Repetitive patterns and other patterns - page 2

[Denis Lazarev]

In this case the ZigZag indicator with a certain period can be very useful,

1) we take the distance between two peaks, such indicators allow you to abstract from the market noise

2) then we compare these values by dividing them by each other and deduce the ratios (the first straight line is twice as high as the following one and represents the amount of the trend, etc.)

3) then write down what exactly should be after this "pattern" and start tracing

the pluses are that the relative values allow to neutralize the market volatility factor, regardless of the figure size, we know for sure there will be a fall for some time

The zigzag indicator would work because it doesn't take into account noise, the fractal indicator would also work

[Vladimir Kustikov]

gpwr:

Can anyone program this interesting thing?

The topic is interesting, but too complex to automate.

Firstly, despite the beauty of the drawings you have published, there are a number of inaccuracies in the graphical construction, for example the last red channel is clearly fitted to the "breakdown usually becomes the middle of a new channel" principle, as its bottom line should have been drawn higher.

The second tricky point is false breakdowns. How can we tell if it is a breakout or a trendline correction? The history shows it very well but there is no such thing in practice.

I myself have been interested in trading in channels for a long time and I am interested in any thoughts on this subject.

[Rustamzhan Salidzhanov]

There is another problem - repeatability and fractality, and although these things should seem to help in analysis, you never really know which channel in the hierarchy you are currently working with (where we are now). Here is an example - and although I drew the channels later (further in time), they all perfectly coincide with history, i.e., in the future the price will move according to the old channels. But which one? And the picture is only a small part of what can be drawn.

[Vladimir]

So, the prediction from the previous page worked out quite accurately. The red pattern is almost over. Now we expect a break upwards.

[Vladimir]

Vladix:

... the last red channel is clearly fitted to the "breakout usually becomes the middle of a new channel" principle, as its bottom line should have been drawn higher.

... How can we tell if this is a breakout or a trendline correction?

Agreed. The lower boundary of the last downward trendline could have been drawn higher as well. It would not make much difference, because we are interested in the touch points and in the break-down of the upper limit of the down trend. The downside boundary usually starts from the third point of the previous uprend touching the upper boundary of the previous channel (the green 3).

It is better not to wait for the second point of touching the upper border of the down trend, but to set an alert at the price of the break in the previous channel (yellow circle with a black border) and to aim at least at half the width of the previous channel. Or you can follow the price movement with a wristwatch. False-breaks do exist, which is why you should always use stops. Let's think about other entry options.

[secret]

gpwr: For example, the two red patterns look the same to the eye, but the second pattern is almost twice as long as the first. The movement along the price axis is also different in amplitude. If we try to take the vertices of the zigzag, one pattern will have more vertices than the other and the net will get jammed. Thoughts?

Do you want channels or patterns? If patterns, then: bring time and amplitude to relative values (normalisation). Zigzags are not needed at all, except for shallow, or better, shallow approximations. Then plug it into any traditional image recognizer, even a two-dimensional correlator. Since the windows are different, you'll have to search through a certain size range, and you'll constantly be wondering if "this is still the previous pattern" or "the next one has already started". All imho. It will have to be tested. I want to make it myself, but have no time.

[Vladimir]

bas:

So do you want channels or do you want patterns?

A pattern within a channel - more on that later. The price pattern itself is meaningless if it is not attached to a channel or support/resistance levels. That is the main drawback of users of networks - they take prices not bound to levels thinking the network will find them. Moreover, the neural network users send a constant part of history to the input of the network thinking the network will find the patterns linearly or non-linearly compressed or stretched in time by itself. Yes, such a network exists, in our brains. Artificial networks are not yet up to it, even dialers, which learn from labeled examples of roughly the same size.

[secret]

>Hence the main drawback of network users - they take prices not linked to any levels, thinking that the network itself will figure it out and find those levels.

I think it will, if you give it that task in a separate way. It's just that the task is usually set differently.

>Besides, neural networkers feed a continuous part of history to the input of the network thinking that the network itself will find the patterns linearly or non-linearly compressed or stretched in time.

Linear distortions are minimized by normalization, nonlinear distortions can be minimized by going to range representation, but it will make momentum/correction indistinguishable, if that is important.

[Vladimir]

The price touched the lower boundary of the current channel, confirming that the boundary was set correctly after all. It is very likely that this down-trend pattern is over and we have started an up-trend, which will be confirmed by a breakout of the upper boundary.

[Vladimir]

bas:

>Hence the main drawback of network users - they take prices not linked to any levels, thinking that the network itself will figure it out and find those levels.

I think it will, if you give it that task in a separate way. It's just that the task is usually set differently.

>Besides, neural networkers feed a continuous part of history to the input of the network thinking that the network itself will find the patterns linearly or non-linearly compressed or stretched in time.

Linear distortions are minimized by normalization, nonlinear distortions can be minimized by going to range representation, but it will make momentum/correction indistinguishable, if that is important.

So we both agree that the network itself is not capable of detecting structures such as channels, we need to set the task accordingly, i.e. prompt it. Note that I drew the bottom line of the last channel before price touched it a second time. You could draw an infinite number of straight lines from the first touch point, but it was the line parallel to the upper limit that turned out to be the lower one.

As proof of the pattern distortions, I stretched the previous downward pattern and overlaid it on the current one:

The problem with automatically finding similar patterns is that we don't know in advance

1. The length of the current pattern. For example, in this discussion the pattern is bound to the channels and its start is more or less defined. If we don't tie the pattern to a channel, where do we get the start?
2. The distortion of the current pattern compared to past patterns. We need to describe the current pattern by some smooth function, compress it by one bar, compare it to all previous patterns, then compress it again by one bar, and so on. Then stretch it by bars. In short, it will take a lot of time. Or create a network working on the principle of our visual cortex.