Machine Learning with algoTraderJo | Page 6 | 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

  • #101
  • Dec 10, 2014 3:55pm Dec 10, 2014 3:55pm
I have used the same input for GBPUSD USD/CHF pairs.I have used 50.000 bars for M30 time frame ,45.000 bars for out of sample test and 5000 bars for in sample test.I increased the commission to get less signals.My experience is it is always better use less inputs.Sometimes math formula gives better result than the neural network formula with GP.

GBPUSD

http://i61.tinypic.com/2uixfh1.png

USD/CHF

http://i62.tinypic.com/34hamit.png
  • #102
  • Dec 10, 2014 5:33pm Dec 10, 2014 5:33pm
Quoting olsen-yersen
I have used the same input for GBPUSD USD/CHF pairs.I have used 50.000 bars for M30 time frame ,45.000 bars for out of sample test and 5000 bars for in sample test.I increased the commission to get less signals.My experience is it is always better use less inputs.Sometimes math formula gives better result than the neural network formula with GP. GBPUSD {image} USD/CHF {image}
What software package are you using? It looks very interesting
  • #103
  • Dec 11, 2014 12:49am Dec 11, 2014 12:49am
I am using Chaos Hunter software.Inputs are more interesting than the software .
  • #104
  • Dec 11, 2014 1:09am Dec 11, 2014 1:09am
What inputs are you using? It looks very interesting
  • #105
  • Dec 11, 2014 2:20am Dec 11, 2014 2:20am
That is my secret receipe , it took me nearly 2 years WITH Chaos Hunter to find the good inputs.Now I want to slaughter it on different time frames and different pairs then appyly sensivity analysis on the data .My problem is I have always seen '' trading'' as hobby and I have to admit I have been lazy to learn coding.I have been working with freelance coders.So it is a tough task.
  • #106
  • Dec 11, 2014 3:06am Dec 11, 2014 3:06am
algoTraderJo posted a method using linear regression. It is perhaps interesting that I post some facts about this method people should be aware of. They aren't obvious when you just use off the shelf black boxes software. To keep the explanations simple I'll use the simplest regression, the line fit y = a.x + b
We are given a set of N training samples (Xi, Yi) and estimate the slope a and the intercept b. With this method we'll find an estimation of these coefficients. These estimations have a variance (they are random variables). If you open a book of stats or google a little bit you'll find the formula for their standard error.

Smallest variance isn't always the best solution.
The linear regression finds the line which minimizes the variance of the residuals. Therefore the estimated line will have a variance smaller than the true line. Financial data are very noisy so the probability if high that the line found is not the true line.

Don't trust the intercept.
By inspecting the formulas you can notice that the variance of the intercept is much bigger than the one of the other coefficient. We always get a poor estimation of the intercept. It is preferable that the trading rules don't rely on it. It is true for high dimensions too.

Beware the outlayers.
The variance of the slope decreases like the inverse of the sum of the squared differences of the inputs to the mean of the inputs. Since the linear regression minimizes this variance it will "over-trust" the outlayers. It makes its best to include outlayers in the model. A single outlayer can have a dramatic impact.

The outlayer isn't the furthest sample.
An obvious method to remove the outlayers is to remove the sample the furthest of the result line and regress again. But because the regression forces the line to stay close of all the samples we can find datasets such that the outlayer is closer to the line than a valid sample. By removing the furthest we make things worse.
No greed. No fear. Just maths.
  • #107
  • Dec 11, 2014 4:09am Dec 11, 2014 4:09am
Quoting olsen-yersen
I am using Chaos Hunter software.Inputs are more interesting than the software .
Thank you...
  • #108
  • Dec 11, 2014 7:29am Dec 11, 2014 7:29am
Quoting PipMeUp
algoTraderJo posted a method using linear regression. It is perhaps interesting that I post some facts about this method people should be aware of. They aren't obvious when you just use off the shelf black boxes software. To keep the explanations simple I'll use the simplest regression, the line fit y = a.x + b We are given a set of N training samples (Xi, Yi) and estimate the slope a and the intercept b. With this method we'll find an estimation of these coefficients. These estimations have a variance (they are random variables). If you open a...
A good graphical example of this is below.

Attached Image (click to enlarge)
Click to Enlarge Name: 990px-Anscombe's_quartet_3_svg.png Size: 50 KB
  • #109
  • Dec 11, 2014 8:13am Dec 11, 2014 8:13am
Yes except for the top right one. Here the problem is the model: the line tries to capture a linear relationship between X2 and Y2 where the relation is actually quadratic.

