Machine learning in trading: theory, models, practice and algo-trading - page 46
![MQL5 - Language of trade strategies built-in the MetaTrader 5 client terminal](https://c.mql5.com/i/registerlandings/logo-2.png)
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
I agree, it's interesting... But there's almost nothing clear to me, starting with the ideology and ending with the code itself, there's a lot and many operators I don't even know
If somebody could explain it all at least by elementary examples how to apply it in trading it would be a good experiment for such inexperienced people as me
You need to look for examples yourself on the Internet.
Very interesting neuronethttp://gekkoquant.com/2016/05/08/evolving-neural-networks-through-augmenting-topologies-part-3-of-4/ Do you think it's possible to make it trade itself and learn from its mistakes? If yes - how, I would like to discuss it.
The feature of this neural network is its adaptive topology. This is not just a set of inputs, hidden neurons and outputs; this is a model where neurons link and disconnect from each other during the evolution, change their weights, thus gradually adapting the network gives better and better results. The result should be a network with unique neural connections and weights, well suited to the particular task.
For forex, no miracle will happen, the network will simply learn from previously prepared examples, just like a normal network. Most likely it will even produce 100% accuracy. But in fronttest it will probably drain the entire balance, why should it not? :)
I once tried to train the neuron in the Expert Advisor itself, retraining it on each new bar. The network increased its balance, but after some time intervals it suddenly lost more than it earned. Then it began to increase the balance again, and after some time, again suddenly lost a lot. As if sometimes events that dramatically change all the internal processes of behavior of the forex pair occur, and the model becomes completely unusable for a while until it is brought up to date again. I've discarded this approach, it is too difficult, it is necessary to adjust the speed of learning new data, to introduce logic like "if the profit has fallen by X points within Y days, then stop trading for Z days", to review and optimize it all. It's easier to train a new network from scratch once a month.
The peculiarity of this neural network is adaptive topology. This is not just a set of inputs, hidden neurons and outputs; it is a model where in the process of evolution neurons are connected and disconnected from each other, their weights are changing, so gradually adapting the network gives better and better results. The result should be a network with unique neural connections and weights, well suited to the particular task.
For forex, no miracle will happen, the network will simply learn from previously prepared examples, just like a normal network. Most likely it will even produce 100% accuracy. But in fronttest it will probably drain the entire balance, why should it not? :)
I once tried to train the neuron in the Expert Advisor itself, retraining it on each new bar. The network increased its balance, but after some time intervals it suddenly lost more than it earned. Then it began to increase the balance again, and after some time, again suddenly lost a lot. As if sometimes events that dramatically change all the internal processes of behavior of the forex pair occur, and the model becomes completely unusable for a while until it is brought up to date again. I've discarded this approach, it is too difficult, it is necessary to adjust the speed of learning new data, to introduce logic like "if the profit has fallen by X points within Y days, then stop trading for Z days", to review and optimize it all. It is easier to train a new network from scratch once a month.
That's interesting.
According to the idea, if you set the experiment correctly (Early Learning Stop!), then such adaptation can be useful.
They seem to be preparing a package for R there. We should take note.
1) For forex no miracle will happen, this network will simply learn on pre-prepared examples, just like a normal network. Most likely it will even give them 100% accuracy. But in fronttest probably will leak the entire balance, why would not it leak? :)
2) I once tried to train the neuron in the Expert Advisor itself, retraining it on each new bar. The network has increased its balance, but after some time it suddenly lost more than it had earned. Then again began to increase the balance, and after some time again suddenly lost a lot.
1) You may be right, but this net is able to learn by itself how to make decisions, it is not the usual classification without a teacher, which means that it is possible to implement the concept I told about long ago - you can teach it not as a standard target in the form of buy-sel-buy or 0001110101011, but in a more abstract way, like simply putting conditions like: "Net! I don't care how you trade there, but I want your daily income to be at least 1% with a 0.5% drawdown, and it will look for the rules and combinations for solving this problem. If I'm wrong somewhere and have said some nonsense then correct me for my own good.)
2) I tried the same thing the day before yesterday but in a slightly different way... On a 5-minute trailing window of 150 candles and at each new candle I trained Forest and traded, then on a new candle I retrained the model, etc... The results were surprisingly good, about 5 times I ran on the same data this kind of trade, the model was always in the black from 8% to 20% per month, I was already excited and thought I would run one more time) and then plum, one more time plum again)) In short it turns out that just by chance the model was earning...
By the way I tried the following thing: after each retraining via "importense" in RF I found the most important features, that is, as if "on the fly" and trained the model only on the significant ones, the model after that started to work about 2 times worse)))) What I was very surprised)))
This is a very interesting topic.
But if we work with NS, then the number of inputs should be reduced as much as possible, in my opinion.
Each extra input "weighs down" the network, reduces its teachability and leads to simple data memorization or, as discussed here, to tossing between inputs/retraining.
This is a very interesting topic.
But if we work with NS, then the number of inputs should be reduced as much as possible, in my opinion.
Each extra input "weighs down" the network, reduces its teachability and leads to simple data memorization or, as discussed here, to tossing between inputs/retraining.
This is not a question. You can select any desired number of inputs before training.
That's true.
But, unfortunately, there is an opinion that the more you submit, the better.
And she, that is, the National Security Service, will take away what is needed.
This approach is fundamentally wrong.
That's true.
But, unfortunately, there is an opinion that the more you submit, the better.
And she, i.e. the NS, is supposed to take away what is necessary.
This approach is fundamentally wrong.
I'll add an intrigue -- you don't have to file a change in the rate of the traded item.
It's like dragging yourself by your hair out of a swamp.
Look for other sources of data as well.
May Profit be with you!
:)