Menu
Sister Sites

Options

Bookmark Thread

First Page First Unread Last Page Last Post

Print Thread

Similar Threads

Neural network (deep learning) EA & some manual trades 70 replies

Best deep learning school/mentor/method 14 replies

Learning at the speed of snail [Newbie Learning Journal] 9 replies

Idea/Strategies for Modular EA Architecture 7 replies

What's your dedicated server architecture? 18 replies

Attachments: Deep learning prediction with DeepMind's Wavenet architecture
Exit Attachments

Deep learning prediction with DeepMind's Wavenet architecture

Quoting sinusgamma
{quote} I trained it on Google Cloud with a single Nvidia K80 GPU. Training a model was some hour, but I trained much more than in the notebook. Last week it became very hard to get google GPU because everybody sitting in home office started using the cloud. Trained some models on CPU, but that was very long. This is a portfolio project. I would be happy to work with a professional group to develop systems based on deep learning models. If not, then I will try to implement other solutions for time series from natural language processing or computer...
super good man, btw I didn't mean to put pressure on you. I just think results are very imoprtant, even backtested as long as you tell us a little bit about your strategy
Quoting sinusgamma
Hello, I built a deep learning model to predict forex prices.
Convolutional layers? I am also working on a system, but I took a different approach. I am using evolutionary strategies , random search and coordinate descent. Something similar to Genetic Algorithms or Beam Search. The system is still in early development but I have good feeling it will be much more profitable than neural nets. The problem with neural nets is that they will never find global minimum, and in finance you need to have your model as close to global optimum as possible, and there is no way you can achieve that without random search. My algorithm evolves programs in x86 assembly instructions , no Python or this kind of "kids stuff". When I confirm success, I will move this to GPU assembly (waiting for AMD RDNA2 to be released this autum) , buy some solar panels to have free energy, and there we go! The only way to make millions in Forex is by using a high quality trading system!!!
Hello ninjakiller. The system must be awesome, but of course, it isn't possible to value a house only from seeing the bricks, and the input features are at least as important as the model. What kind of inputs do you use if you can share that info? Being a meteorologist it would have been too much effort and roundabout for me to learn such a low-level language for solving machine learning problems, so I have to use the "kid stuff" , but I really appreciate your skills.

I don't think that you have to find the single global minimum, it is enough to be close to it in the case of financial forecast. The data you train the system on will always be from a bit different distribution than the data you will get when going live because this isn't some sample from a nice distribution, but life, where the circumstances are ever-changing, and we observe only a part of it. Coordinate descent or gradient descent is needed to go in the right direction to fit our approximatior. But to build models that generalize well on unseen data most of the time we intentionally stop before the perfect fit to the training data, because if you perfectly fit the training data there is a good chance that because of overfitting the model skills on unseen data already started decreasing.

But of course, your model can have different nature.

I agree with you that randomness can help, even for generalization. My last models in the notebook output mixture distributions instead of value predictions. But these kinds of "random" layers can be implemented in the lower layers as well, building a bit variational autoencoder style (ok, I know, this isn't the thing you were speaking about).

Good luck to you, and I really like the solar panels!
Only my scepticism keep me from being an atheist. Be sceptic ever.
Quoting BlackGear
Nice work, I have done similar thing with DL to predict the market direction for the next minute and I achieved accuracy of 60%. But unfortunately this isn't tradable and I couldn't get good results for higher time frame. Overall, No the market is not random.
What did you try to forecast? price, or did you categorized the label for direction sides?
Only my scepticism keep me from being an atheist. Be sceptic ever.
Quoting sinusgamma
Jobushy, if it were a pure random walk the price wouldn't react to the events of the world. From the view of an uninformed observer, it can seem randomish, but even without any information about politics, economy, disasters the price movement has more robust patterns than a random walk (which doesn't have). If you aren't tossing a fair dice to decide trading action, then you are trying to find these patterns, even if you don't think about your actions this way. This article has some nice charts to compare random walk to forex: https://www.dukascopy.com/fxcomm/fx-...15&language=en...
Thank you for your explanation and great links. I agree with some of your points. We do look for patterns from historical data. However, a lot of patterns can also be found in random walk data. Many of these patterns are coincidental or overfitting, there is no guarantee that they will succeed in the future.

