How to form the input values for the NS correctly. - page 21
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
Here's a turkey.
It's too dreamy to be like real life )))) And we don't live in a fairy tale.....
And here's the turkey.
What's the point of adding anything else to this turkey? He's cute on his own. :)
What's the point of adding anything else to this turkey? He's cute on his own. :)
It redraws. We take it to give the signals that the network will choose.
Then we take these signals and forget about this indicator. Then we will need to decide what to describe the inputs given by NS. Choose a range for entering (number of bars and points that are next to the signal, but we will also take them as a training ones).
The same can be done with Klot's indicator (which has GA).
>>The formula for what?
The formula for the target function, if we were talking about an optimisation procedure. But it follows from LeoV's information that this function (Optimal Buy/Hold/Sell ) does not do any optimization, i.e. it has nothing to do with NS. It is quite an ordinary function that really looks into the future. And it absolutely does not care if to do it by Close or by a smoothing indicator. If you compare it to 2ZZ, I still tend to think it will give too many inputs. Or too few, depending on how one interprets on bars prior to ... . In realtime, these inputs will correspond to entries at the opening of each new bar, a point I have already criticized before :).
I'm going to cut in here, even though I am an amateur in neural networks. It seems to me that the target function (TF) must necessarily reflect the probability distribution of the value processed by the neural network. The sum of squares of errors is the best known and most standard form of TF, but this function is only as good as possible for a Gaussian value.
There is a relationship between the probability distribution and the error maximally minimised according to the maximum likelihood function. For a Gaussian distribution this is the square of the difference of the magnitudes, for an exponential distribution it is the modulus of the difference, etc.
We take the most relentlessly overdrawn indicator in the NS, teach the NS to give the right signals. Naturally, we have to get the result
so that our inputs will be just perfect, or nearly so... These values should be taken as a teacher for the NS. The advantage here is that we feed vectors BUY/SELL which the network itself has chosen as optimal ones. But a set of vectors Hold must be trimmed manually. Just to make sure that the sample does not consist of 90 % of Hold vectors and only 5 % on Buy/Sell...
Why take indicators? There is an interesting tool in metastock - Maximum Profit System (MPS), designed to compare the profitability of systems. The MPS is supposed to calculate all probable trades with a positive result. It is very convenient to build training arrays for MLPs on its basis.
It is redrawn. We take it to give the network signals of its own choosing.
Then we take these signals and forget about this indicator. Then we will need to decide what to describe the inputs given by the NS. Choose a range for entering (number of bars and points that are next to the signal, but we will also take them as a training ones).
The same can be done with the Klot's indicator (which has GA).
Unfortunately, you should understand that it is an unreal task to select inputs for the output (not even for the output, but for the buy/sell signals). Fantastic.
You do not know what will happen to these signals in the future? Will they run as correctly in the future as in the training and can you correctly select the inputs that in the future will give the correct information for the opening of the correct entries (as in the training)? On the other hand, will these inputs be opened correctly in the future, so that the outputs can be selected for them? More questions than answers.....
P.S. You won't confuse yourself - nobody will )))))
Unfortunately, you must realise that it is an unrealistic task to match the output (not even the output, but the buy/sell signals) to the inputs. Fantastic.
You do not know what will happen to these signals in the future? Will they run as correctly in the future as in the training and can you correctly select the inputs that in the future will provide the correct information for the opening of the correct entries (as in the training)? On the other hand, will these inputs be opened correctly in the future, so that the outputs can be selected for them? More questions than answers.....
P.S. You won't confuse yourself - no one will )))))
There are always a lot of questions in our business, and the answers are either relative or very few. So it's not surprising.
The selection of inputs is done by statistical analysis of different samples. The sample should be broken down as follows, to be more exact, we should find such samples:
Which contain either Sell/Hold or Buy/Hold. Of course, all three classes may be used, but any of Buy/Sell classes should be minimized.
Thus we should get 3 sets of vectors where the Buy. Buy and Sell (if someone finds sets of vectors where all 3 classes do not intersect each other according to the given signals, we won't need the NS). Exactly these sets are then fed to the inputs for training. Again, vector values should be relative (MACD will do, though even its maximum may change). Then preprocessing of input data, etc.
Of course if our NS will give signals, it doesn't mean there will be a Sell after Buy, but there are many systems that can help in the safety ...
2 rip Thanks! If I understand it correctly this is what we need, though I haven't looked at it yet.
The selection of inputs is done by statistical analysis of different samples. The sample should be broken down in the following way, to be more exact, we should find such samples:
Which contain either Sell/Hold, or Buy/Hold, of course all three, but any of Buy/Sell classes should contain the minimum of them.
Thus we should get 3 sets of vectors where the Buy. Buy and Sell (if someone finds sets of vectors where all 3 classes do not intersect each other according to the given signals, we won't need the NS). Exactly these sets are then fed to the inputs for training. Again, vector values should be relative (MACD will do, though even its maximum may change). Then preprocessing of input data, etc.
Of course, if the NS will give signals, it does not mean that there will be a Sell after Buy, but there are many systems that can help to hedge.
Do you not allow the idea that this could be a flawed theory?
Do you not allow for the idea that this might be a flawed theory?
It's not a theory, it's just one of the ways that I think can lead to good (profitable) results.
Of course I do. Can you prove that it is wrong, prove it, I will be very grateful to you for the time you save!