Machine Learning with algoTraderJo | Page 39 | 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

  • #761
  • Nov 2, 2015 6:57am Nov 2, 2015 6:57am
Another question for you Jo, you enter long if :

the long return is >0 and short return<0
or long return > short return ? (they can be both negative)


Thanks
sam
  • #762
  • Edited 1:37pm Nov 2, 2015 8:06am | Edited 1:37pm
Quoting algoTraderJo
{quote} You can probably go down to 15M before any of these problems become too prevalent with the Asirikuy data. Providing your SL and TP (if any) values are large enough then you should not face any significant issues until you go to something like 5M, where it becomes hellish to develop something that works well across multiple brokers because of these problems. Also using more data in terms of "more bars" does not reduce curve fitting bias. The amount of market conditions inside your test depends on the number of ticks that were needed to construct...
Thanks for the reply.

I did some research last year (before joining Asirikuy) using software designed by a friend. I used SVM, NN and various 'simple' ML models.

Inputs:

The inputs were derived from the currency strength calculations found in the many 'currency strength meters' online. So instead of using simple returns, each element of the input vector was the difference between the two geometric means of the returns of each pair consisting of each of the base and quote currency, within a basket of the 8 most traded currencies. Let me clarify with an example:

Suppose we are calculating the input value of the previous AUDUSD candle, which we call input[1]. Instead of the simple return (Open[0] - Open[1])/Open[1], we compute the following:

We define ret(AUDUSD[1]) to be the simple return of the previous AUDUSD candle. Similar for other pairs.
The basket of currencies is EUR, GBP, AUD, NZD, CAD, USD, CHF, JPY.

AUD[1] = (-ret(EURAUD[1]) * -ret(GBPAUD[1]) * ret(AUDNZD[1]) * ret(AUDCAD[1]) * ret(AUDUSD[1]) * ret(AUDCHF[1]) * ret(AUDJPY[1])) ^ (1/7)

USD[1] = (-ret(EURUSD[1]) * -ret(GBPUSD[1]) * -ret(AUDUSD[1]) * -ret(NZDUSD[1]) * -ret(CADUSD[1]) * ret(USDCHF[1]) * ret(USDJPY[1])) ^ (1/7)

(Of course you cannot strictly calculate AUD[1] or USD[1] since the products contain negative terms, but you can employ some tricks in order to do so. Alternatively, you can use the arithmetic mean instead of the geometric mean which makes the above a lot simpler, but in my experience the latter yields superior results.)

Then input[1] = AUD[1] - USD[1].

Each of these input[i] are then used to construct input vectors in exactly the same way as the simple returns in this thread - number of examples, input vector dimension, trading hour etc are all parameters.

Outputs:

If I remember correctly, the output was the trade outcome (profit or loss) when using a chandelier exit (for a long trade the chandelier exit is the highest high since trade open minus a multiple of the ATR[20]). Of course, trades are only entered when a favourable trade outcome is predicted, and the ATR multiplier is a parameter. I'm not sure about the time period (horizon) over which the trade outcome was considered.

Results:

The first thing to note is that the software I used to conduct this research was nowhere near as reliable as F4, and the historical data was not great to say the least (and only went back to 2002). Therefore I take these results with a boulder of salt
The second is that I have unfortunately lost the files which documented the results of these experiments, so I am just recalling from memory.
In cases where data was not available for a certain pair (such as AUDJPY), the pair was skipped and the calculation adjusted accordingly.
Constant retraining was always done, and 1H timeframe was used.
DMB was measured in the same way as within Asirikuy, except for the fact that only about 10 runs were done due to the immense computational cost of loading data from all these pairs. I did experiment with only loading a few pairs and then creating proxies for each of the pairs in the calculation, but can't remember the results. eg Close(AUDJPY[1]) ≈ Close(AUDUSD[1]) * Close(USDJPY[1]).

There was not a lot of difference amongst the different ML models - they all gave roughly the same number of high quality systems for each pair, and the DMB didn't vary too much either. Hence I gave up on NNs and SVMs in favour of the simpler models.

Contrary to your findings (or at least from what I can tell from this thread), using this input structure did yield better results than when using simple returns in nearly every instance.

With regards to the results for AUD pairs, very few profitable systems were found on the real data and hence DMB was high. The same was true for CAD pairs (I think USDCAD was the only one tested) and some GBP pairs (not sure which ones, but GBPUSD did have DMB < 0.01 for some searches).

I apologise that these results are very hand-wavy, but it's the best I can do from memory. Also, this year has been extremely busy, and so I haven't ported this strategy over to F4. I may do so in the future as it did show some promise, although it is very resource intensive! If there does come a time when measuring DMB becomes more feasible then I will be sure to post my findings here.

