Searching for an arbitrary pattern using a neural network - 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
Do you think there is a difference between 'human describes' and 'neural network counts'?
A neural network is trained by examples, without going into the details of dependencies.
Exactly. First you should prepare 100500 examples of different "head-shoulders" for it and teach it on these examples.
Actually, price patterns can be described by mathematics, you don't need NS for that. But trying to find the signs of a false pattern is exactly the task for NS.
Do you think there is a difference between 'human describes' and 'neural network counts'?
A neural network is trained by examples, without going into details of dependencies.
Dmitry, please explain in more detail your answer, yes they do not go into details of dependencies, but it is mathematics that underlies, I think (imho) that in the basis of human actions also lies mathematics, itself it is more complex but also 1+1=2
it is more complex but also 1+1=2
for NS 1+1=2 , but with a given accuracy
for some types of NS not every training result will be identical to the previous training, but they (the results) will be equal to the accuracy equal to the training error
The basis of NS is not so much mathematics, (yes, NS training is a mathematical calculation), but it is the type of NS, structure of NS. activation function, what do you teach NS classification or regression ... I.e. you can't just say that NS gave me this result - it's right or vice versa NS is lying. NS is a black box, not because you want to call it that, but because NS usesthe black box model
Dimitri, please explain in more detail your answer, yes they do not go into details of dependencies, but it is mathematics that underlies, I think (imho) that human actions are also based on mathematics, itself it is more complex but also 1+1=2
You need to know a little about how neural networks are used. There is an input image, there is an output result. Having a certain large number of such pairs (input image - output result) neuronet is trained. And nobody cares why some image gives such a result, there is just a set of experienced facts and that's all. Then the analyzed image is fed to the input of the trained network and the result is seen at the output.
You need to know a little bit about how neural networks are used. There is an input image, there is an output result. Having some large number of such pairs (input image - output result), a neural network is trained. And nobody cares why some image gives such a result, there is just a set of experienced facts and that's all. Then a trained network receives an image to be analyzed and looks at the result at the output.
You have now described learning with a teacher. When it is possible to collect historical data and ask the network to find the mathematical law (model) that would describe it as accurately as possible. But sometimes there is no historical data as such, but a bunch of patterns that need to be decomposed, sorted out so to speak. All this is done by different networks. That is, a multilevel system of AI is designed in which one network sorts the patterns and another one defines their validity. Question: Why do we need the first NS if we can mathematically arrange them on the shelves WITHOUT NS? Much more interesting is the answer to the other question about the truth of the formed pattern.
Want to give a key tip!!!!!!?????? I can see it in your eyes :-) OK, I'll give it as an example.
If we consider any pattern as a fait accompli (pattern formed), then it only gives us a moment to analyze the market. Suppose we start to make our own calculations within the next bar after the pattern has been formed. It means, in fact, the pattern itself gives us only the time when we should perform the calculation, but we leave it to NS to decide what kind of pattern it is, what conditions have led to its formation and what happened when it was formed. If we consider the condition that any pattern gives us only the moment to analyze it, then having a set of 10 patterns we will obtain much more bars for analysis. No need to train one net for "head shoulders", another for "three soldiers", etc. But it is necessary to inform the NS about what kind of pattern it is and it's done very simply.
So, we have written an indicator that can detect 5 patterns for buying and 5 for selling at the very least. Of course the future order of patterns is not known and they appear chaotically. Believe what kind of pattern it is, the NS does not really care, it looks at the input dataset at the moment of forming any of the patterns and theoretically if patterns differ drastically, the input datasets will also differ significantly to the extent that the net will see it. But this is not enough for us and we want to forcefully tell the net what kind of pattern it is. It is very easy to do. Patterns are encoded from -5 to +5 and the input values are multiplied at the earliest stage of data conversion. Multiplication scatters the data along the Y axis thus making the data of one pattern be multiplied by the same number and shifted by a certain distance. As a result the inputs will be dependent on the pattern type. So what we have in the end:
1. We write a basic indicator that defines patterns, forming signals for the analysis.
2. define the set of input data
Determine the internal structure of the NS, training methods, error analysis methods etc.
4. In the basic indicator make a buffer for the output variable. Remember that we don't know the result of the most recent pattern. The buffer should be such that the future values are attributed to the patterns. When saving the training file we will know the results of ALL patterns except for the last one.
5. We work out a method of checking the obtained models for the presence of generalizing ability.
That's it......
And another interesting point came to mind when looking at the thread title.
Suppose we do want to find arbitrary patterns using NS that we don't know. Question: If we don't know the patterns themselves, then what is known? Correct, we know the reactions to those patterns, or rather we must choose under what conditions to look for the patterns ourselves. Let's formulate the formulation of the problem:
To find patterns of 5 candlesticks after which the rate changes by more than 10% during the next 4 candlesticks. Of course we can arrange an unloading of this kind from the history and generate a training file containing only 5 bars preceding the market reaction for each case. Then we teach the net to show 1 for bars preceding each rise and -1 for all other bars. After training, we start sending data of the last 5 bars to the input of the net bar by bar and when the grid shows 1, then inputs will contain exactly the same pattern or a pattern similar to it as in our training.
With this approach we will not know what kind of pattern it is and what its parameters are. Note that I limited 5 bars on inputs when this number is floating, when during training we also tune input window, then number of optimization results increases many times and number of patterns becomes number of saved data when every record is so unique that out of 1000 records we have 1000 clusters. IMHO!
I believe the approach has its place, although it has certain limitations. At least it doesn't break any critical rules, such as peeping, etc.
Either with or without a teacher - a different perspective on the same thing. The situation and its outcome must be known; if there is no such thing, you cannot teach anything. And you don't need God's gift with the egg - that is, teaching and classification.
When learning without a teacher, you usually don't know the outcome, so what about in this case?
Mihail Marchukajtes:
Michael, the neural network works a little differently. It doesn't work the way you want it to.
When learning without a teacher, the outcome is usually unknown, so how does this work?
Only classification. The network learns to distinguish between situations (images), but it cannot know what to do in which case or what to call which image.