Interested in the experience of developing robots on neural networks. What data do you transmit to the network, on what history and timeframe do you conduct training ?
At the moment, I implemented it in a simple way: we take n-the number of bars and pass it to the grid at the output we get prediction coefficients from 0 to 1 for buying and also for selling
The training was conducted from 2012-2018 on different timeframes risk per trade 2% stop 25/4.
here’s what happened:
greshnik1
Are you sure these are patterns)? Yes, I remember.Of course, it is better to automate the search.Look to the sideVolking – forward optimization. It relies on the fact that patterns must be constant over time. Or, at least, having a small linear “movement” in the space of neural network parameters(in your situation). I hope that I wrote it clearly.
tormovies
Yes, I would like some minimal testing outside the training period
because if you test on the learned period will be just a super fit, and so would have taught the network for the entire period minus 6 months and then separately launched for 6 months, at least some would be an approximate result and then such a fit I also did but there was no sense
ara66676
The fact that a properly assembled NS will find and remember patterns that a person does not see is a fact. The fact that patterns are worked out like all noticeable 50/50 is also a fact. All you did was just find more signals to trade. If you really want to continue, then you need to look a little in a different direction… is input to the neural network such as dates of birth of artists and find a pattern in the development of pricing pair month. That’s when your grid will be worth something. There was a post about decision trees, so this algorithm just makes everything much faster and with much more data. That is, on the trees you collect even more patterns for yourself, but no one guarantees you to work out.
pvitaly
Do you suggest manually searching for patterns ? So after training, she just found them. Maybe in history she just remembered when to enter)?
greshnik1
Vitaly Palkin:
Interested in the experience of developing robots on neural networks. What data do you transmit to the network, on what history and timeframe do you conduct training ?
At the moment, I implemented it in a simple way: we take n-the number of bars and pass it to the grid at the output we get prediction coefficients from 0 to 1 for buying and also for selling
The training was conducted from 2012-2018 on different timeframes risk per trade 2% stop 25/4.
here’s what happened:
This kind of method is not working. There is a lot of “noise”, the network with this approach will not be able to distinguish stable patterns from it.
igoreremenko
Vitaly Palkin:
Interested in the experience of developing robots on neural networks. What data do you transmit to the network, on what history and timeframe do you conduct training ?
At the moment, I implemented it in a simple way: we take n-the number of bars and pass it to the grid at the output we get prediction coefficients from 0 to 1 for buying and also for selling
The training was conducted from 2012-2018 on different timeframes risk per trade 2% stop 25/4.
here’s what happened:
you can become a virtual millionaire in the tester without neural networks, in MT4 you can become a billionaire in General)
this topic is raised on this forum constantly, the last 10 years, something amazing (according to the results) so far no one has shown publicly, maybe you will succeed…
pvitaly
The EA is built on the opening of the 30m bar. ticks are not taken into account
pvitaly
And the meaning is that training only on bars worked with a Bang)
On a real account, the robot is already running, you are interested in who uses what parameters and what training methods are used for neural networks?
-aleks-
Vitaly Palkin:
Interested in the experience of developing robots on neural networks. What data do you transmit to the network, on what history and timeframe do you conduct training ?
At the moment, I implemented it in a simple way: we take n-the number of bars and pass it to the grid at the output we get prediction coefficients from 0 to 1 for buying and also for selling
The training was conducted from 2012-2018 on different timeframes risk per trade 2% stop 25/4.
here’s what happened:
So what’s the point of showing results on training data? Show what went beyond training.