Any success with neural networks in trading | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Neural Networks: Myths And Reality 19 replies

Extreme scalping using neural networks - EA/Manual 646 replies

Cognitive indicators and neural-networks 19 replies

Neural Networks in Trading (Java, Algodeal, Technical Analysis) 3 replies

Price Analysis with Neural Networks 41 replies

Attachments: Any success with neural networks in trading
Exit Attachments

Any success with neural networks in trading

  • #1
  • First Post: Dec 10, 2015 5:05am Dec 10, 2015 5:05am
Hi everyone, I am currently doing some research on using neural networks in trading. I have seen lot of implementations of neural nets with different methods in price predictions in different ways like daily range prediction, predicting close price, etc.
Is there anyone who is successfully using neural nets in trading? Any ideas or tips or information are welcome and would be helpful :-)
  • #2
  • Dec 10, 2015 5:34am Dec 10, 2015 5:34am
Neural Network programs are advanced algorithms which can read and react..

Neural network algorithms have many inputs to get one output..

This technology used in many different sections.

Ex : https://www.ted.com/talks/fei_fei_li...es?language=en

Any indicator based expert advisor similar to neural network systems it just depends on quality of algo's behind them..

If good trading rules can applied in to qualified expert advisor success is possible..
Attached Image (click to enlarge)
Click to Enlarge Name: 6283214_artificial-neural-networks-are-changing_48e377f8_m.jpg Size: 42 KB
Take it easy.
  • #3
  • Dec 10, 2015 5:54am Dec 10, 2015 5:54am
you are right Quants. It always depend on the type of inputs to the neural nets and the method of processing selected with the number of outputs expected. There are some people who are using the neural nets on the go during trading and based on outputs an EA is used to perform trading using live signals. I hope that might take a lot of effort
  • #4
  • Dec 10, 2015 8:46am Dec 10, 2015 8:46am
looks like there is not many people using neural nets for trading :-(
  • #5
  • Dec 10, 2015 12:27pm Dec 10, 2015 12:27pm
Quoting tradeforlife
looks like there is not many people using neural nets for trading :-(
neural networks need inputs, you need to first figure out what makes the market move (thats the hardest part) after that, assuming you have enough computing power you can set up your NN
i lose money for a living
  • #6
  • Dec 18, 2015 2:28pm Dec 18, 2015 2:28pm
Like this thread. I've been investigating NNET with regards of forex for some time now. Computing power is not a big problem with e.g. cloud solution like Azure and its machine learning capabilities. I really love it.
I've been using R for data collection and pre-preparation, basic features definitions etc. Then upload such data set onto azure machine learning and run the experiments. Advantage of using machine learning machinery on azure is that once predictive model is found and created it can be easily published as a simple REST JSON web service. So it's very easy callable in any automated trading program over HTTP requests.
With all that said, it's very hard or even hard and time consuming to find proper feature set being able to be stable enough in forward testing. And it's not just about features but also about proper settings of learning algorithm parameters. I do mean to say that I can easily get model with 97% of Coefficient of Determination (see attached GBPUSD 15 minutes), but I can't get the same or even similar results in forward testing. Cross-validation gets 84-87% percent of the Coef. of Determination. But the reality is yet worse. The reason for that is that NNET are very prone to overtraining and that's apparently my problem. But I'm not about to gave it up (yet). It's quite exciting investigation and I really enjoy it. I'd like to heard about someone else who is investigating or even successfully using such things in real trading. That would be energizing....
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot1.png Size: 54 KB
  • #7
  • Sep 21, 2016 8:31am Sep 21, 2016 8:31am
Quoting mcquak
I've been using R for data collection and pre-preparation, basic features definitions etc. Then upload such data set onto azure machine learning and run the experiments.
Considered using a data factory pipe to organise the data for the input?
  • #8
  • Jun 22, 2018 4:15pm Jun 22, 2018 4:15pm
Quote
Any success with neural networks in trading

Yes..... After considerable research and various neural implementations I have finally generated a reliable and profitable AI trading system. Its a bit of a learning curve but worth it in the end. Keep on experimenting and you may also have success.
  • #9
  • Edited 9:00pm Jun 22, 2018 8:25pm | Edited 9:00pm
The only interesting thing about this topic is to develop an AI that can duplicate and improve a very successful trading method or start completely from scratch and let it figure out bias and momentum by itself, from raw numbers.

