Machine Learning with algoTraderJo | Page 42 | 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

  • #821
  • Jan 14, 2016 8:14am Jan 14, 2016 8:14am
I'm gonna show you a simple initial approach to the "pattern detection" idea. I will use the open of the last 7 bars to define a "pattern" which will be the relationships between these values. I then look into the past to see if there are enough examples to train my machine learning algo. This is a function to perform this task, it returns 1 if there are enough examples in the historical data, 0 if there are not enough:

Inserted Code
int is_viable_examplesByStructure(int examplesNeeded, int frontier){
    int tl = frontier;
    double open1, open2, open3, open4, open5, open6, open7;
    
    open1 = iOpen(0,1);
    open2 = iOpen(0,2);
    open3 = iOpen(0,3);
    open4 = iOpen(0,4);
    open5 = iOpen(0,5);
    open6 = iOpen(0,6);
    open7 = iOpen(0,7);
   int i = 0, j = 0, n=0, m = 0;
   while(m<examplesNeeded && i+tl+500<iBars(0)){
        
      if (
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+2)) == (open1 > open2)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+3)) == (open1 > open3)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+4)) == (open1 > open4)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+5)) == (open1 > open5)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+6)) == (open1 > open6)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+7)) == (open1 > open7)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+3)) == (open2 > open3)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+4)) == (open2 > open4)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+5)) == (open2 > open5)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+6)) == (open2 > open6)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+7)) == (open2 > open7)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+4)) == (open3 > open4)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+5)) == (open3 > open5)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+6)) == (open3 > open6)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+7)) == (open3 > open7)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+5)) == (open4 > open5)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+6)) == (open4 > open6)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+7)) == (open4 > open7)) &&
           ((iOpen(0,i+tl+5) > iOpen(0,i+tl+6)) == (open5 > open6)) &&
           ((iOpen(0,i+tl+5) > iOpen(0,i+tl+7)) == (open5 > open7)) &&
           ((iOpen(0,i+tl+6) > iOpen(0,i+tl+7)) == (open6 > open7))
           )
           {      
       m++;
      }
      i++;
   }
   
   if (m<examplesNeeded){ return(0) ;} else { return(1); }
}

If there are enough examples I then build the input/output array using the matching patterns I have located in the past. In this case the inputs are a past number of returns and the outputs are the outcomes of long and short trades managed with a simple trailing stop where the SL is moved to the starting SL value distance from the current open if this stop is more favorable than the present stop:

Inserted Code
RegressionDataset regression_i_simpleReturn_o_tradeOutcome_examplesByStructure(int period, int barsUsed, double initial_SL, int frontier, double minStop){
   Data<RealVector> inputs(period,RealVector(barsUsed));
   Data<RealVector> labels(period,RealVector(2));
    double openPrice;
    double SL_long, SL_short, new_SL;
    double atr;
    int tl = frontier;
    double open1, open2, open3, open4, open5, open6, open7;
    
    open1 = iOpen(0,1);
    open2 = iOpen(0,2);
    open3 = iOpen(0,3);
    open4 = iOpen(0,4);
    open5 = iOpen(0,5);
    open6 = iOpen(0,6);
    open7 = iOpen(0,7);
    
   int i = 0, j = 0, n=0, m = 0;
   while(m<period){
        
      if (
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+2)) == (open1 > open2)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+3)) == (open1 > open3)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+4)) == (open1 > open4)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+5)) == (open1 > open5)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+6)) == (open1 > open6)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+7)) == (open1 > open7)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+3)) == (open2 > open3)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+4)) == (open2 > open4)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+5)) == (open2 > open5)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+6)) == (open2 > open6)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+7)) == (open2 > open7)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+4)) == (open3 > open4)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+5)) == (open3 > open5)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+6)) == (open3 > open6)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+7)) == (open3 > open7)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+5)) == (open4 > open5)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+6)) == (open4 > open6)) &&
           ((iOpen(0,i+tl+4) > iOpen(0,i+tl+7)) == (open4 > open7)) &&
           ((iOpen(0,i+tl+5) > iOpen(0,i+tl+6)) == (open5 > open6)) &&
           ((iOpen(0,i+tl+5) > iOpen(0,i+tl+7)) == (open5 > open7)) &&
           ((iOpen(0,i+tl+6) > iOpen(0,i+tl+7)) == (open6 > open7))
           ){     

        atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl+1);
        openPrice = cOpen(i+tl);
        SL_long = openPrice+spread()-atr*initial_SL;
        SL_short = openPrice+atr*initial_SL;
        
        n = 0;
        
        while ((SL_long < low(i+tl-n)) && (n<tl)){
            n += 1;
            atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl-n+1);
                
            new_SL  = cOpen(i+tl-n)+spread() - (initial_SL*atr);
            if(new_SL <= SL_long){
                new_SL = cOpen(i+tl-n);
            }
            
          if (cOpen(i+tl-n)-new_SL > minStop){
              SL_long = new_SL;
            }
        }
        
        labels.element(m)[0] = SL_long-(openPrice+spread());
        