All the best.
  • #763
  • Nov 3, 2015 4:52am Nov 3, 2015 4:52am
Quoting ForexSamsam
Another question for you Jo, you enter long if : the long return is >0 and short return<0 or long return > short return ? (they can be both negative) Thanks sam
You need both to happen. Long prediction > Short Prediction & Long prediction > 0 for a long entry and viceversa for shorts
  • #764
  • Nov 6, 2015 3:34pm Nov 6, 2015 3:34pm
I am impressed by all the progress in the cloud enhanced machine learning in Asirikuy. Daniel recently released another 20 or so low DMB systems from these mining experiments. The portfolio has 6 systems on the USD/JPY and 28 on the EUR/USD, all trading on the 1H timeframe using ML algos. I just made a backtest of the full 34 system portfolio with a 0.029% risk per trade per system. It took me a good while to get everything setup but the results in testing are superb.

Attached Image


I definitely enjoy how adding systems improves your Sharpe so fast! The drawdowns become much lower as you add more strategies. This is the big value of trading portfolios, especially bigger and bigger portfolios where correlations between the systems are low. The portfolios I trade are significantly larger than this but they are certainly adding systems quicker than I did when I started doing this. The power of collaborating it seems!
  • #765
  • Nov 6, 2015 3:37pm Nov 6, 2015 3:37pm
Quoting beachboy44
{quote} Thanks for the reply. I did some research last year (before joining Asirikuy) using software designed by a friend. I used SVM, NN and various 'simple' ML models. Inputs: The inputs were derived from the currency strength calculations found in the many 'currency strength meters' online. So instead of using simple returns, each element of the input vector was the difference between the two geometric means of the returns of each pair consisting of each of the base and quote currency, within a basket of the 8 most traded currencies. Let me clarify...
This is interesting. I have tried indexes as inputs before but I have never managed to get them to work better than simple returns. I think Daniel has posted on the Asirikuy forum with a similar experience when using indexes. However I didn't do the pow(1/n) you are doing so that might make some difference. I am going to try it out mate, perhaps you have a jewel here I will post with my findings.
  • #766
  • Nov 6, 2015 4:14pm Nov 6, 2015 4:14pm
Quoting algoTraderJo
I am impressed by all the progress in the cloud enhanced machine learning in Asirikuy. Daniel recently released another 20 or so low DMB systems from these mining experiments. The portfolio has 6 systems on the USD/JPY and 28 on the EUR/USD, all trading on the 1H timeframe using ML algos. I just made a backtest of the full 34 system portfolio with a 0.029% risk per trade per system. It took me a good while to get everything setup but the results in testing are superb. {image} I definitely enjoy how adding systems improves your Sharpe so fast! The...
Hi AlgoTraderJo,
very nice thread , congratulations for all your works done here and also with Daniel

I'm also an old asirikuy member , indeed they are doing a very good job with cloud mining projects , just think that also for pkantu (price based systems) cloud mining reached around 1000 systems with low data-mining bias with high linearity (R2)!

I've never applied myself to ML and related stuff mostly because I lack the underlying math notions even if i'm quite comfortable with programming (mostly C#/Mql), i have to start seriously to study it , I guess pdf materials posted at the beginning of this thread could be a good starting point to grasp basic concept.
Btw I've also backtested portfolio as you did but only using 6 USDJPY ML systems derived from asirikuy cloud mining , I used your R script posted back here (thx!) :

Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot2.png Size: 37 KB


Note that my result are far higher than your simply because I forgot to change risk and I used 1% for each systems (so 6% total portfolio risk ).
You can image latest orders in backtest comes with lot size around 100.000... a little bit unrealistic to be placed in live trading

Will follow your thread with renewed interest

Cheers,
Skyline
  • #767
  • Nov 6, 2015 4:50pm Nov 6, 2015 4:50pm
Quoting algoTraderJo
{quote} This is interesting. I have tried indexes as inputs before but I have never managed to get them to work better than simple returns. I think Daniel has posted on the Asirikuy forum with a similar experience when using indexes. However I didn't do the pow(1/n) you are doing so that might make some difference. I am going to try it out mate, perhaps you have a jewel here I will post with my findings.
Great! It will be interesting to see whether or not my results had any underlying truth; like I said before, pretty much everything about my previous setup was suboptimal.

Just in case you find yourself having to compute the geometric mean for negative numbers, here's what I did:
Calculate the geometric mean (GM) for the positive and (modulus of) negative terms separately, and then compute the arithmetic mean of GM(+) and -1*GM(-).

