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
As I was writing the post remembering the details of the implementation, there were some inaccuracies.
I looked at the code and it turns out that I was influencing FA on weights once when switching to a new forecast, i.e. not every epoch, but when it was necessary to retrain the network when new data came in.
One more thing. Correction of the weight is the product of the error from neuron output by FA derivative and by neuron output (amplitude, taking into account the sign) from which the signal is input.
Here's what it looks like for one perseptron with a non-linear output (for example):
Here epochs are numbered by the index L. I purposely show it in MathCad, it is more clear. In - number of inputs, x - input vector. The rest is seemingly clear.
I think, how to use BUY/SELL as input vector ?
For example, we take three or four indicators, preferably correlated ones, put them on the input and mark the price chart with something - like "buy here", "sell here"... but how should we do it? Another question: what if we mark differences (on each bar) between 200 values with step 2 on grid inputs, will it find something there?
I'm wondering how to make BUY/SELL as input ?
For example, we take three or four indices, preferably poorly correlated, set them on the input and mark the price chart somehow - like "buy here", "sell here"... but how to do it? Another question: if I want to enter the grid with differences (on each bar) between 200 values with step 2, will it find anything?
Oh, this thing called gradient method of optimization is complicated. I mean, algorithm implementation is not very complicated, but there may be some big problems with learning.
I know it's the holidays, but don't take it so hard on your chest:)
I do not understand you gentlemen.
So it is better to redirect the thoughts of SUBOMEGAMOUS brains to quantum analysis. Example - just decompose an ordinary graph into a quantum graph (i.e. not in real time, but in quantum time), we can ALREADY see the regression and dependence. It is possible to avoid sharp angles by using chaos theory, according to which even the most chaotic motion of particles can be combined into a system, but at different times (i.e. the system is defined in different time ranges on the basis of quantum physics, i.e. not in real time, but in quantum time). And it's enough to convert quantum time to real time and get the long-awaited SELL BUY.
>> I'm in!
I remember in one of the threads, someone called Prival asking fellow citizens to guess the number... that controls the market. I knew it was the Constant Plank. Just what a poor trader needs in the hour when the Forex Bar drops... and there comes Heisenberg Uncertainty on the DC server and only by the Doppler effect one can guess the trajectory of a deposit... just kidding -:)
I'm in!
...I knew right then that it was the Constant Plank. Just what the poor trader is missing in the hour when the Forex Plank drops... and Heisenberg Uncertainty sets in on the DC server and only the Doppler effect can suggest the trajectory of a deposit... joke -:)
+5
That's right, paralocus, screw them, obscurantists. Also I forgot to mention torsion fields. Strong thing for a sore head - it makes the market go round and round.
By the way!
In that article the perceptron searches exactly for Buy/Sell situations. It's just that I have some semantic inconsistencies with the transition to ORO.
I.e. the grid input is a clustered turkey... why is it there? And it is there to define the moment of market entering: OUT of the indicator is higher than zero - buy, lower than zero - sell.
The thing is that earlier the teacher was a geneticist who corrected weights according to the results of perceptron working on the training sample (in the tester). The geneticist used the profitability of the Expert Advisor that worked, in which signals for Buy/Sell were generated using OUT of the perceptron. Therefore, everything was working properly to get the maximum profit at the training section of the history. That is, by abandoning the geneticist, I abandoned the teacher for the grid. As a matter of fact, I'm only now making concrete sense of it. Remember, I asked you in the beginning if we can use the value of losses obtained in a transaction as a corrector? That's what we need to do.
Neutron, I also wanted to ask about Hebb's training (read by Wasserman). It seems that the formula for correction of weights is very simple there:
Wij(t+1) = Wij(t) + [OUTi(t) - OUTi(t-1)]*[OUTj(t) - OUTj(t-1)] and no gradient drops. Is this going to work?
I haven't messed around with Hebb-m. I've had enough of what I've got. I, in general, am a supporter of simplifying a problem as much as possible, practically to the level, until I myself begin to understand what to what:-) From this point of view, usual two-layer NS is a universal approximator, capable to solve almost any problem of extrapolation of input data, it is proved strictly. So why all the frills, if I have enough means and power to retrain this magic and simple Grid on each of my transactions? That's right, no reason! As for Wasserman, if he says it's right, then it's right! Of course it will work.
You're solving the problem of optimal NS input now. Of course, you can just feed it with various indukes in the hope that the grid will decide what is better for it... But it's better to think "what is the optimal TS in the market? Maybe you should predict its moments?
Read this work. Of course, there are glitches, but they are not fundamental:
I haven't messed around with Hebb-m. I've had enough of what I've got. I, in general, am a supporter of simplifying a problem as much as possible, practically to the level, until I myself begin to understand what to what:-) From this point of view, usual two-layer NS is a universal approximator, capable to solve almost any problem of extrapolation of input data, it is proved strictly. So why all the frills, if I have enough means and power to retrain this magic and simple Grid on each of my transactions? That's right, no reason! As for Wasserman, if he says it's right, then it's right! Of course it will work.
You're solving the problem of optimal NS input now. Of course, you can just feed it with various indukes in the hope that the grid will decide what is better for it... But it's better to think "what is the optimal TS in the market? Maybe you should predict its moments?
Read this work. Of course there are some glitches, but they are not principal:
Yes, thank you, I'm reading it now, I'm really stuck.