Discussion of article "Thomas DeMark's Sequential (TD SEQUENTIAL) using artificial intelligence" - page 3
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
Read.
Well done for deciding to post your TS and explain roughly how it works, plus explain a bit of theory in terms of your thoughts on the work of neural networks directly with the Forex market.
However, the article is too much "trader can guess", "trader should make a decision based on his own experience", etc. All very vague for such a loud title of the article.
As I understand, you are not a programmer, otherwise the article would be more informative, systematic, and you would refine your TS so that you do not reverse positions.
Then you would have posted the results of trading in the tester for a year or two, pre-training the network every fortnight.
Reversal is always at your own risk, there is no clear algorithm - this is a big disadvantage.
It's not even an article about neural networks, it's just a description of your TC.
I expected more specifics, more words about detailed work of the proposed neural network, examples of test samples, training examples, examples of work after training.
Metadological articles are not needed here, beginners will never get it right anyway, and knowledgeable people are not interested in it. People who have been in forex for a long time, who know programming, often look for an interesting idea with a detailed explanation of why it is a good idea, how it works, and how to use it. Then they adapt it to their requirements, build it into their Expert Advisors, indicators, etc.
For example, I have a data clusteriser based on Kohonen Neural Network in C++:
In it, the picture on the left is the original data, the picture on the right is after clustering, with each class signed with a two digit class name, and the minimum distance in the class. The network consists of 7 neurons. There are 49 classes in total.
You don't have a single example, what exactly you give input data and in what format, what you get as output, etc., description of the learning algorithm, etc.
Thanks for your interest. I posted a description of the optimiser on the previous page. There is a link to the optimiser's organisation written by the author himself.
I am programming in MQL4. All the grids are in the files attached to the article. The point is that the task is to construct such a polynomial, which carries the effect of generalisation of the output variable. The polynomial itself may not be large, the main thing is that it understands the essence of the market. I walked in step with it.
Since the classifier divides the sample into two areas. So we compare only the last two signals, knowing the result of the previous one and seeing what group the current signal has fallen into, we can draw certain conclusions.
Yes indeed the article was a bit dry, but its essence was to familiarise the readers with the METHOD of building trading systems. Try to apply the "context of the day" for your TS and it is likely that the training will include paterns that appear during the day.
Regarding reversals, I explained it in the article how it is done.
Looking at the inputs, it is not difficult to guess what data are used to enter. The exit is organised by profit.
I will tell you one secret, only you to nobody.... well????
The point is that it is possible to do without NS and it is done very simply.
There is such a concept as a delta profile. When we have received a signal and know its window. We need to build a delta profile for this window. If the "Context of the day" coincides by volume and OI, i.e. both of them rose or fell at the same time, then we move in accordance with the delta profile, if the delta profile for the maximum volume in the window for buying is positive, then we buy, if it is negative, then we consider this buy signal "False". But if the "Context of the day" is different. That is, the volume has fallen, but the OI has grown. Then we need to work from the opposite direction. Usually in such days the price grows from the negative delta profile and falls from the positive one. It's a clusterfuck. You don't need any Ns here, everything works quite well as it is. Because the Sequenta itself is adaptive, and also the delta profile, which is the reason for the price change. Unfortunately, there is no delta profile indicator for a certain window, so the TS work is good but not very good. But if it would be possible to provide this data to the input of the EA, the issue would be solved coordinately.
If there is a desire to write an Expert Advisor according to the conditions, we can try, but the issue of the delta profile in the window should be solved somehow.....
Not every sample is a second, only those where there is a target, but what can I explain to you... my "Sharp-ratio" has not fallen below 3 for 3 years and this is taking into account that I very often redo everything, reorganise and from this a lot of system transitions, and on the current model, on the real, I have 8 "Sharp-ratio". So, tell someone else about your 30 lines in the "context of the day".
You can put out your dataset of course, it is always interesting, but I will tell you at once that at quantisation more than a minute, there are no fish on the market for a long time already (except for classical insider), and also I have quite effective techniques of detecting overtraining, my system "30 samples" will not take seriously, it will say "not enough data".
Are you familiar with statistics at all? Have you heard of the central limit theorem? About the minimum sample size, that 30 samples is the minimum threshold value of potential representativeness of the sample, for ONE FIGURE, in case of quasi-normal distribution, and it is better from 200 -t . In a multimeter, the minimum sample size limit grows to 30^(D\2) where D is the number of fiches.
It seems to me that you are doing a simple fitting sir, quit this unpromising activity, there is less and less meat, soon it will be like in "developed countries" qualified investors and all that, the suckers will not be allowed to the market and you will have to fight directly with hedge funds and institutional, and those are not 30 lines to learn their models, you will be robbed with this approach)))).
I see that you and I have different approaches to working on the market. I am not going to change anyone's mind or prove anything. I put out the essence of the approach and possibilities of strategy improvement, I hope that it will be useful for someone (not for you, of course, what can I say). However, the meaning of polynomial construction is the ability to divide a multidimensional space by a universal line, which will keep its efficiency for a certain period of time.
Are you aware that I have about 100 columns or inputs in my training file, I will post it, I want to see how your AI will build a model on it. If you would be so kind!!!!!
