Machine Learning with algoTraderJo | Page 21 | Forex Factory

Options

Go to
Permalinks
Print Thread

Similar Threads

Algorithmic Quant Trading (Machine Learning + Stat-Arb) 25 replies

Forex with Machine Learning Software Project 1 reply

Machine Learning + Retail Forex = Profitable? (Quant) 1 reply

Potential new machine learning style software. 79 replies

My most recent advancements into machine learning 16 replies

Attachments: Machine Learning with algoTraderJo
Exit Attachments

Machine Learning with algoTraderJo

  • #401
  • May 13, 2015 7:51am May 13, 2015 7:51am
Quoting levkly
Hi algoTraderJo, Thanks for your posts ,It's very helpful. 1. You told that you combine other fin markets to ML input. How do you choose that markets ,Any known tests? 2. Do you use any other inputs except last price returns ? 3. Do you have recommended book to read about ensembles that can be useful for fin markets. 4. In order to reduce long drawdowns , What do you recommend to improve the ML algo, use lower TF, Use ensembles ? Thanks Lev

  1. We will get into this much later into the thread. It is not very useful to go into this with so much about the basics still not properly addressed.
  2. Yes of course, but still too early to say anything about this. We haven't even seen half of what can be done with simple price returns as inputs.
  3. No. You won't find any books with the info that you require for that, not even academic publications address this properly. Daniel wrote a blog post recently about this (I think it came from a discussion we had about the subject). You do need a solid formation in statistics, to carry out experiments and understand what is going on in ensembles that work and ensembles that don't work. The info already on this thread should give some hints about this but we'll keep working on it.
  4. To reduce drawdown lengths you need both ensembles and portfolios. Going to lower timeframes won't guarantee anything, if anything you will mostly have to deal with more noise and predictions will become harder to make, the lower you go the smarter you need to get.

  • #402
  • May 13, 2015 10:11am May 13, 2015 10:11am
Quoting KaBo
I am very interested in how you go about choosing systems for ensembles. From what I understand, you 1. Have an idea about target (mse/mle regression in the last case), inputs (bars returns) and setup (hourly / daily with 4 Parameters to optimize) 2. Generate lots of systems with different parameters optimizing for Sharpe / R2 3. Pick individual systems that perform good 4. ... but how do you pick systems that perform good together as ensembles. As Asirikuy generates the system and does the Backtest at the same time, do you save generated systems...
KaBo,

You might start with this wiki page on Model Selection.
  • #403
  • May 14, 2015 1:10am May 14, 2015 1:10am
Quoting algoTraderJo
{quote} Any observation that you carry out on historical data to build a trading system is some form of data snooping and introduces mining bias. I account for all of this when I measure data mining bias.
Back to these threads after 5 years. How do you measure data mining bias and what package do you use? It looks like R or something?
  • #404
  • May 14, 2015 8:02am May 14, 2015 8:02am
Quoting NIandP
{quote} Back to these threads after 5 years. How do you measure data mining bias and what package do you use? It looks like R or something?
Still too early to talk about data-mining bias, but I measure it using my own C based custom software hooked up with R. I gather all the data with C and then use R for the hypothesis tests.
  • #405
  • Edited 10:59am May 15, 2015 4:31am | Edited 10:59am
I am curious as to the data aspects of what I have been reading.

Many things affect a currency pair:
- Another currency pair
- News of importance
- The combination of currency pairs that make up its index, which is basically the trading countries where most of the exports go.
- Also looking exclusively at 1 daily bar gives us only a summary of what happened, looking inside the bar gives us the feeling.
- Another thing is market's opening and closing times, this causes fluctuations.

All of these would be variables in a multidimensional model, just feeding on closing price sounds somewhat naive.
A combinatorial analysis of multiple currencies like a logistic regression that will give a direction for all the currencies would be interesting. I mean this because when a purchase is done somewhere, something else was sold, there is a flow of money, and although we cannot estimate real volumes in the market, we can estimate where the money is going via the candle shapes.

