Machine Learning with algoTraderJo | Page 18 | 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

  • #341
  • May 4, 2015 2:59pm May 4, 2015 2:59pm
Hello algoTraderJo, glad to see you back
I am also following with great interest!

As I subscribed to the Asirikuy framework I wondered sometimes if algoTraderJo was Daniel himself since, if I am not mistaken, you left us with similar questions and ML achievements/results as realised in the community

There is a lot of valuable info in your thread as is in Daniels and the community forum - great! But the community seems to put more effort on the price action systems. I enjoy more the approach with ML and try to learn from you and reproduce your results.
Thank you for sharing!
When it comes to choosing ensambles I am still stuck with lots of bias (and drawdown )- same as with any trading system, choosing ensambles on the IS often does not result in good OOS.

Cheers
KaBo
  • #342
  • May 4, 2015 3:19pm May 4, 2015 3:19pm
Quoting KaBo
Hello algoTraderJo, glad to see you back I am also following with great interest! As I subscribed to the Asirikuy framework I wondered sometimes if algoTraderJo was Daniel himself since, if I am not mistaken, you left us with similar questions and ML achievements/results as realised in the community There is a lot of valuable info in your thread as is in Daniels and the community forum - great! But the community seems to put more effort on the price action systems. I enjoy more the approach with ML and try to learn from you...
Yes, Daniel and the Asirikuy community in general are much more focused on the production of price action based systems. I am working with Daniel on some machine learning setups and this is mainly the reason why I resumed posting here (Daniel contacted me a few days ago to resume our work together and posting here helps me get focused on the stuff). I also hope that the community moves more in favor of machine learning but to be truthful I am far more advanced in my own developments than what is shared there (I have developed some large personal F4 libraries for ML I/O functions and the like). I do try to use F4 available functions to the whole community on my posts here so that you folks can reproduce. I am interested here in building a sort of path that can help you advance along the same lines I have.

Ensembles are tricky. What you want to reduce curve-fitting bias is to use as much data as you can, when I have developed systems in 20 or more years of data I often find that ensembles work out equally well in a considerable real OS (new live market conditions). You can use the Asirikuy 1987 data set for this, I find it works quite well and matches my other 1987-2015 data set nicely. If you want share with me what you are doing more specifically and I will try to give more practical advice.
  • #343
  • May 4, 2015 3:23pm May 4, 2015 3:23pm
Nice thread!

I did my dissertation that I submitted 2 weeks ago on Do Machine Learning Techniques Add Any Value Over Traditional FX Trading Strategies, and I explored OLS, elastic net, random forest, GB regression trees in a portfolio approach over 30 years of daily aggregated into weekly data. I'm an econs major so it's a bit of a stretch .

Just curious, have you tried looking at the returns compared to a benchmark? For example, no predictability/carry trade benchmark where you simply buy if the interest rate differential is positive and sell if the interest rate differential is negative using the same rules as you have now. How will that perform and how much does your strategies outperform that benchmark?
Build good relationships with others.
  • #344
  • May 4, 2015 3:24pm May 4, 2015 3:24pm
Going back to the EUR/JPY machine learning strategy I just posted, here is what happens when you change the number of inputs:

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_1h1d_step_eurjpy_diffinputs.png Size: 37 KB


What does this tell you? Any comments?
  • #345
  • May 4, 2015 3:34pm May 4, 2015 3:34pm
Quoting algoTraderJo
Going back to the EUR/JPY machine learning strategy I just posted, here is what happens when you change the number of inputs: {image} What does this tell you? Any comments?
What was the 9th input? Adding it over the 8 input model not only improved the returns significantly but also reduced the drawdown by a huge amount.
Build good relationships with others.
  • #346
  • May 4, 2015 3:43pm May 4, 2015 3:43pm
It's good to see you back on this thread. You have posted details on the inputs and target outputs so that the rest of us can try to reproduce, but you have provided little detail on what you actually do to "build a machine learning model using linear regression". More details on this aspect would be appreciated to at least get me thinking in the right direction.