Suppose we are calculating k = (x * y * u * v * w)^(1/5) where x,y > 0 and u,v,w < 0.

Then k ≈ ((x * y)^(1/2) - (-u * -v * -w)^(1/3))/2
  • #768
  • Nov 6, 2015 5:05pm Nov 6, 2015 5:05pm
Just being nosy, and a new dog trying to find a bone. A broker from a certain platform suggested me join and learn from here before I lose anymore on the binary business. I'm so new to this that my hairs still wet. can i please catch a tip from anyone?
  • #769
  • Nov 6, 2015 5:40pm Nov 6, 2015 5:40pm
Quoting quickdrw45
Just being nosy, and a new dog trying to find a bone. A broker from a certain platform suggested me join and learn from here before I lose anymore on the binary business. I'm so new to this that my hairs still wet. can i please catch a tip from anyone?
This thread is probably a bit too advanced if you are just starting out.

http://mechanicalforex.com has a lot of great information, and takes a very scientific approach to trading the forex market. I'm not saying that this is the only way to trade, but it may help you to avoid some of the most common pitfalls. Go back a few years and there should be some slightly more 'basic' articles describing the underlying techniques we are using in this thread. Also bear in my mind that we are doing algorithmic trading - designing and developing software to trade the markets.

But in any case, my advice would be to stay away from binary options. In my opinion it is just gambling, so you might as well go to a casino and have fun losing money I wouldn't trust those unscrupulous binary brokers!
  • #770
  • Nov 7, 2015 5:57am Nov 7, 2015 5:57am
Quoting skyline
{quote} Hi AlgoTraderJo, very nice thread , congratulations for all your works done here and also with Daniel I'm also an old asirikuy member , indeed they are doing a very good job with cloud mining projects , just think that also for pkantu (price based systems) cloud mining reached around 1000 systems with low data-mining bias with high linearity (R2)! I've never applied myself to ML and related stuff mostly because I lack the underlying math notions even if i'm quite comfortable with programming (mostly C#/Mql), i have to start seriously...
OMG definitely 100,000 lots is a bit unrealistic When doing these portfolios I always like to keep total risk at 1-2% to give some more realistic simulations. But in any case the Sharpe ratio and other risk adjusted statistics should be very similar regardless of how much risk you use in the portfolio. In the latest newsletter Daniel said he's working on some scripts to automate the backtesting of the ML systems, hopefully this makes everything easier to setup!

I'm also very interested in the pkantu systems and their results, it's just amazing to be able to trade sooo many systems on live accounts! I trade 100-500 system portfolios (from my privately mined collection) and the results do start to get incredible as you go up and up the ladder.

If you're interested in learning ML then you cannot go wrong with the coursera machine learning courses. The Stanford course in ML is awesome.
  • #771
  • Nov 7, 2015 7:35am Nov 7, 2015 7:35am
I now want to talk about neural networks. Up to this point we have explored simpler machine learning algos like linear regression and we have only briefly explored SVMs, NNs and random forests (when we were doing daily data experiments if you remember well). We have seen that linear regression algos are good enough to get profitable results on many different pairs but there are still many pairs where we cannot find satisfactory solutions and other pairs where we would want to diversify to systems that attack non-linear relationships. Neural networks can give us an ability to find non-linear relationships, this means that we can get to find profitable results in pairs where we haven't up until now and even find different less correlated results on pairs where I have already showed successful linear regression algos.

The problem is of course that the neural networks can be very computationally expensive meaning that NNs can cause difficulties when evaluating things like DMB. Can we get around this problem? We can try! Below you can find a function for making predictions using NNs in F4 that takes advantage of the functions that we have already used to create the datasets for the linear algos:

Inserted Code
double NN_Prediction_i_simpleReturn_o_tradeOutcome(int learningPeriod, int barsUsed, int tradingHour, int BE, double initial_SL, int frontier, double minStop, int dsl_type)
{
    //get problem data
    RegressionDataset dataset = regression_i_simpleReturn_o_tradeOutcome(learningPeriod, barsUsed, tradingHour, BE, initial_SL, frontier, minStop, dsl_type);
    FFNet<FastSigmoidNeuron,FastSigmoidNeuron> network;
    unsigned numInput=barsUsed;
    unsigned numHidden=(int)((barsUsed+2)/2);
    unsigned numOutput=2;
    unsigned numberOfSteps=50;
    unsigned step;
    network.setStructure(numInput, numHidden, numOutput);
    initRandomUniform(network,-0.5,0.5);
    SquaredLoss<> loss;
    ErrorFunction error(dataset, &network,&loss);
    RpropMinus optimizer;
    optimizer.init(error);
    for(step = 0; step < numberOfSteps; ++step){
      optimizer.step(error);
    }
    network.setParameterVector(optimizer.solution().point); // set weights to weights found by learning
    Data<RealVector> testOutput;
    RegressionDataset datasetInput = p_regression_i_simpleReturn(barsUsed);
    testOutput = network(datasetInput.inputs());
    if (testOutput.element(0)[0] > 0 && testOutput.element(0)[1] < 0) return(1);
    if (testOutput.element(0)[1] > 0 && testOutput.element(0)[0] < 0) return(-1);
    return(0);
}

