Article: Price forecasting with neural networks - page 4

New comment
 

Good day to you all !
I joined this forum for the first time.

I am interested in the topic of using neuro-analysis for stock operations,

In particular, for FOREX.

Can I ask you a question? I must clarify at once that I am not a programmer, but just a beginner trader...

A neural network can predict the price movement on what time intervals?
Only within a trading day or beyond it as well?

If it is possible, if you want, try to check these predictions on the EUR/USD chart in the next week.

How will this indicator behave during the sharp price movements?

Suppose with a certain probability that the pair's exchange rate from 1. 3465 in 1-2 days will fall to 1.3320-1.3350,

and in the following 1-3 days it will go up to 1.3500-1.3550 ?

Will the neuro-indicator work with such price zigzags ???

I would like to see the results of its testing!

I most likely may not appear on this forum anymore, so if it's not too much trouble,

that I look forward to comments to myself on Mile : Skywind99@mail.ru

 
Better, you have worked with different DCs and know many of them from working on real accounts.
Could you share some information/statistics about them? Slippage, requotes, minimum time of position holding...
And most importantly, the ability to withdraw honestly earned money :)
This forum is forbidden to discuss brokerage companies, so please contact me at favoritefx@mail. ru.
Help a novice pirate :))))
 
Skywind:

Good day everyone !
I have come to this Forum for the first time.

And I am interested in the topic of using neuro-analysis for stock trading ,

FOREX in particular.

Can I ask you a question? I must clarify at once that I am not a programmer, but just a beginner trader...

A neural net can predict price movement on what time intervals???
Is it only during the trading day or beyond as well?

If it's possible, you may try to check these predictions on the EUR/USD chart in the coming week.

How will this indicator behave during sharp price fluctuations?

Suppose with a certain probability the pair's rate will fall from 1.3465 in 1-2 days to 1.3320-1.3350,

and in the following 1-3 days it will go up to 1.3500-1.3550 ?

Will the neuro-indicator work with such price zigzags ???

I would like to see the results of its testing!

I most likely may not appear on this forum anymore, so if it's not too much trouble,

that I'm waiting for comments to myself on MIL : Skywind99@mail.ru

A neural network can make predictions on any timeframe, but it tends to make better predictions on smaller timeframes. During sharp fluctuations the neural network usually gets very confused, so it is better not to trust it when zigzagging.
 

ANN is a tool, just like the LR channel. Given that the trader makes decisions through his NS (which is in his head), what prevents him from making decisions through ANS. IMHO the conversation should be directed towards "how to use NS". For example try Pesavento models to recognise. Hmmm....

 
Piligrimm:


Neural networks may do many things, the only problem is in the preparation of initial data, as an example - EURUSD M5 chart, but I have not optimized the network to the end yet, I am working on it now. All lines on the chart are the result of the neural network operation without using any indicators.

I have a question about the graph. As far as I understand, the curves drawn on the left side of the graph are "ex post facto" calculations, that is, they are based on past data, shifted back about half a period, and the last half a period is drawn by the neural network. Did I get it right?

The question is: do you use mql4 codes or do you use third-party programs for neural network calculations?

In general, I'm somewhat skeptical about such methods because my numerous attempts to predict signals ahead using neural networks gave such coarse results that it was almost impossible to use them in real trading. More or less decent results were obtained by Fourier transform adjusted to resonance with past oscillations and extrapolated forward. I wrote a little about it on the forum in the section about Fourier. The harmonic balance method is also quite good and quite accurate in forecasting the future. But I'm still hoping that neural network methods can be useful for something else besides classification. If it is not too much of a secret what are you working on, could you describe it at least in a few lines?

Thank you in advance for your answer.

Regards. Gr. Al. Nik. (ANG).

 
ANG3110:
Piligrimm:


Neural networks may do a lot, the only problem is in preparation of the initial data, as an example - EURUSD M5 chart, but the network has not been optimized yet, I'm working on it now. All lines on the chart are the result of neural network operation without using any indicators.

I have a question about the graph. As far as I understand, the curves drawn on the left side of the graph are "ex post facto" calculations, that is, they are based on past data, shifted back about half a period, and the last half a period is drawn by the neural network. Did I get it right?

The question is: do you use mql4 codes or do you use third-party programs for neural network calculations?