Another example would be when the samples are packed in a small region of their space. Any line fits equally bad as any other.
=> the samples must spread as much as possible across the range of possible values.
Attached Image
No greed. No fear. Just maths.
  • #110
  • Dec 11, 2014 8:27am Dec 11, 2014 8:27am
Has somebody been able to reproduce ATJo result on E/U daily? My result is always a loser strategy.
Of course I checked that if I feed the data in the excel file (post #99) I get the same predictions. I noticed that the week-end bars are absent from the file. I removed them as well but the result is still the same.

@ATJo: what is the exit condition to take profit? I tried stop-and-reverse, trailing the SL, setting a TP equal to the SL or twice the SL.... Nothing worked.

My data aren't the same: the returns I find are different from yours. But if it is the reason of the failure there is a major problem with the robustness of the method... The overall result of the regression is yet the same with the same under-estimation of the magnitude. Over the last 450 days the regression gets 56% of correct sign yet it isn't enough to beat the market.
No greed. No fear. Just maths.
  • #111
  • Edited 9:01am Dec 11, 2014 8:44am | Edited 9:01am
@PipMeUp I get the same losing results.

I normally keep quiet on FF and only appear when there is something interesting being discussed regarding machine learning. This thread is only a couple of pages long but I hope that it does not turn into yet another thread where results are presented with without any sort of feedback when others try to reproduce (eg threads by EURUSDd, gg53 etc etc).

@ATJo what is the purpose of this thread? Are you sharing thoughts, ideas or a system? Do you want other users to participate or are you simply going to present post after post without any sort of detail behind the methodologies used? If you are wanting other users to help improve you current implementation then you will need to post more details with each post. If you are simply sharing ideas without wanting any feedback, that is also fine, I'm always interested in hearing new thoughts and ideas.
  • #112
  • Dec 11, 2014 1:14pm Dec 11, 2014 1:14pm
Quoting PipMeUp
Has somebody been able to reproduce ATJo result on E/U daily? My result is always a loser strategy. Of course I checked that if I feed the data in the excel file (post #99) I get the same predictions. I noticed that the week-end bars are absent from the file. I removed them as well but the result is still the same. @ATJo: what is the exit condition to take profit? I tried stop-and-reverse, trailing the SL, setting a TP equal to the SL or twice the SL.... Nothing worked. My data aren't the same: the returns I find are different from yours. But if...
It is very important for the data to be GMT +1 non-DST and GMT +2 DST. It is also important for your data to start Monday at 4 and finish on Friday at 19. The timestamp alignment of the data is very important for any of these methods to give you the same results as I get. If your data has a different structure you will get something different (worse or better). Since hourly volatility is cyclical the structure of your daily candles really matters (where they close/open). A huge advantage I have found with the Asirikuy programming framework is that you can automatically correct any historical data to the GMT shift you wish and get whichever weekly opening/closing times you desire. This makes dealing with different live broker DST/GMT settings a breeze.

Here are some other details worth mentioning:

  1. A signal in the opposite direction means that a trade is closed and a new trade is opened in the new direction.
  2. A signal in the same direction resets the stoploss as if a trade was just opened (stoploss is 60% of the ATR20)
  3. A spread of 3 pips is used
  4. Swap is included (credited/debited at 17 every day, 3x on Wednesday). Swap is derived from historical interest rates (taken from Oanda I believe)

I attach the detailed trade log (100K USD starting balance) of the linear regression strategy in case you want to take a detailed look at the results. I have also posted the strategy code here so that those who are Asirikuy members can reproduce. The machine learning functions used are already a part of the default framework. I also include the set file and the configuration file needed to run this on the Asirikuy simulation software.

I believe the above should be enough to reproduce but feel free to ask any questions you might have.

Attached File(s)
File Type: txt detailed_trade_log.txt   451 KB | 810 downloads
File Type: zip machineLearningExample.zip   2 KB | 946 downloads
  • #113
  • Dec 11, 2014 1:25pm Dec 11, 2014 1:25pm
Quoting beachloup
@PipMeUp I get the same losing results. I normally keep quiet on FF and only appear when there is something interesting being discussed regarding machine learning. This thread is only a couple of pages long but I hope that it does not turn into yet another thread where results are presented with without any sort of feedback when others try to reproduce (eg threads by EURUSDd, gg53 etc etc). @ATJo what is the purpose of this thread? Are you sharing thoughts, ideas or a system? Do you want other users to participate or are you simply going to present...
The purpose of this thread is not crystal clear to me but I have some general ideas. I want it to be the following things:

  1. A journal to keep details about basic machine learning experiments I run so that I can take a look at them in the future easily (I am a bit disorganized)
  2. A place to attempt to teach and learn from others by asking questions about my experiments and getting their answers
  3. A place to get concrete, evidence supported ideas from others in machine learning

I don't want to it to be:

  1. A holy grail search
  2. A place where people expect full blown systems to just plug-and-play

In summary I want this to be a learning experience for everyone, searching for a better way to understand machine learning and what we can do with it.