Quoting algoTraderJo
Long time no see folks! Life happens, we get busy, but now I am back! I want to start posting again on this thread to get the ball rolling again on machine learning in FX trading. To start things fresh I want to start sharing with you a machine learning technique that is between the 1D and 1H timeframes (since we use 1H data but trade only once per day), that is quite successful in generating systems for most FX trading pairs. This machine learning technique does the following: On each day, at a given trading hour (A) build a machine learning model...
  • #347
  • May 4, 2015 3:47pm May 4, 2015 3:47pm
Quoting shellsnail
{quote} What was the 9th input? Adding it over the 8 input model not only improved the returns significantly but also reduced the drawdown by a huge amount.
It is an additional bar return. The 8 input curve uses 8 bar returns on each example, the 9 input curve uses 9 bar returns, etc. The difference between the 8 and 9 input curves is simply the use of one additional bar return for the building of examples.
  • #348
  • May 4, 2015 3:52pm May 4, 2015 3:52pm
Quoting beachloup
It's good to see you back on this thread. You have posted details on the inputs and target outputs so that the rest of us can try to reproduce, but you have provided little detail on what you actually do to "build a machine learning model using linear regression". More details on this aspect would be appreciated to at least get me thinking in the right direction. {quote}
I use the Shark machine learning library which is implemented as part of the F4 framework. I use the linear regression shark class as explained here. The Shark library is completely free and open source so you can implement this within your project if you wish and build your models using the exact same machine learning. Here is the function in F4:

Inserted Code
double LR_Prediction_i_simpleReturn_o_mfemaeDifference(int learningPeriod, int barsUsed, int tradingHour, int barThreshold)
{
 
    RegressionDataset dataset  = regression_i_simpleReturn_o_mfemaeDifference(learningPeriod, barsUsed, tradingHour, barThreshold);
    LinearModel<> model;
    LinearRegression trainer;
    trainer.train(model, dataset);
    Data<RealVector> testOutput;
    RegressionDataset datasetInput = p_regression_i_simpleReturn(barsUsed);
    testOutput = model(datasetInput.inputs());
 
    if (testOutput.element(0)[0] > testOutput.element(0)[1]) return(1);
    if (testOutput.element(0)[1] > testOutput.element(0)[0]) return(-1);
    return(0);
}

The regression_i_simpleReturn_o_mfemaeDifference and p_regression_i_simpleReturn are the functions used to build the input/target structures but the whole machine learning happens in this part:

Inserted Code
LinearModel<> model;
LinearRegression trainer;
trainer.train(model, dataset);

The shark library is really compact in its machine learning function calls, that's one of the reasons why I like it so much.
  • #349
  • May 4, 2015 3:55pm May 4, 2015 3:55pm
Best result is: We take the last 12 hours returns to predict the next 12 hours move - seems logical

The next few days, I will go back and look over my systems and results - I know I had quite a few questions back then - especially about mixing ensambles.
  • #350
  • May 4, 2015 4:10pm May 4, 2015 4:10pm
Quoting KaBo
Best result is: We take the last 12 hours returns to predict the next 12 hours move - seems logical The next few days, I will go back and look over my systems and results - I know I had quite a few questions back then - especially about mixing ensambles.
The observation about the 12 hour relationship is interesting indeed Is the 12 hour target horizon and the best 12 input result a coincidence or something of more substance? How would you test this to find out?

Looking forward to your questions on ensembles.
  • #351
  • May 4, 2015 4:30pm May 4, 2015 4:30pm
Time is also an incredibly critical factor... Try to train/predict at the wrong hour and you get much worse results (remember that examples are also built using previous days at the same hour). It is no coincidence that the best results for this strategy are at a specific hour around the Asian open (remember that this is EURJPY).