And about fitting, I will say that I consider data and analyse the model's work exclusively on the out-of-sample part, what fitting is there, I don't understand.....Greetings all, neuronka interested long ago very much and understood and now I also think that behind it the future and the present already, of course with its own peculiarities, I have such a question of the essence of neuronka, if I understand correctly that when teaching it certain algorithms it looks for similar in the future and trades on them, of course in the digital and structural movement in the market is never identical, and there are always shifts, ie.For example, a signal is given and on the history it is repeated 100 times, but out of 100 times naturally only a small amount brings profit, the rest are in the negative. Why there is no neuronka which is made exactly on patterns or on clear signals, what is a clear signal, i.e. at certain settings of any indicator neuronka should find on history such repetition which will give a positive efficiency, as a result here it is the number of separate signals on me should give the result with good profitability. I.e. any indicator and profitable and not profitable in skilful hands, why not to cross 1 or more indicators in one strategy, further by selection of the best results of adjustment through neuronka not to deduce the best results and trade further on them. Is there anyone who can do such a thing or can do it, I would participate.
You are absolutely right, there is a variant of the network when it looks for patterns in the future, studies them and then makes the same move in the future. However, it is very difficult to find a pattern in the past that would be interpreted unambiguously. That is, out of 100 patterns in the past (one and the same pattern), 50 will give a plus and 50 will give a minus. What then? What should the NS do? It is important when preparing data that there is no contradiction when the same pattern produces different results. Therefore, the process of generalisation is the solution, when at the appearance of a pattern, based on the information of inputs, a conclusion is made about the truth or falsity of the signal.
I see that you and I have very different approaches to working on the market. I am not going to change anyone's mind or prove anything. I have laid out the essence of the approach and possibilities to improve the strategy, I hope that it will be useful to someone (not to you, of course, what can I say). However, the point of building a polynomial is the ability to divide a multidimensional space by a universal line, which will remain functional for a certain period of time.
Are you aware that I have about 100 columns or inputs in my training file, I will post it, I want to see how your AI will build a model on it. If you would be so kind!!!!!
And about fitting, I'll tell you that I consider data and analyse the model's work exclusively on out-of-sample participants, what fitting is there, I don't understand.....Pardon for a slight aggression, I rarely communicate on forums and social networks, so sometimes I can behave not politically correct. Reactions and quotes to my harsh posts can be removed or edited if you want.
For sharing your approach, thank you, in our business it is rare, even if the approach is original, actually it MUST be original, how else, if we rob each other and the doll. Here you have to be original and inventive))))))
About the second dataset that you posted to say nothing I can not say anything alas, for 100 chips need at least 10k samples for training and at least a thousand on the test that would say something definite. But my model can cope with both one chip and 10k.
I will output the linear model above, trained on 50 samples, in the evening I will check, purely for the sake of experiment, although IMHO the probability that it will give at least 51% on 14k test samples is close to zero.
I will post the result if you like.
Pardon the slight aggression, I rarely interact on forums and social media, so I may not behave politically correct at times.
For sharing your approach, thank you, in our business it is rare, even if the approach is peculiar, in fact it MUST be peculiar, how else, if we rob each other and the doll. Here you have to be original and inventive))))))
I can't say anything about the second dataset that you posted, alas, for 100 chips you need at least 10k samples for training and at least a thousand on the test that would say something definite.
I will output the linear model above, trained on 50 samples, in the evening I will check, purely for the sake of experimentation, although IMHO the probability that it will give at least 51% on 14k test samples is close to zero.
I'll post the result if you like.
Definitely post it. I will be only glad!!! Interesting results.....
Pardon the slight aggression, I rarely interact on forums and social media, so I may not behave politically correct at times.
For sharing your approach, thank you, in our business it is rare, even if the approach is peculiar, in fact it MUST be peculiar, how else, if we rob each other and the doll. Here you have to be original and inventive))))))
About the second dataset that you posted to say nothing I can not say anything alas, for 100 chips need at least 10k samples for training and at least a thousand on the test that would say something definite. But my model can cope with both one feature and 10k.
I will output the linear model above, trained on 50 samples, in the evening I will check, just for the sake of experiment, although IMHO the probability that it will give at least 51% on 14k test samples is close to zero.
I'll post the result if you like.
It's a bit unclear with definitions, but I noticed that with a 100 by 100 matrix, when we have 100 columns (inputs) and 100 records (signals) the model gets an increase in generalisability. That is, a 100 X 100 matrix will train better than a 10X100 matrix and worse than a 100X10 matrix, where the first digit is the number of inputs and the second digit is the number of signals. That is, when the number of inputs significantly exceeds the number of signals, the NS, as they say, has a lot to choose from and therefore the model gets a higher level of generalisation. When the number of inputs is much less than the number of signals, then the model is not very good, because it becomes difficult to choose because of the emergence of contradictory data. Also, I can tell you one trick, but this is already in private. The trick is not very important, but the effect it brings is significantly interesting.
So... please explain a couple of things, for example, I take a toad file, is it a trained model? There are four methods, there are no references to references and in the methods themselves coefficients are shardcoded, but they take 5 chips, and there were 15, what chips did you use or how did you reduce the dimensionality of 15->5?
And do I need *.vrm binary for something, if I just need to run a test?