Merely looking at a single currency pair will not yield a proper system.
I have only started machine learning this year, but I have been looking at trading and forex for the last 5 and you learn a lot looking at what the market is doing and when it is doing it.
My plan is to mix machine learning with OpenCL(im a c developer so shouldnt be too hard), but I need to do some tests, and although I am alright using R, I am still in the dark on how to integrate R and Metatrader. Otherwise I would like to start feeding into this thread when I am up to speed.
  • #406
  • May 16, 2015 6:32am May 16, 2015 6:32am
Quoting bruce_loco
I am curious as to the data aspects of what I have been reading. Many things affect a currency pair: - Another currency pair - News of importance - The combination of currency pairs that make up its index, which is basically the trading countries where most of the exports go. - Also looking exclusively at 1 daily bar gives us only a summary of what happened, looking inside the bar gives us the feeling. - Another thing is market's opening and closing times, this causes fluctuations. All of these would be variables in a multidimensional model, just...
You should definitely try all those things and share your results with us! I am interested to see if you will come to the same conclusions I did when I attempted to solve the problem in the manner you describe. About integrating R and MT4, that is a total nightmare (at least it was when I attempted it a few years ago). The best thing is to do so through a library that is in C++ or some other language, do something like MT4<->C<->R to have the most freedom.
  • #407
  • May 16, 2015 6:39am May 16, 2015 6:39am
I also wanted to say that the previously posted EURJPY results were not correct . These results were for another method (which we will be talking about later on) and I posted them erroneously (confused the results, very sorry about that). KaBo has reproduced the USDJPY results (which are indeed correct for the mentioned method) but the EURJPY results are not those mentioned on the past few pages. The results for the EURJPY are as follows:

Attached Image


The settings for these results are A = 10 (GMT +1/+2), B = 80, C = 7 and D = 12. I will try to work with KaBo to verify he can reproduce everything I post here so that this does not happen to us again on the thread. The above results use a spread of 7 pips, as I said before. I hope that with these changes other users attempting to reproduce outside the F4 framework will be able to have better success. Thanks for the patience folks
  • #408
  • May 16, 2015 6:56am May 16, 2015 6:56am
As with our previous machine learning exercises, results are the best for the EURUSD. However the EURUSD results do benefit from a longer horizon of 20 vs the 12 value we tested on the EURJPY and USDJPY (although if you remember our USDJPY tests they also were better with a horizon of 16 instead of 12). These are the results for the EURUSD:

Attached Image


For this test I have used settings of A = 10 (GMT +1/+2), B = 130, C = 7 and D = 20. The stop for this test is also a bit shorter, at 40% of the daily ATR(20), although I will share with you results for different SL values soon as well. I used a spread of 3 pips for this test. I hope KaBo can help me verify to make sure I am not making any more dumb mistakes!
  • #409
  • May 16, 2015 7:05am May 16, 2015 7:05am
Concerning reproduction of my results using the F4 framework

If any of you have access to F4 here are the files you will need to reproduce the above EURUSD results (other results, just change the settings). Take into account that I have set this up as a system called system_1d1h and have set it up to have systemid 22 (you need to replace this in the ast.conf with whatever id you have assigned the system on your framework build). I am using F4.4.14 (latest unstable branch) compiled on the Release-64x configuration in Linux (moved to Linux recently to work on a project with Daniel). I use the 1987 data set for testing and do the tests using the ast.py (command line tester for the asirikuy python backtester). The commandline tester is a much better option than the UI tester IMHO.
Attached File(s)
File Type: zip system_1d1h.zip   3 KB | 437 downloads
  • #410
  • May 16, 2015 8:26am May 16, 2015 8:26am
Quoting algoTraderJo
I also wanted to say that the previously posted EURJPY results were not correct . These results were for another method (which we will be talking about later on) and I posted them erroneously (confused the results, very sorry about that). KaBo has reproduced the USDJPY results (which are indeed correct for the mentioned method) but the EURJPY results are not those mentioned on the past few pages. The results for the EURJPY are as follows: {image} The settings for these results are A = 10 (GMT +1/+2), B = 80, C = 7 and D = 12. I will try...
This is what I get with EUR/JPY and this new set of parameters. (didn't try E/U yet)
Attached Image (click to enlarge)
Click to Enlarge Name: newParameters.png Size: 33 KB
No greed. No fear. Just maths.
  • #411
  • May 18, 2015 6:20am May 18, 2015 6:20am
Hi algoTraderJo

How about a decision tree to add bias based on fundamentals? Wouldn't this be reduce the drawdown seen sometimes?

