Searching for an arbitrary pattern using a neural network - page 8
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
Yes, I know maths within the school curriculum. I once asked a teacher in analytical geometry class (the one dealing with functions and coordinate axes), "If a function draws a curve on a graph, can you draw a function from a curve on a graph? " and got an unambiguous answer - "No. It's impossible." From this, I concluded that patterns can be described mathematically but cannot be identified because you cannot get the formula that generated them from the values.
Maybe there are some other mathematical tools. Tell me if you know them.To be precise and unambiguous, no. But in general there are means to get a function from arbitrary data. It's called "approximation". In addition to the fact that in general any function can be fitted in the middle of data, there are special functions that allow you to almost completely replicate the original data - polynomials, series...
By the way, methods of calculation of any function via Taylor and McLaren series have existed for a long time.
Yes, I know maths within the school curriculum. I once asked a teacher in analytical geometry class (the one dealing with functions and coordinate axes), "If a function draws a curve on a graph, can you draw a function from a curve on a graph?" and got an unambiguous answer - "No. It's impossible." From this, I concluded that patterns can be described mathematically but cannot be identified because you cannot get the formula which generated them from their values.
Maybe there are some other mathematical tools. Tell me if you know any.To be precise and unambiguous, no. But in general there are means to get a function from arbitrary data. It's called "approximation". In addition to the fact that in general any function can be fitted in the middle of data, there are special functions that allow you to almost completely replicate the original data - polynomials, series...
By the way, methods of calculation of any function via Taylor and McLaren series have existed for a long time.
By the way, it's an MO problem too, I'll have to try to sketch something like that sometime, it's an interesting problem.
2Peter Konow: I'd advise you to get a handle on the MO at least up to a point so as not to cause so much fun. Anyone who can read and do math can understand the basics of MO, you just have to put in the effort.
Кеша Рутов:
...
2 Retag Konow: I would advise you to at least get to the bottom of the MO so as not to amuse people too much, anyone who can read and do math can understand the basics of MO, just take some effort.
Well, that's how I do it. There's a lot of laughing, crossing, and slapping on the forehead, but no answers... Making a smart face and laughing is not an answer.
In a field I understand, I always explain, and only then send you to read.
1. Which data is more suitable for pattern recognition - OCHL or pictures? Is there or is there no difference?
2. Why isn't NS used for pattern recognition in algotrading for a long time now? Where are the functions like "bool Head_n_showlders()" working on the NS and fixing the pattern found?
Do you know?
Well, that's how I figure it out. There's a lot of laughing and crossing and forehead slapping, but no answers... Making a smart face and laughing is not an answer.
In a field I understand, I always explain, and only then send you to read.
1. Which data is more suitable for pattern recognition - OCHL or pictures? Is there or is there no difference?
2. Why isn't NS used for pattern recognition in algotrading for a long time now? Where are the functions like "bool Head_n_showlders()" working on the NS and fixing the pattern found?
Do you know?
1. It depends on what kind of patterns you want to find. Preferably, the format of the data fed for analysis should be as close as possible to the intended set of patterns.
For example, to describe a "head-shoulders" pattern, it is better to use a zigzag indicator, or rather several indicators, only with different parameters:
You can see that the pattern description is very similar to the Zigzag indicator. But the patterns can be of completely different nature, e.g. through different indicators, their combination, their transformed values, etc.
2. As written above, there are more appropriate methods to look for patterns.
The further away the patterns are from the data itself, the harder it is to find them.
Well, that's how I figure it out. There's a lot of laughing and crossing and forehead slapping, but no answers... Making a smart face and laughing is not an answer.
In a field I understand, I always explain, and only then send you to read.
1. Which data is more suitable for pattern recognition - OCHL or pictures? Is there or is there no difference?
2. Why isn't NS used for pattern recognition in algotrading for a long time now? Where are the functions like "bool Head_n_showlders()" working on the NS and fixing the pattern found?
Do you know?
1 Certainly OCHL, but pre-processed e.g. with wavelets, at the worst a bunch of indicators such as Momentum or Stochastic with geometrically increasing period.
2) price patterns are easier to search "directly" with a sliding convolution (multiply and add) with a reference; NS are not very useful here. We need NS and MO when it is not quite clear how to make an exit from the input, and MO itself creates something, a kind of model, but unfortunately within rather narrow boundaries.
PS And in general the "head-shoulders" and so on one chart is nothing, it is easy to find them, but they do not predict anything.
1 Certainly OCHL, but pre-processed e.g. by wavelets, at least a bunch of indicators like Momentum or Stochastic with geometrically increasing period.
2) price patterns are easier to search "directly" with a sliding convolution (multiply and add) with a reference, NA is not very necessary here. We need NS and MO when it is not quite clear how to make an exit from the input, and the MO itself creates something, a kind of model, but unfortunately within rather narrow limits.
PS and in general the "head-shoulders" patterns on one chart are nothing, they are easy to find, but they do not predict anything.
1. It depends on what kind of patterns you want to find. Preferably the format of the data you want to analyse should be as close as possible to the intended set of patterns.
For example, to describe a "head-shoulders" pattern, it is better to use a zigzag indicator, or rather several indicators, only with different parameters:
You can see that the pattern description is very similar to the Zigzag indicator. But the patterns can be of completely different nature, e.g. through different indicators, their combination, their transformed values, etc.
2. As written above, there are more appropriate methods to look for patterns.
The further away the patterns are from the data itself, the harder it is to find them.
I see. Thank you. I'll look into it.
In fact. You first ask yourself how can I use the answer I get from NS? Does it need to be transformed to make a decision etc. And only then you will realize that it's not the pattern itself that is important, but the reaction of the market to its appearance. And it's this reaction that should be traded. Otherwise you get a mess, not a TS.
When you have some substance in your statements, I will seek it out then. If you are unable to understand something, it does not mean that it is heresy.
And where you have gone - very well seen - what wild conclusions.
A neural network can be trained to classify and catalogue images, but on the basis of what criterion of probability of correlation of this or that image to a certain category of the catalogue it will happen, that is the probable error of any neural network.
Already memorized images with time will come to screening out with worsening results of reaction on them, thus success of any neural network will consist of a momentary content of that very catalogue of images on its memory, which are effective at that very second. Its extended base would be a minus rather than a plus.