among all their possible variants, and it is equivalent to selection at once on the whole history area.
2) Nevertheless, the use of two models (in my case, two grids) is, in my opinion, the best that can be applied among currently available methods of "machine learning". It's not training or coaching, it's a way to optimize the model.
3) There is no real training available today. Recognizing the same or similar patterns is not a result of learning, it is a result of remembering. Learning must imply some thought process (albeit primitive), which would allow reasoning and making conclusions when receiving new information, as well as the ability to generate new information independently. The market requires just such an approach - thinking, which as far as I know does not exist today. And what we use today - memorization, not thinking, unfortunately.
1) The thought is deep and correct. But not complete.
The CV (crossvalid.) is learning and testing on the same parameters, at different training and testing sites. Even 10 different fouls. If the machine is learning noise, the average quality metric will be weak. The method itself is very strong.
But, if the data is noisy, there may be a CV fit, which is what you're talking about, but you don't finish the thought in technical terms and get bogged down in pessimism. There has been a nested CV (nested CV) for a long time now. All your selected models can be validated on unique out-of-sample data. If there is consistency of results, the model is good, if not, bad. Everything is solvable.
2) It's not clear why this is the case.
3) It is. But machine learning is an industry that relies on generalizable understanding. Fighting overtraining is 90% of the effort.
The "bad machine guys" take it into account. The time is fed to the input of the machine. In addition, the price behaves differently not only at night, but also by sessions.
"The Manin Boys... Cool!
It is good that they take into account, I read the thread - did not notice any mention of this, decided to share my thoughts. True, the clear characteristic signs to identify individual sessions, I have not found, so I apply only the limitation by the hour, from now to now.
2) It is not clear why this is so.
In continuation of my small publication https://www.mql5.com/ru/forum/86386/page98 about spectral analysis and adaptation to the real market parameters.
I made a small experiment, just to reinforce the theory with practice, the essence of the experiment is to check whether the indicator will be more effective if you change the period of the indicator to the one that is objectively present in the market
the indicator took the "RSI" (just out of luck), trading rules are elementary more than 70% sell, buy less than 30%, stupid trading reversals, no stops
First I took a regular RSI indicator with a period of 14 (this period is the most common in all books and articles) just to compare it with something
the indicator didn't lose, i'm honestly surprised ....
now the adaptive RSI
Conclusion: the adaptive approach is much more affective conventional
I followed the link, but did not understand how you dynamically change the period of the indicator.
Please explain in more detail.
In continuation of my small publication https://www.mql5.com/ru/forum/86386/page98 about spectral analysis and adaptation to the real market parameters.
I made a small experiment, just to prove the theory with practice, the essence of the experiment is to check if the indicator will be more effective if every time you change the period, which is objectively present on the market
I followed the link, but I do not understand how you dynamically change the period of the indicator.
Please explain in more detail.
Everything is clear about the adaptation. And where do you take the period, "which is now objectively present in the market"
The advantage is a decrease in the number of trades over time since the start of trading on the OOS, not an increase in the percentage of wrong signals. More and more often there are contradictions between nets, at the same time one says to sell and the other says to buy, and this is signal 0, that is, instead of trading losses on completely unfamiliar data, the model stops trading.
The idea is interesting.
Ato.
In fact, I use this effect in the validation plots as an indicator of the correctness of the training, rather than the ratio of correct/incorrect answers (error). This is an important trade property and an indicator of the quality of training. If the model gives wrong signals on OOS - this is an indicator of wrong training, not the fact of market changes.