It makes some simplifications to make things far faster than a regular NN:

  1. Training is very restricted (just 50 iterations, very little for an NN)
  2. Used the RPropMinus optimizer, the cheapest available in Shark for an NN.
  3. Number of hidden neurons is the average between input and output neurons, a good rule of thumb.
  4. You want to keep the number of inputs (barsUsed) to 10 or less. Go higher and it will take ages to run.

This simplified NN approach allows us to perform retraining on every trading decision (as with all the ML experiments I have showcased) while we get some ability to capture non-linear relationships (see we're using sigmoid functions for the hidden and output layers). If we run the optimizations we would hope to find something that is able to generate good results using this restricted training and cheap optimizer. If you want to use it and you have access to F4 you can simply add it to your AsirikuyMachineLearningCWrapper.cpp file (also the header for the definition) and you can then replace the linear regression function in Munay with this one.

I am gonna show you some experiments on the EUR/USD first so that you can see what we can get with this. You will also see what happens when we increase or decrease the number of training iterations. So... We're into NN territory now! Brace yourselves

  • #772
  • Nov 7, 2015 7:55am Nov 7, 2015 7:55am
Quoting quickdrw45
Just being nosy, and a new dog trying to find a bone. A broker from a certain platform suggested me join and learn from here before I lose anymore on the binary business. I'm so new to this that my hairs still wet. can i please catch a tip from anyone?
Daniel's blog can be a bit overwhelming to the beginner. Try quantstart as well, the guy is more organized than Daniel and has lots of insightful articles. I also fully agree with beachboy44, binary options are terrible IMHO.
  • #773
  • Nov 7, 2015 8:50am Nov 7, 2015 8:50am
These are some initial sample EUR/USD results using the before mentioned NN function for predictions, the trailing stop as explained before and the input and output structure used for previous linear regression 1H strategies. As in all cases the ML algo is retrained before making each trading decision. These are the parameters and results:

Inserted Code
SL = 160% ATR20
A = 8 GMT +1/+2
B = 130
C = 2
D = 115

Attached Image


I have put together results for different numbers of training iterations for the NN on the graph above. This means that when the NN is trained before each trading decision the NN weights are adjusted across this number of steps using the RpropMinus optimizer. If the number we had chosen was too low we would see a significant change when increasing the number of iterations but taking it from 50 to 100 shows that results only vary slightly. When we go to lower numbers for trading iterations we still get similar results with 20 but at 10 the NN completely breaks down and no longer has any predictive powers. As we would expect for an NN with predictive powers there is a critical threshold for training iterations.

With this we can see that the proposed fast-NN algo works and can generate systems with some suspected edge (DMB analysis reveals if it is spurious or not). The training is good enough despite being really constrained and the optimizer is also working properly. This is also only a quick sample, we can do much better. I will show you some more results soon.
  • #774
  • Nov 7, 2015 10:54am Nov 7, 2015 10:54am
Quoting algoTraderJo
{quote} OMG definitely 100,000 lots is a bit unrealistic When doing these portfolios I always like to keep total risk at 1-2% to give some more realistic simulations. But in any case the Sharpe ratio and other risk adjusted statistics should be very similar regardless of how much risk you use in the portfolio. In the latest newsletter Daniel said he's working on some scripts to automate the backtesting of the ML systems, hopefully this makes everything easier to setup! I'm also very interested in the pkantu systems and their results, it's just...
Indeed backtest should be always taken with a grain of salt , one thing is backest another thing is live trading even if the same methodology is applied.
Yes next week Daniel will release new python script (as he did for pkantu systems) to speed up ML backtest process , this is why I just stopped to backtest ML waiting for him

I saw you talked about that Stanford ML course it seems really interesting, I don't understand how it works, if there's the need to follow some live webinars or it could be done 'offline' , infact my biggest concern is how much time is needed to follow properly that kind of online course considering that I have a full time job

Thx,
Skyline
  • #775
  • Nov 7, 2015 1:20pm Nov 7, 2015 1:20pm
Another case on the EUR/USD where the number of returns used to build each example (C parameter) is 5 instead of 2. As in every case retraining of the NN is done before each trading decision.