For that you have exactly know what to look for and then increase the field for opportunities.

That's what i'm trying to do manually (side project) in the USD and YEN pairs.
  • #10
  • Jun 23, 2018 10:11am Jun 23, 2018 10:11am
With the technological advancement, foreign currency exchange traders are moving towards more updated technical systems and trading methods that can actually help a trader to engage in trading more cost effective way. Neutral network is a community of artificial intelligence that is being developed to summarize available data to give one single conclusion by analyzing all aspects. As human brain takes time, neutral network can complete your analysis work in a less time consuming way.
  • #11
  • Jun 23, 2018 11:33am Jun 23, 2018 11:33am
Quoting The-Flipper
For that you have exactly know what to look for and then increase the field for opportunities.
I both agree and disagree with your statement

Agree
For shallow neural networks i.e. less than 2 hidden layers and 10-50 inputs it is advantageous to to have detailed prior knowledge of the problem space and present the network with "intelligent inputs" that have been derived from human selection and specific system ideas.

Disagree
For deep neural networks i.e. 3-10 hidden layers, with convolution matrices and 50-1000 inputs the general industry agreed process is to literally "throw" as much data as you can at the network, even data you feel maybe irrelevant. You then let the network decide what is important or not to solve the reward function you have given it.
  • #12
  • Jun 23, 2018 12:46pm Jun 23, 2018 12:46pm
Quoting go4it
{quote}Disagree For deep neural networks i.e. 3-10 hidden layers, with convolution matrices and 50-1000 inputs the general industry agreed process is to literally "throw" as much data as you can at the network, even data you feel maybe irrelevant. You then let the network decide what is important or not to solve the reward function you have given it.
I completely understand the problem.
If you feed the network with more and more data, it can maybe come up with a much better trading approach.
I'm not denying that.

But my goal is to develop an AI which i fully understand, based on certain parameters and rules.
It should fit my style/method so that i can learn from it and possible improve my trading.
My manual findings showed enough results and excellent opportunities during a normal week.
And for (now) i don't need more, the playing field is already big enough.
  • #13
  • Jun 23, 2018 1:32pm Jun 23, 2018 1:32pm
The-Flipper I came across something a few months ago that you may be interested in..... I have poor discretionary trading skills so it was not of much use to me, but if you are trying to replicate one of your manual systems it may help.

Spoiler alert - sorry if it distracts from the meat of the thread there.

https://www.alpaca.ai/alpacaalgo/ is an online automated trading service that allows you to highlight lots of portions of a chart that you would normally trade, it then trains itself to find these historically and backtests your strategy for you. It might save you the effort of doing it yourself.

I think it uses SVM ( support vector machines ) and decision trees as its machine algo.

Here is a vid from the CEO

Inserted Video
  • #14
  • Jun 23, 2018 7:28pm Jun 23, 2018 7:28pm
Quoting go4it
The-Flipper I came across something a few months ago that you may be interested in..... I have poor discretionary trading skills so it was not of much use to me, but if you are trying to replicate one of your manual systems it may help. Spoiler alert - sorry if it distracts from the meat of the thread there. https://www.alpaca.ai/alpacaalgo/ is an online automated trading service that allows you to highlight lots of portions of a chart that you would normally trade, it then trains itself to find these historically and backtests your strategy...
Sadly this site has nothing to offer what i need.
I don't trade on chart patterns or some stupid indicators.

My idea is to create a self-learning AI that mirrors and optimizes my discretionary trading method.
I wanna set specific parameters, standards, examples and rules and it has to learn it/figure out by using historical data.

Basically i wanna teach and give input how i see the market and how normal bias and strong momentum phases look like.
And later even how to trade against normal bias in specific moments/sessions.
  • #15
  • Jun 24, 2018 9:39pm Jun 24, 2018 9:39pm
Neural networks in trading = waste of time.

The general rule with neural networks is that if you want to extract valuable results you have to feed the system with valuable data.

Such data does not exist in the Forex space because this market is decentralized.