I definitely want people to be able to reproduce the things I do and give them feedback, if I have the time to do so. I will do this to the extent that I am able to but since I use a specialized programming framework it may be hard for those who don't have access to it to do the things it does (doing data refactoring, using the same machine learning libraries, etc), however hopefully this won't constitute a significant obstacle for the thread.

  • #114
  • Dec 11, 2014 1:54pm Dec 11, 2014 1:54pm
Quote
It is very important for the data to be GMT +1 non-DST and GMT +2 DST.
I didn't at all though about this "details". I knew I would someday have to bother coding this GMT -> any TZ+DST.... :-(
As you're from the USA I suppose you refer to the US DST, right? Europe and US don't shift on the same week.
I'll try again take this into account. Thanks.
No greed. No fear. Just maths.
  • #115
  • Dec 11, 2014 4:00pm Dec 11, 2014 4:00pm
Quoting PipMeUp
{quote} I didn't at all though about this "details". I knew I would someday have to bother coding this GMT -> any TZ+DST.... :-( As you're from the USA I suppose you refer to the US DST, right? Europe and US don't shift on the same week. I'll try again take this into account. Thanks.
Yes, in FX the devil is really in the details. I use the European DST because most of my strategies depend on having the European market open accurately timed. Since the market opening times also have DST you will see that this also becomes important when we move to lower timeframe ML strategies. The lower timeframe volatility in fact changes as a function of DST as well. This also creates only a minor issue if you also trade US market open since although the DST changing weeks are different they are in fact very close together (so you have a one week offset which does not cause any issue).
  • #116
  • Dec 11, 2014 4:13pm Dec 11, 2014 4:13pm
We use a daily sampling therefore, as long as we sample at the exact same time (forgetting the DST days), we still have a 24 hour period. Each period of the trading session is represented. I mean if I sample every 1:37am GMT, the asian session, the London open, etc are all in there in the same "quantity". Just they aren't at the same offset. According to you why is this so important for this strategy? Especially because you use exclusively the open price of each candle. It sounds as if the very first tick would be the only one meaningful. BTW London and Frankfurt are 1h different TZ. GBP is the currency of the UK. What does your backtest of G/U look like if you shift you dataset by 1h?
No greed. No fear. Just maths.
  • #117
  • Dec 11, 2014 4:30pm Dec 11, 2014 4:30pm
Quoting PipMeUp
We use a daily sampling therefore, as long as we sample at the exact same time (forgetting the DST days), we still have a 24 hour period. Each period of the trading session is represented. I mean if I sample every 1:37am GMT, the asian session, the London open, etc are all in there in the same "quantity". Just they aren't at the same offset. According to you why is this so important for this strategy? Especially because you use exclusively the open price of each candle. It sounds as if the very first tick would be the only one meaningful. BTW London...
We will hopefully move into this issue of the daily starting time later on, you will see that the results change depending on how you actually start/end your daily candles. I believe that this is because the return of each day must encompass the return of certain consecutive sessions in order to work on a given pair. Remember that I am not looking simply at one open, but at the comparison of each open with the next (the return). If you use GMT 1|2 each day ends with the Asian session and you effectively have sessions in the order European->American->Asian if you start at other times your returns represent the returns of a different sequence of sessions.

True, I don't have a very satisfactory answer of why it makes such a huge difference, but it strongly changes results. This change is also not spurious, the results are not distributed randomly between losing/profitable as a function of the TZ but profitable outcomes cluster strongly at some daily candle structure (week start/end, GMT TZ).

The G/U simulations change significantly depending on how you shift the TZ, I will probably post these results as we move along the thread.
  • #118
  • Dec 12, 2014 1:19am Dec 12, 2014 1:19am
Quoting algoTraderJo
Here are two plots showing you the actual returns on the EURUSD and the returns that are predicted by the linear regression model at each point: {image} {image} As you can see the predictions made by the linear regression model fall in line with the increases/decreases of volatility of the actual data but the size of our predictions is actually lower. Since most daily bars do not go above 1% (0.01) the model almost never predicts changes of this magnitude. An issue with machine learning models is always that our returns depend largely on accurately...
I will test your prediction results tonight with chaos hunter to see if I can improve it further.
  • #119
  • Dec 12, 2014 1:39am Dec 12, 2014 1:39am
AlgotraderJo,

Can you post , actual data(eur/usd)/perdicted values in csv file as well ?
  • #120
  • Dec 12, 2014 5:48am Dec 12, 2014 5:48am
Quoting olsen-yersen
AlgotraderJo, Can you post , actual data(eur/usd)/perdicted values in csv file as well ?
I already posted a csv file containing actual/predicted data for the EUR/USD linear regression test.
1 trader viewing now