- This topic has 68 replies, 20 voices, and was last updated 10 years, 8 months ago by
.
- Posts
Let’s say that we got acurate X years of historical data.
We also got all available parameters for currency pairs: OHLC, Bid, Ask, Volume (ticks), Swap, etc.
And we got all kind of indicators (each with std. params), many of those.
How would you test to find out:
which one (or combination of several) will give the most accurate prediction for 1 year of “out of sample” data for a specific TF (let’s say H4)?
What is the prefered “machine” or “method” to do this task (preferably available and open-source)?
I made an interesting experiment:
I fed a Neural-Network algorithm almost all conceivable known parameters:
Stochastic, RSI, CCI, Bolinger, Pivots, S/R, MA’s, Swap, OHLC, Bid/Ask, Volume, Currency Strength, etc. etc…. to run on each and on combinations of several params.I Let it run on 3 yrs historical data to find and sort most influetial parametrs (what are the most significant parametrs) and combinations of several params to “predict” one year of “out of sample” data …for a specific TF.
It ran for more than 17 days (!) on fairly strong computer, before spitting out the results.
Some of the highly significat results are known to me from experience, and some are quite surprising (because I previously used to ignore them).I have to further analyse the results, and let you know.
Hey G!
Very excited to see your results.
Reminds of a thread from ff where proximus did a much more simple version of this http://www.ff.com/showthread.php?t=487974 (change ff to forexfactoy)
-Life’s Good & Why Wouldn’t It Be? I’m a Pirate After All!Interested in the results, and the methods/tools used to get them. Seems an army of testers each testing one tf and strat/indy and compiling the results could be a huge win.
Without overheating our processors. lol
Awesome to see this thread G! Can’t wait to hear about your results and discuss further about the method. I have been thinking about this as well during my breaks. Like ways of combining TZs with other indicators to arrive at probabilities and turning points.
Very interesting! Subscribed!
Focus, Patience, Determination & Order in chaos
Yeah, but he did it on individual indicator at a time, with fixed parameters in changing market conditions.
We all know that this doesn’t work.
Another thing is the method of “Always in” – enter on signal and exit on opposite signal. You can’t expect a single indicator to provide both accurate entry AND accurate exit.
His research is priceless in putting the whole “indicators” is proper perspective – but I’m looking for something much more complicated – using combinations of data/indicators.
G.
Interested in the results, and the methods/tools used to get them. Seems an army of testers each testing one tf and strat/indy and compiling the results could be a huge win. Without overheating our processors. lol
I think I’ve already said that the target is just one TF (Preferably H4).
G.
Yeah, but he did it on individual indicator at a time, with fixed parameters in changing market conditions. We all know that this doesn’t work. Another thing is the method of “Always in” – enter on signal and exit on opposite signal. You can’t expect a single indicator to provide both accurate entry AND accurate exit. His research is priceless in putting the whole “indicators” is proper perspective – but I’m looking for something much more complicated – using combinations of data/indicators. G.
I totally agree with you. What he shared is just the tip of the iceberg in terms of what can be done with testing indicator probabilities, this is why I am very excited to see your findings :)
-Life’s Good & Why Wouldn’t It Be? I’m a Pirate After All!I very much like all kind of “channels” because they give you a definite “area” with boundries to focus on.
So if you are consistantly aiming at the center of that channel – you are not far away from target – and that’s a start. From that point it’s just a matter of refining the accuracy.
I tested many types of “channels” (Bollinger, Keltner, various regression, envelope, ATR, Price, and more).
Among the “surprising” results are two correlated ones: Weekly S/R and Weekly Pivot.
It’s something that I never used before… I’m trading much lower TF’s, H4 and below.
I still need to double and triple check the results, because it’s first run and probably have some bugs and logic flaws – but interesting outcome.
G.
Great thread, gg53!
Definitely follow
I am coding a tiny project on price prediction by using resilient backpropagation algorithms.
It predict high, low, close price.
I use ATR 20 to see if close > high or close < low, it like break out, i will buy or sell.
It just an idea i plan to code.
Weekly S/R and Weekly Pivot sound interesting, i will try to code too.
Keep posting gg53 !
Great thread, gg53!
Definitely follow I am coding a tiny project on price prediction by using resilient backpropagation algorithms. It predict high, low, close price. I use ATR 20 to see if close > high or close < low, it like break out, i will buy or sell. It just an idea i plan to code. Weekly S/R and Weekly Pivot sound interesting, i will try to code too. Keep posting gg53 !Try to multiply your ATR by 4, i.e. 4*20=80
G.
Try to multiply your ATR by 4, i.e. 4*20=80 G.
Awesome thread. Thanks G, subscribe. About multiplying the ATR with 4, is it because of 1 month has 4 weeks ?
Try to multiply your ATR by 4, i.e. 4*20=80 G.
Awesome thread. Thanks G, subscribe. About multiplying the ATR with 4, is it because of 1 month has 4 weeks ?
I’m getting better results with ATR between 80-100.
G.
Thanks all for your comments, but I asked for suggestions/ideas on how to achieve the task presented in the first post.
There are several NN aproaches, and there are some other ways of doing it.
I’m looking for such – not relying on just one “engine” which might be not very suitable for the mission.
G.
Here is ATR=4 on 10K bars of H1 AUDUSD from Dukascopy.
