Machine learning in trading: theory, models, practice and algo-trading - page 132

 
Alexey Burnakov:
What else can you say? Are you waiting for ready results?

What can I say? If I've never worked with a convolutional network. I'm not Sanych....

All I can say is that it's interesting...

===================

Youri Tarshecki and where did your post disappear to?

 
Youri Tarshecki:
Why such a need? If a pattern has an analogy in the history, then it should also correspond to the duration. At least I was looking for commensurate sections when I did a pattern search.

sss

I'm working on the following concept: from the current pattern "B" in the history we look for a pattern analogous to "A". With the dtw algorithm we look for similarity...

The sad thing is that we don't know what size "B" as well as "A" may end up in our search and that's a lot of headache,

The sad thing is that we don't know what size "A" and "A" might end up being, so it's a lot of headache, so besides the search we need to dynamically shrink/expand these patterns...

If anyone has any ideas how to make such search as effective as possible, I'd be very interested to hear them...

 
mytarmailS:


The sad thing is that we do not know what size the search may end up as "B" or "A". And because of this a lot of headache,

If anyone has any ideas how to maximize this search with interest will listen ...

So there is no size, there is a pattern and we should look for a pattern.

Personally, I use ratios and sequence of extrema for this purpose.

Although I also use the duration of an event, but just not in the unique sense, but on the contrary, in the average sense. I.e. if the duration is longer than the average, I increase the chance of occurrence, and vice versa.

But it's so to speak learning from nearest patterns, not searching through the history.

 
Youri Tarshecki:
So there is no size here, there is a pattern and we need to look for a pattern.

How not? I don't understand your point...

The dimensions are there, but we will know them only when we find patterns "A" and "B" and if we take the picture above as an example, it turns out that "B" will consist of, say, 13 candles and "A" consists of 53 candles

 
mytarmailS:

How not? I don't understand your point...

The sizes are there, but we recognize them only when we find patterns "A" and "B" and if we take the above picture as an example, it will turn out that "B" consists of, say, 13 candlesticks, and "A" consists of 53 candlesticks.

That's why I simply correct for the volatility. Higher volatility - more expectations from the average. And vice versa.

But the pattern itself is a certain sequence of extrema (as I perceive it). That's if to put it in a nutshell.

I've experimented a lot with such sequences in my time and concluded that only the simplest laws work, but we must consider levels and the duration of levels existence, volatility and correlation at the same time. Then something just starts to work.

In your example even if you reliably learn to identify these patterns, the price will not necessarily go in the direction of the dotted line, because the pattern is too complex (although it's easy enough to catch it)!

(the last two is just dividing by two, not an extremum, you can just take some coefficient instead -))

 
Alexey Burnakov:
What's the point of your words, passenger? You don't want to try, or your own way. I'm working on my own task and I'm interested.
You yourself are a passenger. I didn't ask to be your traveling companion. You want to get free testers?
 
Alexey Burnakov:

Who has tried? My colleagues and I want to train a convolutional NS. There's some mapping going on. We hope.

For this purpose, LSTM would be better suited.

 
Youri Tarshecki:

That's why I just adjust for volatility. Higher volatility means we expect more from the average. And vice versa.

But the pattern itself is a certain sequence of extrema (as I perceive it). That is if to put it in a nutshell.

I've experimented a lot with such sequences in my time and concluded that only the simplest laws work, but we must consider levels and the duration of levels existence, volatility and correlation at the same time. Then something just starts to work.

In your example even if you reliably learn to identify these patterns, the price will not necessarily go in the direction of the dotted line, because the pattern is too complex (although it's easy enough to catch it)!

(the last two is just to divide by two, not an extremum, you can just take some coefficient instead -))

I see your point.... but all your model easily breaks, for example, some random zigzag wave inside any of the waves 1...5, the human eye will ignore it, dtw too and thus will save the image, but your algorithm will immediately make something else out of "head and shoulders"...

p.s. I'm already slowly giving up on dtw because it's not living up to my expectations

 
mytarmailS:

But your algorithm will immediately make something else out of "head and shoulders"...

So that was the "pathos" of my post. -) It's us humans who see some symbols and signs. But if we formalize them into simple rules, they are

1. For the most part will disappear in our perception as recognizable patterns.

2. It will turn out that even if we can detect them more or less accurately, it will do absolutely nothing, because complex patterns don't work the way people expect them to work. Belief in patterns is a phenomenon of human psychology. One of the myths of the market.

This, by the way, does not remove the problem of finding specific chart properties on a particular section at all - For example, it can be useful for optimization during a jumpshare.

 
Youri Tarshecki:

Belief in patterns is a phenomenon of human psychology. One of the myths of the market.

The standard indicator "Bill Williams' Fractals" in MT is also a search for a certain pattern, and without dtw, but just by bars. It worked very well once, until it was lost because of its popularity (on some symbols on D1 it can still be used, the profit is minimal, though).

But trading strategy with this indicator is more complicated than "buy/sell on 1 bar". It uses pending, tp and sl, so apart from looking for patterns we need to look for trading strategies to which they are applicable.