In general, I'm somewhat skeptical about such methods because my numerous attempts to predict signals ahead using neural networks gave such coarse results that it was almost impossible to use them in real trading. More or less decent results were obtained by Fourier transform adjusted to resonance with past oscillations and extrapolated forward. I wrote a little about it on the forum in the section about Fourier. The harmonic balance method is also quite good and quite accurate in forecasting the future. But I'm still hoping that neural network methods can be useful for something else besides classification. If it is not too much of a secret what are you working on, could you describe it at least in a few lines?

Thank you in advance for your answer.

Regards. Gr. Al. Nik. (ANG).

In this variant, a neural network makes a forecast based on a single time slice of a group of symbols submitted to the input without using lagged arguments. Correspondingly, all lines on the graph are the results of forecasts at different neural network outputs recalculated by history points and passed through filters with different settings, that is why they show different leading. I didn't make any artificial shifts, if there wasn't any prediction the lines after filtering would have moved to the right relatively to current quotes, like moving averages for example.

In this example I used external neural networks, in general I use formalized neural networks that are represented as polynomials after training and programmed with mql4, however accuracy in this case is much lower, besides retraining in dynamics is not possible.

I can understand your scepticism, I too wasted many years in a futile exercise before I got anywhere.

 
Piligrimm:

In this variant the neural network makes a forecast based on one time slice of a group of instruments submitted to the input without using lagging arguments. Accordingly, all lines on the graph are the result of prediction on different neural network outputs recalculated by history points and passed through filters with different settings, hence they show different lead. I haven't made any artificial shifts, if there was no prediction the lines would have moved to the right after filtering relative to the current quotes, like moving averages for example.

Thanks for the reply!

So it's a few instruments filtered out. Probably the main six EURUSD,GBPUSD,USDCHF,AUDUSD,USDCAD,USDJPY?

But there are 8 lines on the chart and 2 others ? Or am I getting it wrong again ?

About the polynomials - is it probably from NeuroShell2?

I tried that method, it's a bit better than backpropagation nets, but it's also very crude. Besides, polynomials often lead to additional extrema and you can easily get confused with them. All these methods do not stand next to what I have described above. Especially the harmonic balance method. But it is quite difficult to implement in software to achieve unambiguous results. It is very difficult to make all constructions by one person, quickly.

I even tried to automatically search for similar plots on the history and extrapolate them using the classification of forms. I will only get 50-60% coincidence, which is clearly not enough for successful trading.

And what allows you to be ahead? It's not quite clear to me yet.

 
ANG3110:

I use 15 instruments. All outputs are EURUSD but have different filter settings. I haven't tried your methods - I can't say anything in this regard. I used PolyAnalyst to build the polynomials.
 
njel:

ANN is a tool, just like the LR channel. Given that the trader makes decisions through his NS (which is in his head), what prevents him from making decisions through ANS. IMHO the conversation should be directed towards "how to use NS". For example try Pesavento models to recognise. Hmmm....

As a pessimist I will say - "To the madness of the brave we sing a song ..." :))
What prevents the NS from making decisions?

1. For instance, their lack of brains.
Our brains contain approximately 100 million neurons (correct me if I am not mistaken).

2. For example lack of learning.
Learning is exponential, the first three years a child learns just to move his/her limbs and understand/say the simplest words.
And yet his brain is almost the same as yours in terms of number of neurons.

3. For example, the notorious Curve fitting.
Otherwise, just remembering patterns (which do not have to work).

Perform an experiment:
- Create a neural network of 50-100 million neurons.
- Train it for 10-20 years.
- And you'll get a pathetic semblance of Homo Sapience ... :))

If you think that a modern computer is much faster than your brain, you are seriously mistaken.
A computer keeps one thread of calculations (let it be two today - not the point).
Your brain can run thousands and millions of threads in parallel ...
(it's called subconscious, associative thinking ...)
 
There are about 10 to 15 three billion neurons in the brain. Each one has an average of 10,000 synapses. And this whole bunch controls the whole body. So when you create a sales network you don't need information at all from your foot and ear receptors. and that kind of nonsense.

When I was young, I foolishly built networks with 50 million neurons. Not 10,000 connections per neuron... but still. And I trained that network for a week. The problem was that I didn't want much from it, and my mini-brain cheekily memorized the training sample. Eventually I got to reurent networks... etc.

In short, it's not about creating an analogue of a trained human brain. You need to create a brain that helps you make decisions. Or if you're confident enough, trade. This is a completely different task.
1234567891011...17
New comment
Reason: