What to feed to the input of the neural network? Your ideas... - page 27
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
No super computer will help, just a minimal knowledge of the nature of markets.
A bee without a super computer flies thousands of flowers several kilometres away and carries the profit to its hive without error.))
No super computer will help, a minimal knowledge of the nature of markets is enough.
Abee without a super computer flies thousands of flowers several kilometres away and carries the profit to its hive without error.))
the bee had a head start of millions of years and billions of bees not flying to the flower or back to the hive.
the bee had a head start of millions of years and billions of bees that didn't make it to the flower or back to the hive.
I agree that millions of years and billions the bee has a head start on history. Humans may have just as many.))
===========
If you just look at the price chart it is not clear why it is reversing.
But if you look closely at OHLC, you start to notice certain regularities in the price behaviour.
From the symmetry of OHLC the behaviour changes and it is possible to predict future behaviour.
Here is an example.
The figures show variants of price behaviour in the future. The final version of the idea is not yet complete, but I think the essence is clear. In the picture only OHLC prices are usedwithout taking into account the NS.
Things have changed.
"Hi 😊 I'm looking for a way to create a trading advisor that can analyse requests via Telegram and automatically determine the parameters to trade. For example, if I write 'want to buy an iPhone' in Telegram, GPT will independently determine the appropriate parameters for the trade. Does anyone have experience in integrating Telegram with MQL4 or using artificial intelligence to analyse such queries? I would appreciate any information or advice! It's just an idea, thanks! 😄✨"
Now your query looks even more welcoming and friendly!
Just three entrances
in[1] = Close[1] - Close[2];
in[2] = Close[2] - Close[3];
in[3] = Close[3] - Close[4];
Scale optimisation for 2021
Forward almost 2 years: from 2022 to 2023-10-29
What's the catch. - It's one of hundreds or thousands of sets. You'll never find it amongst the flush sets.
What's the good news? - Neuronics can work on any data, as long as it reflects a little bit of what's happening on the chart
Same conditions: 3 inputs, only normalised to a range of -1 to 1.
Learning in 10 years
Forward almost 2 years: from 2022 to 2023-11-26
More trades already, though it looks unnatural against a more chaotic optimised period.
Still the same: you won't find a working set among hundreds and thousands of other fits, but once again confirms that it is possible to work in forex, looking only at the last three candlesticks.
Same conditions: 3 inputs, only normalised to a range of -1 to 1.
Learning in 10 years
Forward almost 2 years: from 2022 to 2023-11-26
More deals already, though it looks unnatural against a more chaotic optimised period.
The same problems: you won't find a working set among hundreds and thousands of other fits, but it once again confirms that you can work in forex, looking only at the last three candlesticks.
4 inputs
Optimisation since 2000 in 22 years
Forward almost 2 years: from 2022 to 2023-11-27
I'm going to try to solder RNN to this MLP. Maybe sets like this will show up more often on the list.
I encountered one effect: when optimising the weights of a regular MLP (2 layers of 3 neurons, and other architectures), the neuron opens only one type of positions: either buy or sell. That is, those positions that it considers necessary to make money.
Do you have any ideas on how to "force" it to study both bullish and bearish moods of the chart, and "guide" it gently?
I tried forcibly putting a flag on alternate opening (if now I opened a sit, then next time look for a buy and open, and so on in a circle). It didn't bring any results, and it's rude and clumsy.
I tried to add the SoftMax activation function to the output, for which I let 3 outputs: buy, sit and wait. Then I twisted it so that waiting played the role of position closing, so that the neuron (optimiser, that is) tried to somehow rearrange the weights in such a way that it was forced to open buy positions.
In the sets at the bottom of the optimiser you can find such settings where two kinds of positions are opened, but in the top - none. If you teach a year in which a bearish trend prevailed, it will persistently open only sel. It either skips decent corrections or settles on them.
I understand that neuronka itself is something like a mashka - it averages the weights so as to take the maximum. I mean, it takes the easy way out. But the hard way, it's going to make more money. And it's on the optimised period, you can take what you want.
But no, as soon as you force her to open two types of positions, she doesn't even want to be retrained.
Playing with the third state in SoftMax, the optimiser selected the weights in such a way that the SELL state just flirted with the HOLD state, and the BUY neuron was put on the BUY neuron in general.
Here, I'm curious, how to persuade the neuron to open in both directions during optimisation? I don't want to take a q-table or forcibly introduce penalties, or lower the weights leading to the SELL neuron by a coefficient.
There is either something more elegant, simple and clear.
In sets at the bottom of the optimiser you can find such settings where two kinds of positions are opened, but in the top - none. If you teach a year in which a bearish trend prevailed, it will persistently open only sel. It either skips decent corrections or settles on them.
I understand that neuronka itself is something like a mashka - it averages the weights so as to take the maximum. I mean, it takes the easy way out. But the hard way, it's going to make more money. And it's on the optimised period, take what you want.
But no, as soon as you force her to open two types of positions, she doesn't even want to be retrained.
Playing with the third state in SoftMax, the optimiser so selected weights that the SELL state just flirted with the HOLD state, and the BUY neuron was not even considered.
Here, I'm curious, how can I persuade the neuron to open in both directions during optimisation? I don't want to take a q-table or forcibly introduce penalties to reduce the weights leading to the SELL neuron by a coefficient
There is either something more elegant, simple and clear.
train two grids - one only in buy the other in sell.
switch on both :-)
then add a collision resolution network (or just an alg.) so that they don't trade in different directions at the same time.