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_1h1d_step_eurjpy_hour.png Size: 30 KB


There is a volatility cycle caused by market opening and closing times, a cycle that makes it different to predict at hour 0, 1, 4, 15, etc. Any effort to make predictions in the lower timeframes must account for time. Time, is critical. Time is information.
  • #352
  • May 5, 2015 1:44am May 5, 2015 1:44am
Hi AlgoTraderJo

Can you post the data preparation function for this strategy.
  • #353
  • May 5, 2015 5:00am May 5, 2015 5:00am
Ok, after looking at the shark library I still am a bit lost on how your model works. I understand that using the input data you generate a regression line that gives you a slope and an intercept, but how does that apply to the MLE and the MSE?
  • #354
  • May 5, 2015 6:57am May 5, 2015 6:57am
Happy to see Joe is back.

Quoting beachloup
Ok, after looking at the shark library I still am a bit lost on how your model works. I understand that using the input data you generate a regression line that gives you a slope and an intercept, but how does that apply to the MLE and the MSE?
You don't regress the samples to get a line. Each sample is made of C=9 returns. It is a 9 dimension vector. You have B=130 such vectors. The target is made of two values (mle and mse). It is a 2 dimension vector. You also have 130 of them. By the mean of least square regression you search (9+1)x2 coefficients. The plus one is for the intercept.

Then you use the last 9 returns to forecast:
Inserted Code
(mle, mse) ~ (L0 + L1 * return_1 + L2 * return_2 + ... + L9 * return_9, S0 + S1 * return_1 + S2 * return_2 + ... + S9 * return_9)
No greed. No fear. Just maths.
  • #355
  • May 5, 2015 8:00am May 5, 2015 8:00am
Quoting PipMeUp
Happy to see Joe is back. {quote} You don't regress the samples to get a line. Each sample is made of C=9 returns. It is a 9 dimension vector. You have B=130 such vectors. The target is made of two values (mle and mse). It is a 2 dimension vector. You also have 130 of them. By the mean of least square regression you search (9+1)x2 coefficients. The plus one is for the intercept. Then you use the last 9 returns to forecast: (mle, mse) ~ (L0 + L1 * return_1 + L2 * return_2 + ... + L9 * return_9, S0 + S1 * return_1 + S2 * return_2 + ... + S9 * return_9)...
Couldn't have said it better
  • #356
  • May 5, 2015 8:33am May 5, 2015 8:33am
Quoting levkly
Hi AlgoTraderJo Can you post the data preparation function for this strategy.
Here is the code, note that F4 functions are used for getting price values, etc.

Inserted Code
RegressionDataset regression_i_simpleReturn_o_mfemaeDifference(int period, int barsUsed, int hourFilter, int successBarThreshold){
    Data<RealVector> inputs(period,RealVector(barsUsed));
    Data<RealVector> labels(period,RealVector(2));
    double mle = 0;
    double mse =  0;
    struct tm timeInfo;
    int tl=successBarThreshold;
    int i = 0, j = 0, n=0, m = 0;

    while(m<period){
        safe_gmtime(&timeInfo, openTime(i+tl));
        if (timeInfo.tm_hour == hourFilter){        
        mle = 0;
        mse =  0;
        for(n=0;n<tl;n++){        
            if((high(i+tl-n)-cOpen(i+tl) > mle)){
                mle = (high(i+tl-n)-cOpen(i+tl)) ;
            }
            if((cOpen(i+tl)-low(i+tl-n) > mse)){
                mse = (cOpen(i+tl)-low(i+tl-n)) ;
            }
        }
 
         for(j=0;j<barsUsed;j++){
         inputs.element(m)[j] = (cOpen(tl+j+i)-cOpen(1+tl+j+i))/cOpen(1+tl+j+i);
         }
 
         labels.element(m)[0] = mle/cOpen(tl+i);
         labels.element(m)[1] = mse/cOpen(tl+i);
         m++;
        }
        i++;        
    }    
    RegressionDataset dataset(inputs,labels);
    return dataset;
}
  • #357
  • May 5, 2015 9:18am May 5, 2015 9:18am
Ok thanks, I think I now have it. I am basically searching for the best 10 coefficients that when applied to every vector give the smallest target error. I do the same for mse and mle and end up with 20 coefficients. Once found I apply those to the current data to make the next prediction. Thanks I have it now, seems like an ideal GA optimisation problem (if I am to roll my own solution), look's like it is time for some gpu coding..