I see predicted High and Low act as High and Low band like BB.
But the good thing is that they can adapt to current market trend not like other indicator.
But i dont see how i can trade them yet
I need more time to play with it to see.
But gg53, do you use other pairs like AUDUSD, GBPUSD as parameter to predict EURUSD?
Hi gg53 and Zelo
pls check this
I think you will like itBest wish.
Hi gg53 and Zelo pls check this I think you will like it https://lgtrades.wordpress.com/about/ http://forums.babypips.com/free-forex-trading-systems/42052-what-really-turned-my-trading-around.html backup https://web.archive.org/web/20130102021335/http://forums.babypips.com/free-forex-trading-systems/42052-what-really-turned-my-trading-around.html Best wish.
I’ll sure read those links, Thanks.
G.
Here is ATR=4 on 10K bars of H1 AUDUSD from Dukascopy. I see predicted High and Low act as High and Low band like BB. But the good thing is that they can adapt to current market trend not like other indicator.
But i dont see how i can trade them yet I need more time to play with it to see. But gg53, do you use other pairs like AUDUSD, GBPUSD as parameter to predict EURUSD?I sure do, but in indirect way – through Currency Strength and Correlation indies.
G.
To illustrate the method, imagine:
1. We found an adaptable “channel” that includes at least 90% of the bars, has the narrowest boundries, and can be accurately extended a few bars forward.
2. We “aim” our canon ball roughly at this area…
3. We know the (avg?) size of the next bar(s) and our current location. So we add and substruct that size from our current location (provided it’s still within the channel boundries).
4. We can further reduce that size if we knew the direction of next bar, or use a different method of “extending” from current location. I’ll offer other ways of doing it later on.
5. Now we are left with the question: When it will get to that price area? Simple: ROC (Rate of Change).
6. It’s time to fire that canon ball…
We dont have to “guess” accurate price, just a range of very high probability prices and their location relative to current bar close.
Simple concept, harder to implement but have some logic in it.
Suggest tools, indies, scanners, algo, or whatever drive this wagon forward.
Debate the concept and suggest solutions to that.
Note:
Every tool or indi MUST be adaptable to changing market condition, otherwise it will show terrific results today and fail tommorow.
G.
Oh, if you’re discouraged by the adding and subtructing full bar sizes from current location – don’t be.
You can add/subtruct only the Open/Close distance of the “future” bar(s).
G.
Great post. I have experimented with NN and at a high freq level, there is a good signal i could generate (using order book info from primary markets). In last few years there has been very significant advancement in neural networks particularly techniques for deep learning which i think have a lot of potential in trading too. look into convolution networks, RBMs etc. I want to experiment this on retail level too so would be very interested in working together on this.
Great post. I have experimented with NN and at a high freq level, there is a good signal i could generate (using order book info from primary markets). In last few years there has been very significant advancement in neural networks particularly techniques for deep learning which i think have a lot of potential in trading too. look into convolution networks, RBMs etc. I want to experiment this on retail level too so would be very interested in working together on this.
It took me nearly 6 months to develope my own NN to handle this, and I’m quite sure it’s not pefect and probably has many bugs and logic flaws.
I don’t want to repeat or debug it – I’m looking for something ready or maybe something that needs small refinments.
That’s why I asked if there is something “engine” that can handle that many parameters and indicators effectively.
Good to see you here.
G.
Great post. I have experimented with NN and at a high freq level, there is a good signal i could generate (using order book info from primary markets). In last few years there has been very significant advancement in neural networks particularly techniques for deep learning which i think have a lot of potential in trading too. look into convolution networks, RBMs etc. I want to experiment this on retail level too so would be very interested in working together on this.
It took me nearly 6 months to develope my own NN to handle this, and I’m quite sure it’s not pefect and probably has many bugs and logic flaws. I don’t want to repeat or debug it – I’m looking for something ready or maybe something that needs small refinments. That’s why I asked if there is something “engine” that can handle that many parameters and indicators effectively. Good to see you here. G.
yes, it does take a long time to get all things working (partly because of the learning curve). We have used lot of open source code thats available. I am not sure if you are using GPUs at this point but if you do then you will want to use cudalib which has many algorithms implemented already.
Otherwise look for packages from http://deeplearning.net/software_links/ – again there will be some learning curve but you wont have to implement all the algos yourself from reading papers etc.
you certainly dont have to develop your NN from scratch – there are implementations in opensource that will be error free and will do almost anything you want to do.
btw if you are looking for some commercial platform/code, check this http://newsite.asirikuy.com/?page_id=3420. algotradingjo at FF uses this.
That’s why I asked if there is something “engine” that can handle that many parameters and indicators effectively.
Have you looked into this G?
http://www.heatonresearch.com/encogThey have libraries for almost all languages.
Also I have used FANN library and it was pretty good.
Focus, Patience, Determination & Order in chaos
Definitely follow 
I am coding a tiny project on price prediction by using resilient backpropagation algorithms. It predict high, low, close price. I use ATR 20 to see if close > high or close < low, it like break out, i will buy or sell. It just an idea i plan to code. Weekly S/R and Weekly Pivot sound interesting, i will try to code too. Keep posting gg53 !
I need more time to play with it to see. But gg53, do you use other pairs like AUDUSD, GBPUSD as parameter to predict EURUSD?
- You must be logged in to reply to this topic.