What to feed to the input of the neural network? Your ideas... - page 33
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
If you're doing all of this in Every Tick mode, I don't recommend continuing.
You know very well that in this mode tick values are modelled (generated) according to certain laws.
And any middle-class Expert Advisor, through optimisation, will be able to find such combinations of input parameters that you can get unrealistic results.
And it is pointless to spend time on this.
Did you say what is the maximum drawdown on funds?
I test at opening prices.
This is enough for a system that works only with closed candlesticks.
As I mentioned above - I have no idea how to work with real ticks, what to do with them.
Maximum drawdown on funds - I do not look at numbers, I look at the green line of the report chart. It is not critical there.
I'm testing at opening prices.
This is enough for a system that works only with closed candlesticks.
As I mentioned above - I have no idea how to work with real ticks, what to do with them.
Maximum drawdown on funds - I don't look at numbers, I look at the green line of the report chart. It is not critical there.
How is it not critical???
It is obvious that for 10 000 you use a fixed lot, probably 0.01 and you get a gain of less than 10% for 3 years.
With such a lot you don't even have to use a stop loss.
How is it not critical???
I can see that for 10,000 you use a fixed lot, probably 0.01 and get a gain of less than 10% over 3 years.
And why?
Now it is the task to start the machine so that it gives a stable upwards.
By stable upwards I mean either single training (optimisation), after which the network always works, or periodic additional training (optimisation) after a certain period with further workability.
At the moment neither of the two directions is implemented.
Issues of drawdown are secondary.
Drawdowns, risks, lots, mani-management, all this is secondary. The car is not driving yet, there is no point in polishing it.
is this neural network stamping news in the news?? it appeared literally 3 days ago..... It's just that in zen, the news is in order on the home page, the sources are different!!!
A little more, and the realisation will come that neurons are not needed at all)
Hmmm.
Bold, sure.
I'd say the truth is somewhere in the middle. Neurons can fulfil a function, basically they are - a big such function. A function and inside it little functors, which also multiply and sum up.
In the end, there is some set of weights and there is some activation function, which together with the right inputs will give something that works on all currencies and for quite a long time.
This is all reasoning, of course, but given the results above on the branch - yes, sometimes it seems so too.
UPD
Activation feature. It's been bugging me too. Why the tangent? Why a sigmoid?
Why not a curved line? Sine, cosine, a set of exp(x) multiplied by each other, which eventually give all sorts of squiggles on the activation graph. These are the rules of the trade. When the input is 0.7 and the output is 0.8, and 0.7000001 - then the output can be a sharp jump.
And such strange complex activation functions together create infinite rules for the function output.
I tried sigmoid and tangent - typical results, almost the same.
But as soon as I created a curved arc by multiplying exp(x) by each other 150 times, when optimising the same(!) architecture, the same(!) inputs - the optimiser started to overtrain. That is, to fit the trade to the price chart very, very well. Thus modification of the activation function essentially improved the neural network in the task of path memorisation. With the same initial data.
That is, playing with the activation function is another infinite field of possibilities for training (optimisation).
And yesterday I also had a thought that suddenly a neural network (one) should not be responsible for two or more signals (buy, sell, wait, exit, etc.). Maybe for the exit we need a separate NS, which is in no way connected with the NS that "enters".
The same is the division into buy and sell. One NS for buy, the second for closing buy, the third for sell, the fourth for closing sell.
In general, I'll play around too
.
The same is the division into buy and sell. One NC for buy, second for close buy, third for sell, fourth for close sell.
I don't know about forex, but stock markets are usually asymmetrical for ups and downs, and skewed, so it is quite natural to consider buy and sell separately.
The same is true for exiting a position, but here it is also a good idea to consider a trailing stop, i.e. these are also separate algorithms.
Hmm.
Bold, of course.
I mean, neural networks in general).
Without the influence of news (or FA as anyone prefers), neural networks would work just fine. Very fine, in fact.)