What to feed to the input of the neural network? Your ideas... - page 74

New comment
 
mytarmailS #:
Google it
I can't find it.
 
osmo1717 #:
I can't find it.
It's sad.
 
mytarmailS #:
That's sad.
Well, explain it to me.
 
osmo1717 #:
So explain.

Explained.

mytarmailS #:
When a neural network is trained, its weights are unchanged. Only the input data changes.
 
osmo1717 #:
Well, explain it to me.

watch this


 
osmo1717 #:
So explain.
These are coefficients of significance of parameter valuesThere are articles like neural networks it's simple. Read articles about networks here. Perceptron single-layer type..... )


 
Thanks for the explanations everyone!
 
Hello.

I'll put my five cents on the subject.
Do not judge strictly.

1. Use the "teaching by examples" methodology.
The number of training examples should be limited, not the entire foreseeable history.
Otherwise, the "brains" of the network will be a mess, while the answers of the trained network should correlate well with the known correct answers at least for the examples it was trained with.
Actually, this is the purpose of training any network, when we are talking about training on some obviously available data.

2. Training examples are prepared from a certain limited number of h histories by the "crawling window" method having the size of a certain number of k values.
In other words, the number of training examples (each of size k) is equal to h, where k and h are preset.
Since the number of training histories is limited, when a new history appears on the right and an old history disappears on the left, the contents of the training packet also changes, which means that the network must be retrained completely anew.
In this way, the network never remembers too old and is regularly retrained on the new.

3. As a rule of thumb, when learning from examples, each training example must be accompanied by a known correct answer.
In the context of an antipodean system (buy or sell), the value of the known correct answer should be between -1 and +1.
Where with a positive value of the answer a learning example for buying is provided, and with a negative value of the answer a learning example for selling is provided.
What exactly should act as a known correct answer, especially in the range from -1 to +1, perhaps, I will not spread now.
However, I will repeat that after training, the network's answers should correlate well with the known correct answers, at least when the network is presented with the examples it was trained with.
Checking the presence of such a correlation (network test) is a mandatory procedure, otherwise it is impossible to state that the network has learnt something adequate and not "pissed off".

4. As an image of a training example, we can use a certain number of k values of closing and opening price differences normalised to the range from -1 to +1, relative to the maximum (modulo) of these k values.
Such a training example contains positive and negative numbers, which have every right to be of the opposite sign as a similar, but mirrored, example.
Based on the considerations of mirroring, those signs that are suitable for purchase are not suitable for sale, but will be suitable for it being in a mirror form, and vice versa.
The concept of mirroring suggests that the network should give answers in the range from -1 to +1, as antipodal, either selling or buying.
If the value of the network response is above the threshold, e.g. +0.8, then it is a buy, or if the value of the network response is below the threshold, e.g. -0.8, then it is a sell.

5. In addition to the price image, the image of volumes, even tick volumes, should not be neglected.
It is not important whether the volume is tick or real, it is only important that the volume scale has pronounced sessional and daily fluctuations, which means that it is not useless at least.
However, the volume values are unipolar and do not fall under the concept of mirroring, and one should think carefully about how they should be used in a network with a bipolar response.
For example, two situations, price goes up/volume goes up, price goes down/volume goes up, are mirror-like, but the mirror component here is only price, and volume remains itself.
But for example, the delta of volumes (if any) is fully consistent with the principle of mirroring.
By this point I just want to emphasise that not all data are antipodal in nature, but some of them should remain themselves (not mirrored), at least for buying, at least for selling.
However, the formula of error back propagation does not know about such peculiarities ...

 
Evgeny Shevtsov #:
Hello.

I'll put my five cents on the topic.
...

Thank you, substantive

 
Search Paradigm:

Target - not the next ordinal candle or the next ordinal reading.

Target is the result of a trade from the current time.

That is,

1) We take any strategy. We wait for its triggering (position opening). Record an input set (any). Monitor the position until it closes, and record the result of the trade as a target.

2) We train it using any method.

3) Run our (pomoichny) strategy

4) Add a filter to it - our NS.


As a result, we need to put a little intellectual, creative effort, patience and time to pick up the strategy.

The rest will be done by the neural network.

We remove from the neural network the hard intellectual labour simply because in the current realities the NS is not even able to determine the levels, if it does not mark them "by itself".

And it will do what it is supposed to do: to grind the TS.
1...676869707172737475767778798081...86
New comment