Machine learning in trading: theory, models, practice and algo-trading - page 106
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
SELL BUY Interpretation
-1 0 sell
0 0 fence
0 1 buy
-1 1 fence
Here is the switch table. From it you can see that the signals occur when the grid signals are there and do not contradict each other.
Thanks! So, like in my ternary classifier - inconsistency of two binary classifiers - sit on the fence.
It turns out that we arrived at the same solution in parallel: one ternary can be assembled from two binary classifiers. This is easier to implement than a three-output ternary, since with a three-output ternary it is not clear how to interpret disagreements, since all outputs must be mutually exclusive, and in fact this does not always happen.
And our conclusions are the same, too: a ternary of two binaries has a higher generalization ability than individual binaries.
Thank you! So, like my ternary classifier, the inconsistency of the two binary classifiers is sitting on the fence.
It turns out that we came up with the same solution in parallel: one ternary can be assembled from two binary classifiers. This is easier to implement than a three-output ternary, since with a three-output ternary it is not clear how to interpret disagreements, since all outputs must be mutually exclusive, and in fact this does not always happen.
And our conclusions are also the same: a ternary of two binary has a higher generalization ability than individual binarys.
Actually, adding two binary classes to a ternary does not equal one ternary.
When you add a third class you get as a contradiction of two classes. And what if the original ternary? by-invariant-market-settle? For example in terms of a zigzag, what does it look like? Is the out-of-market a sideways move in a zigzag? I.e. a completely different target variable than yours.
With these explanations, it sounds plausible. And how rare patterns do you trade? Say, if there are training examples with the classes"open buyposition " and "close all trades", what will be the percentage ratio of these classes? I can assume that the "buy" class corresponds to the sharp price movement upwards by hundreds of points, i.e. the amount of "buy" class is about 10% of all the training examples?
I don't apply "close" signals. The neuronics knows only how to enter the market, and how to exit - the trading system decides. This is better than applying this very TS to just random signals in general.
In 4th forum I wrote earlier that I am deeply convinced that for neuronkey and similar machine learning systems only TSs with time-limited action of trade signals are good enough. Initially there appear insurmountable contradictions for neurons (to buy or sell with counting on infinite future equally and without advantages over each other), that is why for neurons I do not allow to decide when to close trades.
I don't use "close" signals. Neurons only know how to enter the market, and the trading system decides how to exit. This is better than applying this very TS to just random signals in general.
In 4th forum I wrote earlier that I am deeply convinced that for neuronkey and similar machine learning systems only TSs with time-limited action of trade signals are good enough. Initially there appear insurmountable contradictions for neurons (to buy or sell with counting on infinite future equally and without advantages over each other), that is why for neurons I do not allow to decide when to close trades.
Your post made me happy.
I've been preaching here that considering machine learning algorithms outside of a trading system is a hollow exercise. You are the first one to make the same point.
You have to dance from the trading system and use machine learning models to try to improve the parameters of the REAL trading system. I put in an rf and reduced the drawdown. In fact the real TS has much more characteristics than the buy/sell.
1) There is a network that can be trained if it receives data that were not previously known to it. You may be interested to read about it, the network is called SOINNhttps://www.google.com.ua/webhp?sourceid=chrome-instant&ion=1&espv=2&ie=UTF-8#q=soinn
2) I'm tired of writing about it, moreover I practically proved that the market goes against my own statistics and even explained the mechanics, why it happens and all the classical trainings are not applicable to it, but nobody is interested, everybody does the same thing as I did
1. thank you, I will read. though. see p2.
2. the problem of changing patterns over time is not solvable, in my opinion, even with Pre-Learning, because you not only need to add new patterns to the knowledge base, you need to find the ones that have changed and whether they have changed at all... This is a very difficult and seemingly insurmountable task. It's a section of AI, of thinking, of intelligence, it's like learning the multiplication table and being able to apply math to all problems, even new unconventional and unfamiliar ones. It is a section of thinking. capable of making discoveries, that is, capable of independently generating useful information and knowledge. But, clever uses of neurons are not forbidden and are possible, of course: I think it is simply from the section of statistics and probability calculation, statistics for memorization and probability for safety net. A successful combination of statistics and probability calculation would give mo more spread, maybe even be, at least all hope is for that alone.
Your post made me happy.
I've been preaching here that considering machine learning algorithms outside of a trading system is a hollow exercise. You are the first to make a similar point.
You have to dance from the trading system and use machine learning models to try to improve the parameters of the REAL trading system. I put in an rf and reduced the drawdown. In fact the real TS has much more characteristics than the buy/sell.
Glad to have pleased you, quite a bit.
I'm surprised that machine learning can be considered by anyone at all in isolation from any particular TS...
For forex it is also possible, but it requires just as much effort.
Yes, but not on the configuration of the neuronics.
Thanks! So, like my ternary classifier, the inconsistency of the two binary classifiers is sitting on the fence.
It turns out that we came up with the same solution in parallel: one ternary can be assembled from two binary classifiers. This is easier to implement than a three-output ternary, since with a three-output ternary it is not clear how to interpret disagreements, since all outputs must be mutually exclusive, and in fact this does not always happen.
And our conclusions are also the same: a ternary of two binary has a higher generalization ability than individual binarys.
You're welcome).
Well, yes. A simple ternary doesn't have the wonderful effect of reducing the number of trades while decreasing market recognition (in addition to the difficulty of interpreting the signals). But it's through the ternarnik that I came to this (true to the word learned here).
2. the problem of changing patterns over time is unsolvable, in my opinion, even with Pre-Learning, because you not only need to add new patterns to the knowledge base, you need to find the ones that have changed and whether they have changed at all...
The issue of pre-learning is well illustrated in randomforest.
Example.
Fitting a model on a sample of 5000 bars. The model itself gives a graph that shows how the error changes as the number of trees grows, and the tree is the pattern. With my predictors I can use 100 trees. I take 300, so that the error is completely straight. I start fitting (not applying) on a sample of 20,000 bars. The surprising thing is that the error graph has not changed! The number of trees remains the same. I.e. the pattern variants were already exhausted at the first 5000 bars.
But the most unpleasant thing is that this does not solve the problem. And the problem is the retraining of the model, i.e. some particularities are taken and the trees are built on them.
What is the problem?
The problem is the predictors, not the model.
I know how to measure the predictive power of predictors. So not only do I have to discard predictors that have no predictive ability (noise predictors), but the predictive ability must be stationary. I could not find such predictors. And this problem persists.
You're welcome).
Well, yes. A simple ternary doesn't have the wonderful effect of reducing the number of trades while decreasing market recognition (in addition to the difficulty of interpreting the signals). But it is through the ternary that I arrived at it (verve learned that word here).
Ternary means that it can take three mutually exclusive states. Another name for it is ternary.
A grid with three outputs, each binary, can produce 8 mutually exclusive states of which only three are interpreted unambiguously, like a ternary. And the remaining 5 states are unclear how to interpret them?