Inserted Code
SL = 160% ATR20
A = 1 GMT +1/+2
B = 90
C = 5
D = 90

Attached Image (click to enlarge)
Click to Enlarge Name: machinelearning_1h_nn_eu_dti2.png Size: 98 KB


This time the results improve significantly when going from 50 to 75 and then remain fairly constant at 100. This means that the NN was not trained properly on average with only 50 iterations. Since we have more inputs per example in this case (5) it means that there are more weights and relationships can therefore be more complex. This translates into more iterations needed to properly train the NN. More inputs almost always means that more training iterations are demanded. The NN also deteriorates much faster at even lower iteration numbers, in this case just going down to 20 scrapes all predictive ability.
  • #776
  • Nov 7, 2015 1:24pm Nov 7, 2015 1:24pm
Quoting skyline
{quote} Indeed backtest should be always taken with a grain of salt , one thing is backest another thing is live trading even if the same methodology is applied. Yes next week Daniel will release new python script (as he did for pkantu systems) to speed up ML backtest process , this is why I just stopped to backtest ML waiting for him I saw you talked about that Stanford ML course it seems really interesting, I don't understand how it works, if there's the need to follow some live webinars or it could be done 'offline' , infact my biggest concern...
This course is actually taken "on your own time" now, also offline with all the lectures available when you want to watch them. So don't worry about it mate, you can just join and go through the lectures, exercises and homework as you see fit. I think they have people with full time jobs in mind when they design these courses. It's also not overly complicated so you'll definitely be able to handle it, you seem like a smart guy
  • #777
  • Nov 7, 2015 1:37pm Nov 7, 2015 1:37pm
We can also pay a 10x penalty in simulation speed and train for 1000 iterations per NN training, before every trading decision. This would heavily impair DMB analysis. Look how it affects the last posted system.

Attached Image


Is it worth it? You tell me if it is.
  • #778
  • Nov 7, 2015 2:41pm Nov 7, 2015 2:41pm
Quoting algoTraderJo
{quote} This course is actually taken "on your own time" now, also offline with all the lectures available when you want to watch them. So don't worry about it mate, you can just join and go through the lectures, exercises and homework as you see fit. I think they have people with full time jobs in mind when they design these courses. It's also not overly complicated so you'll definitely be able to handle it, you seem like a smart guy
Thx for infos and i'm an average guy not that smart at all
I'll definatelly give a try with that ML course , so that hopefully be able to follow your post here with a more solid ML background

Cheers,
Skyline
  • #779
  • Edited 4:47pm Nov 7, 2015 3:39pm | Edited 4:47pm
Quoting algoTraderJo
We can also pay a 10x penalty in simulation speed and train for 1000 iterations per NN training, before every trading decision. This would heavily impair DMB analysis. Look how it affects the last posted system. {image} Is it worth it? You tell me if it is.
I would say it is not worth it at all considering the amount of time it takes to measure DMB for NN systems - even more so if the system is going to be traded within a large, uncorrelated portfolio.

Just out of interest, have you done any experiments with images as inputs (such as in Daniel's 'Tapuy' system)? Do they often outperform simple returns, and if so, are they worth the extra computational effort (both from processing the images and from the added parameters)? There is something 'satisfying' about having a neural network (brain) processing data from images (eyes) - almost like a human trader, but my experience with the markets has been that preconceptions don't seem to match reality.
  • #780
  • Nov 8, 2015 6:18am Nov 8, 2015 6:18am
Quoting beachboy44
{quote} I would say it is not worth it at all considering the amount of time it takes to measure DMB for NN systems - even more so if the system is going to be traded within a large, uncorrelated portfolio. Just out of interest, have you done any experiments with images as inputs (such as in Daniel's 'Tapuy' system)? Do they often outperform simple returns, and if so, are they worth the extra computational effort (both from processing the images and from the added parameters)? There is something 'satisfying' about having a neural network (brain)...
I tend to agree with that. Even if you have some improvement in Sharpe ratio with more training it's not worth the additional time because the DMB analysis would be impractical. With the NN training 50 iterations it takes about 2x the time of a linear regression algo, which makes it feasible for system searches and DMB analysis.

To answer your other question, I can't imagine how expensive it must be to do the DMB of this image based systems. Such systems are by their very nature complex and slow, imagine having to search parameters that vary the image resolution and all other NN related variables. I can't see it being feasible. Have you researched any systems like this?
2 traders viewing now