So with technical NN applied to Forex you will get the classic GIGO. (garbage in, garbage out).

Even if the back testing is showing positive result the actual live trading is always random.

You cannot eliminate the human factor.
  • #16
  • Jun 24, 2018 9:54pm Jun 24, 2018 9:54pm
Quoting The-Flipper
{quote} Sadly this site has nothing to offer what i need. I don't trade on chart patterns or some stupid indicators. My idea is to create a self-learning AI that mirrors and optimizes my discretionary trading method. I wanna set specific parameters, standards, examples and rules and it has to learn it/figure out by using historical data. Basically i wanna teach and give input how i see the market and how normal bias and strong momentum phases look like. And later even how to trade against normal bias in specific moments/sessions.
Thats kind of a tall order dont you think?

My opinion is that it is possible to teach a nn to either mirror your trading method. But to optimize it takes some creativity. Irregardless, the first step is to define, and then quantify the signals/situation whereby you will enter/exit a trade. How to do that is the golden question aint it? Thats where the creativity aspect comes in.
I look for value wherever it can be found
  • #17
  • Jun 24, 2018 10:14pm Jun 24, 2018 10:14pm
I guess the main point that most people overlook is the fact that machine learning algorithms such as neural networks are nothing more than just tools. It is the practitioner's ability to use these tools that makes the difference. Quantitative trading is not simply about feeding data to a algorithm - anybody can do that! There is an "art" part to it. One which is based on the situation at hand and relies on a fair amount of subjectivity.

Of course it can be argued that in the financial markets (or nearly anywhere else), that the past does not reflect the future. But it does give us a rough estimate! And that brings us back to the situational and subjectivity point. Under what circumstances do we trust our forecast? And under what circumstances do we become skeptical of it? IMO it really boils down to knowledge and experience.

Unfortunately in this greed driven industry, most people are looking for the easiest route - even in quantitative trading. They just throw a bunch of random data into a few algorithms, churn out the best performing ones. And implement them. It may work for awhile, or even a long time. But such recklessness will in my honest opinion, only lead to an eventual downfall.
I look for value wherever it can be found
  • #18
  • Jun 25, 2018 2:03am Jun 25, 2018 2:03am
Neutral Networks need more input for a single output and that's the hardest part to use it.
  • #19
  • Jun 25, 2018 3:53am Jun 25, 2018 3:53am
Quoting alphaomega
Neural networks in trading = waste of time. The general rule with neural networks is that if you want to extract valuable results you have to feed the system with valuable data. Such data does not exist in the Forex space because this market is decentralized. So with technical NN applied to Forex you will get the classic GIGO. (garbage in, garbage out). Even if the back testing is showing positive result the actual live trading is always random. You cannot eliminate the human factor.
also; NN does not work because it does not explain the results.

would you as trader trust a black box of managing your account; without knowing why it does what it does?

Not knowing why a decision is made to buy or sell; also has a negative impact on the trader moral; and confidence in the system.

A very good example of this failure; is Tesla auto pilot. It works until it does not.
  • #20
  • Jun 25, 2018 7:12am Jun 25, 2018 7:12am
Quoting synicz
neural networks are nothing more than just tools. It is the practitioner's ability to use these tools that makes the difference. Quantitative trading is not simply about feeding data to a algorithm - anybody can do that! .
Excellent comments!
It is about feature selection, selecting the data sets for training,regularization,selection of inputs-outputs etc...Neural networks and deep nets suffer from two major problems:
1-They are susceptible to noise.Check this (Anonymous) https://www.quora.com/How-can-I-hand...chine-learning
2-They are stationary.I mean you can feed the neural networks or deep nets with the cat face data , they may be able to predict the cat face with up to 99 percent accuracy.But if you feed the Neural networks with continous financial data for prediction ,it is a failure.

Interview of Prof. Judea Pearl father of tragic Wall Street Journal reporter Daniel Pearl and famous critic of Deep Nets :

“The way you talk about curve fitting, it sounds like you’re not very impressed with machine learning.
No, I’m very impressed, because we did not expect that so many problems could be solved by pure curve fitting. It turns out they can.” https://www.quantamagazine.org/to-bu...fect-20180515/
0 traders viewing now