Bruno
  • #412
  • May 18, 2015 6:39am May 18, 2015 6:39am
Quoting bruce_loco
Hi algoTraderJo How about a decision tree to add bias based on fundamentals? Wouldn't this be reduce the drawdown seen sometimes? Bruno
How do you quantify the fundamentals over the historical data?
No greed. No fear. Just maths.
  • #413
  • May 18, 2015 6:47am May 18, 2015 6:47am
Quoting PipMeUp
{quote} How do you quantify the fundamentals over the historical data?
This data is available. You know that at x date, there were y events with z values, like interest rates, jobs data, etc.
You also know that if forecasted is lower than actual depending on the data type the effect will be positive/negative.
A large deviation from forecasted and actual value will induce an effect in the variation, and the algorithm knowing this would adjust somehow. It's what machine learning is supposed to be used for(I am not being arrogant, just been through 3 courses of the thing and I think this is what we are supposed to do).
  • #414
  • May 18, 2015 7:02am May 18, 2015 7:02am
Quoting bruce_loco
{quote} This data is available. You know that at x date, there were y events with z values, like interest rates, jobs data, etc. You also know that if forecasted is lower than actual depending on the data type the effect will be positive/negative. A large deviation from forecasted and actual value will induce an effect in the variation, and the algorithm knowing this would adjust somehow. It's what machine learning is supposed to be used for(I am not being arrogant, just been through 3 courses of the thing and I think this is what we are supposed...
OK for the recurrent events (NFP, FOMC...) but how to code the punctual events like bankrupty of a big bank, earthquake, explosion in a Japanese nuclear power plant, war in whatever country, embargo because of a previous event...? And how do you model their effect in time?
No greed. No fear. Just maths.
  • #415
  • May 18, 2015 7:07am May 18, 2015 7:07am
Quoting bruce_loco
{quote} This data is available. You know that at x date, there were y events with z values, like interest rates, jobs data, etc. You also know that if forecasted is lower than actual depending on the data type the effect will be positive/negative. A large deviation from forecasted and actual value will induce an effect in the variation, and the algorithm knowing this would adjust somehow. It's what machine learning is supposed to be used for(I am not being arrogant, just been through 3 courses of the thing and I think this is what we are supposed...
The suggestion is too broad. What exact fundamental data? How would you exactly act on it? I would encourage you to carry out this exercise and let me know what you get. After you do this let me know if you still think fundamental data is so important for the prediction of financial time series.
  • #416
  • May 18, 2015 7:11am May 18, 2015 7:11am
Well

Those you can't. What you know about these types of events is that immediately there are negative consequences on both currency and stock markets. To begin with they are local, then global.
When I was doing the quant course in coursera the guy did a small correlation on how long it would take a stock to absorb the impact of lets say a department windup, a significant lawsuit, loss of competitive edge, stuff that would drag the stock down 10-20% and sometimes even more, and this you cannot predict, what you can predict is what to do next, which was a short term recovery followed by a downtrend.
But the crux here is:
Severe news comes out that says that the jobs have fallen way bellow expected value; Your knn algo says buy but you know you have to sell. That is why I was pointing out that a model cannot just be fitted straight on price action of the currency alone.
  • #417
  • May 18, 2015 7:17am May 18, 2015 7:17am
Quoting algoTraderJo
{quote} The suggestion is too broad. What exact fundamental data? How would you exactly act on it? I would encourage you to carry out this exercise and let me know what you get. After you do this let me know if you still think fundamental data is so important for the prediction of financial time series.
Data that has had significant effect in the past.
You can do linear regression on economic data(the ones with values) and extract the effect that it produced and feed that to a predictor as a parameter.
A study can be done on the types of news, like the FF orange and red
  • #418
  • May 18, 2015 7:19am May 18, 2015 7:19am
Quoting bruce_loco
Well Those you can't. What you know about these types of events is that immediately there are negative consequences on both currency and stock markets. To begin with they are local, then global. When I was doing the quant course in coursera the guy did a small correlation on how long it would take a stock to absorb the impact of lets say a department windup, a significant lawsuit, loss of competitive edge, stuff that would drag the stock down 10-20% and sometimes even more, and this you cannot predict, what you can predict is what to do next, which...
You remind me of when I first finished college (too long ago!). It seems that you have a lot of ideas about how things "should be" in machine learning but you still don't have any practical experience in the field. You are saying "a model cannot just be fitted straight on price action" but this is merely a hypothesis you are making based on how you think things "should be" according to the things you have learned in theory. What you should do now is try all these things so that you can see some of the differences between what is "supposed to work" and what actually does work. Building a profitable machine learning algorithm is much more than simply throwing a lot of data that is "supposed to be relevant" at the problem, it is much more difficult. Do some experiments and share your conclusions with us!
  • #419
  • May 18, 2015 7:24am May 18, 2015 7:24am
Quoting bruce_loco
{quote} Data that has had significant effect in the past. You can do linear regression on economic data(the ones with values) and extract the effect that it produced and feed that to a predictor as a parameter. A study can be done on the types of news, like the FF orange and red
I absolutely encourage you to try this and share your findings with us!
  • #420
  • May 18, 2015 8:44am May 18, 2015 8:44am
Parameters that generate positive results on the EURUSD are also not limited to a single set. You have several different combinations that yield good returns with good correlation coefficients. The graph below, for the A=10 GMT +1/+2, B=120, C=11, D=20 parameters using a stop of 40% of the ATR(20), gives us good results with a different distribution of returns.

Attached Image
3 traders viewing now