I have read the linked you provided, it is a very interesting article, but, that article is published on the broker's website. The brokers will never let traders know that foreign exchange is a random walk, and the objectivity of the article is questionable.

There is also a big issue in the article: the author's random data is generated based on a uniform distribution, but the real-world data are normal distribution, they cannot be directly compared.

I plotted some charts with two kinds of random walk data. The 2D chart with normal distribution is very similar to the chart of real data provided by the author. This proves that prices movement are random walks.
Attached Images (click to enlarge)
Click to Enlarge Name: rand-walk with standard normal distribution.png Size: 202 KB Click to Enlarge Name: rand-walk with discrete uniform distribution.png Size: 221 KB
Attached File
File Type: txt random-vs-forex.py.txt   1 KB | 49 downloads
Jobushi, nice research and charts. I agree that the charts in the linked article arent satisfactory to prove the non-random walk theory. But your charts arent satisfactory either to prove the random walk theory. There is a continous debate about this even in academia. So I only can say: in my opinion forex is not pure random walk.
Only my scepticism keep me from being an atheist. Be sceptic ever.
Quoting jobushi
I have read the linked you provided, it is a very interesting article, but, that article is published on the broker's website. The brokers will never let traders know that foreign exchange is a random walk, and the objectivity of the article is questionable.
Checking the source and reproducing the expirement is a very good attitude but the article was written by a forum member like you and me not the broker.

Quoting jobushi
There is also a big issue in the article: the author's random data is generated based on a uniform distribution
The article never says that the synthetic data was generated from a uniform distribution. It just says "random walk". Without any indication of the distribution it is customary to understand a discrete brownian motion which implies a normal distribution.

Quoting jobushi
..., but the real-world data are normal distribution, they cannot be directly compared.
Price movements are real-world and they are clearly not normally distributed.

Quoting jobushi
...This proves that prices movement are random walks.
No this is not a sufficient proof. If price movements were just random walk they would be iid. How do you then explain the volatility clustering? You need the distribution to change over time. I do perfectly admit these changes are random(1) themselves but if the volatility changes there is no reason for the mean to stay zero over time(2). If the mean is non-zero during long enough periods you get trends (stochastic trends but not spurious trends). Therefore even if the price is random, trends should be an exploitable pattern and there are certainly other such patterns(3).

-----
(1) They are mechanically derived from real life interactions, interest rate, news, fear, speculation, network latency, duration of the pee break of a bank trader... But the complexity of the system makes it so chaotic that it is essentially random for us.
(2) Interest rates are an example of fundamental which can create a non-zero mean due to carry trading (see the good old days of GBP/JPY)
(3) For instance the volatility clustering is tradable through options.
No greed. No fear. Just maths.
Quoting PipMeUp
I do perfectly admit these changes are random(1) themselves but if the volatility changes there is no reason for the mean to stay zero over time(2). If the mean is non-zero during long enough periods you get trends (stochastic trends but not spurious trends). Therefore even if the price is random, trends should be an exploitable pattern and there are certainly other such patterns(3). ----- (1) They are mechanically derived from real life interactions, interest rate, news, fear, speculation, network latency, duration of the pee break of a bank trader......
Hi, PipMeUp, Thank you for pointing out so many facts.

I just started writing EAs, and have read some of your old posts, they have benefited me a lot, thank you again.
Quoting PipMeUp
{quote}Therefore even if the price is random, trends should be an exploitable pattern and there are certainly other such patterns(3).
Who are you, and what have you done with the real PipMeUp?
  • Post #30
  • Quote
  • Last Post: Jun 29, 2020 12:02pm Jun 29, 2020 12:02pm
My experience is neural networks fits well into clusters of low volatility bars.
0 traders viewing now