Neural network - page 8

 
storm >> :
Personally, I have my own networks (:

You say A, you say B, but you're smiling mysteriously :)

What is the cardinal difference between your networks?

 
Look, to predict the time series we can use the difference of, say, the opening and closing price! And then the corridor will be smaller and will not jump as much!
 
xweblanser писал(а) >>
Look, to predict the time series you can use the difference of, say, the opening and closing price! And then the corridor will be smaller and will not bounce so much!

It's not that simple...

 

Question for network experts

Suppose there is some network that ideally should give a 1 to buy (or let's say >0.7 to buy), (-1) to sell, the rest we wait for. There are a number of indicator nets inputs. Some of inputs is the indicator that at crossing 0 (i.e. at changing of sign from negative to positive) kind of gives a signal to buy. That is, the maximum value of the buy signal by this indicator is exactly at the moment of crossing from 0 (further the signal remains, but the potential profit decreases).

Now - a neuronet, it is roughly a function of the sum of products of inputs and weights (well, inner layer neurons are also considered). If we consider the formula f=F(w1x1+w2*x2+...) then if x1=0, then regardless of other inputs and activation function at the moment this input is simply excluded from the final output. It turns out that the signal will simply be ignored.

This situation reminds me of a sort of real case (from Wikipedia) - It is known a case where a network was trained to recognize images of tanks on photos, but later it turned out that all the tanks were photographed against the same background. As a result, the network "learned" to recognize this type of landscape, instead of "learning" to recognize tanks.

So, the actual question is. Does it make sense in this case to transform the value of that input in such a way (though another question - how), that the maximum value of the buy signal by such an indicator is not at crossing 0, but when, say, this indicator = 1.

For example, we can split this indicator into two:
- The first (of the 1-x type) shows the degree of approach to zero.
- The second is binary - just the sign of this difference (+1, -1).

Does this manipulation have any fundamental importance for the network?

 
If the signal is a crossing with zero from the bottom upwards (and for sure there is also a crossing in the opposite direction, which gives an opposite signal), then the 0 itself can not be a significant value for the entry. The signal is just the signal: code them as +1 buy, -1 sell, irrespective of how it is displayed on the chart of a particular indicator. And in general, zeros should not be used - the values must be symmetrical in order to use the power of the net weights as much as possible. In the case of indicators, in which signals are given when crossing zero, you can take the derivative of their output (not in analytical form, of course).
 
GrooovE писал(а) >>

For example, it is possible to split this indictor into two:
- The first (of the 1-x type) shows the degree of approximation to zero.
- The second is binary - just a sign of this difference (+1, -1).

Does this manipulation make any difference to the network?

For a neural network, the first option is more informative than the second one.....

 
Urain >> :

If you say A, say B, then you smile mysteriously :)

What's the fundamental difference between your networks?


What's the cardinal difference ... It is highly specialised (recognition of combinations of waves, fractals), and thus the simplest in execution. For example, weights are simply picked up by the tester (as in Reshetov's perceptron), at the same time my perceptron, with the same range of input parameters, is able to memorize a particular pattern, which Reshetov's perceptron cannot. To the credit of Reshetov's perceptron, his design perfectly finds a flat, which in skilful hands can and probably does bring profit.

 
LeoV >> :

For a neural network, the first option is more informative than the second one.....

And if we compare (under the above conditions, of course) the original signal and the derivative, the choice is the derivative?

And if one extends the situation to crossing some threshold by specifying an offset relative to zero, should the "threshold" signals be amplified in this way...?

 
The coding of the signal must be chosen by the trader, based on its meaning. In particular, if the grid will have outputs trained virtually on probability (x and 1-x), then no derivatives are needed. If the outputs are binary (buy/no-buy/sell/no-sell), then a single-valued signal is needed. But whether it is necessary to calculate the derivative - dI/dt on some number of bars - depends on the specific indicator. In particular, crossing zero in each direction, imho, it is easier to mark +1 and -1, as I suggested. Regarding the threshold - the question is general - it is valid not only in the context of grids, but in general. If the system implies working off the threshold, you should use it. You can train the grid itself to pick up the threshold.
 
marketeer >> :
The coding of the signal must be chosen by the trader, based on its meaning.

>> I agree. The idea of the TS should be present. NS is a tool only. Therefore, it is paramount to choose signals to enter and be aware of what we want to get in the output.