Quoting PipMeUp
Happy to see Joe is back. {quote} You don't regress the samples to get a line. Each sample is made of C=9 returns. It is a 9 dimension vector. You have B=130 such vectors. The target is made of two values (mle and mse). It is a 2 dimension vector. You also have 130 of them. By the mean of least square regression you search (9+1)x2 coefficients. The plus one is for the intercept. Then you use the last 9 returns to forecast: (mle, mse) ~ (L0 + L1 * return_1 + L2 * return_2 + ... + L9 * return_9, S0 + S1 * return_1 + S2 * return_2 + ... + S9 * return_9)...
  • #358
  • May 5, 2015 9:48am May 5, 2015 9:48am
Quoting beachloup
Ok thanks, I think I now have it. I am basically searching for the best 10 coefficients that when applied to every vector give the smallest target error. I do the same for mse and mle and end up with 20 coefficients. Once found I apply those to the current data to make the next prediction. Thanks I have it now, seems like an ideal GA optimisation problem (if I am to roll my own solution), look's like it is time for some gpu coding.. {quote}
No GA is needed. This has got a closed form solution. In matrix notation:
Attached Image

Where q is the query vector (the last 9 returns), S is the samples matrix (130x10), T is the target matrix(130x2), ŷ is the forecast.

EDIT: q and S must have their first column set to 1 for the intercept.
No greed. No fear. Just maths.
  • #359
  • May 5, 2015 9:59am May 5, 2015 9:59am
I was not really able to reproduce the results. I could obtain a positive return but nothing as great as the equity curve posted. Jo, what are you using for the spread? the real ask-bid or some fixed value? If you use 3 pips like in the previous posts it can explain the difference. 0:00 GMT+1/+2 is low liquidity and the spread was much bigger in 2007-2010 than it is now. The spreads in my dataset are more like 4 pips and can go as high as 12 pips.

I also have a couple of questions:
1- How did you determined the values for A, B, C and D? visually, using some cross-validation or something else?
2- It happens that the forecast for MLE/MSE I get is such that the predicted high is lower than the predicted low. Do you take care of this?
No greed. No fear. Just maths.
  • #360
  • May 5, 2015 11:50am May 5, 2015 11:50am
Quoting PipMeUp
I was not really able to reproduce the results. I could obtain a positive return but nothing as great as the equity curve posted. Jo, what are you using for the spread? the real ask-bid or some fixed value? If you use 3 pips like in the previous posts it can explain the difference. 0:00 GMT+1/+2 is low liquidity and the spread was much bigger in 2007-2010 than it is now. The spreads in my dataset are more like 4 pips and can go as high as 12 pips. I also have a couple of questions: 1- How did you determined the values for A, B, C and D? visually,...
I am using 7 pips as the spread. As you point out the spread on this pair can be bigger, so I used 7 as a compromise of the historical average. I also did a small simulation using tick data (2007-2014), results are very similar. Now to answer your questions:

1- I did a genetic optimization of the variables selecting for high R2. This has worked great for me when the amount of possible parameter combinations is too great. Since it's done over such a long period I also do not get any significant increase in curve-fitting bias due to this process. As far as my experience goes things have continued to work well in live market conditions after selecting in this manner. Note that my process for live trading selection is more complex, there is a lot on data-mining bias we're yet to discuss here.

2- This never happens for me. I predict a return MLE and MSE so they are not price values that can be higher/lower in an illogical manner. My predictions are, for example, an MLE of 0.2% and an MSE of 0.1%, there is no way in which the MLE can be lower than the MSE in an inconsistent way. Structuring the outputs in this way is crucial for reproducing my results.
1 trader viewing now