        n = 0;
        
        while ((SL_short > high(i+tl-n)+spread()) && (n<tl)){
            n += 1;
            atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl-n+1);
            new_SL  = cOpen(i+tl-n) + (initial_SL*atr);
            if(new_SL >= SL_short){
                new_SL = cOpen(i+tl-n);
            }
          if (new_SL-cOpen(i+tl-n) > minStop){
              SL_short = new_SL;
            }
        }
              
        labels.element(m)[1] = openPrice-SL_short;
        
       for(j=0;j<barsUsed;j++){
       inputs.element(m)[j] = (iOpen(0,tl+j+i)-iOpen(0,1+tl+j+i))/iOpen(0,1+tl+j+i);
        }
       m++;
      }
      i++;
   }
   
   RegressionDataset dataset(inputs,labels);
   return dataset;
}

With these two functions I can build the machine learning functions. This time I decided to use a neural network as the machine learning algo:

Inserted Code
double NN_Prediction_i_simpleReturn_o_tradeOutcome_examplesByStructure(int learningPeriod, int barsUsed, double initial_SL, int frontier, double minStop, int trainingEpochs)
{
    
    int is_viable = is_viable_examplesByStructure(learningPeriod, frontier);
    
    if (is_viable == 0) return(0);
    
    RegressionDataset dataset = regression_i_simpleReturn_o_tradeOutcome_examplesByStructure(learningPeriod, barsUsed, initial_SL, frontier, minStop);
   
    FFNet<FastSigmoidNeuron,FastSigmoidNeuron> network;
    unsigned numInput=barsUsed;
    unsigned numHidden=(int)((barsUsed+2)/2);
    unsigned numOutput=2;
    unsigned numberOfSteps=trainingEpochs;
    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);
}

In this case the algorithm is retrained and relaunched on every hourly bar where there are enough examples. I then performed an optimization to see if I could find anything worth it on the EUR/USD 1H and found some interesting stuff. Here is a linear system example I found:

starting SL = 150% ATR20
B = 75 (number of examples used per training)
C = 8 (number of input returns used per example)
D = 185 (frontier for the trade output)

Attached Image


This algo does not use any hourly filters on the 1H chart, for this reason it's very different from everything I have shared so far on the 1H. Thoughts, always appreciated...
  • #822
  • Jan 14, 2016 12:19pm Jan 14, 2016 12:19pm
Quoting algoTraderJo
I'm gonna show you a simple initial approach to the "pattern detection" idea. I will use the open of the last 7 bars to define a "pattern" which will be the relationships between these values. I then look into the past to see if there are enough examples to train my machine learning algo. This is a function to perform this task, it returns 1 if there are enough examples in the historical data, 0 if there are not enough: int is_viable_examplesByStructure(int examplesNeeded, int frontier){ int tl = frontier; double open1, open2, open3, open4, open5,...
Wow, this opens up a whole new avenue for 1H ML systems!

I suppose the first question we should ask is how much more computationally demanding this method is than building sequential examples at specific hours. We no longer have the trading hour parameter but we now have to (potentially) retrain on every hour. There must also be some overhead from the pattern finding. Couldn't we store the bar index for each valid example from within the is_viable_examplesByStructure function? Then we don't need to find the pattern again in the dataset creation function.

Another thing is that you have introduced an extra parameter - the number of bars used to define the pattern. Since the fundamental reason for searching for these patterns in the first place is to find examples that are more similar to the current one, it seems reasonable to define each pattern using 'x = barsUsed' bars.

We could also simplify the pattern search procedure by only comparing adjacent open values - so we only train examples where each return has the same sign as the one in the example for the current bar. Though if I had to hazard a guess, I would say that this would yield far inferior results.

It might also be possible to combine the 'trading hour' method with the 'similar example' method in a couple of ways:
1) Retrain on each hour by searching for similar examples sequentially,
2) Retrain on each hour by searching for similar examples on the trading hour only.

We could run into problems with number 2) as we may have to go too far in the past (or may not be able) to find enough similar examples.

Thanks a lot for sharing your code! I was actually going to implement this tomorrow, so you've saved me a lot of effort Much appreciated.
  • #823
  • Jan 15, 2016 2:20am Jan 15, 2016 2:20am
I don't want to spoil the euphoria about having a new approach, but aren't the net results factor 5 worse than even the simplest ML experiments before?
By which metric do you value the success of the experiment? And at what point or value do you continue researching the approach?
  • #824
  • Jan 15, 2016 4:51am Jan 15, 2016 4:51am
Quoting KaBo
I don't want to spoil the euphoria about having a new approach, but aren't the net results factor 5 worse than even the simplest ML experiments before? By which metric do you value the success of the experiment? And at what point or value do you continue researching the approach?
You are right that the example shown here is nothing to write home about, but I'm always interested in looking for extra layers of diversification. From what I understand, having these 1H systems only trade at a specific hour was vital to their success. Hence the mere fact that we can obtain linear results over the last 28 years is significant, and so - for me at least - this warrants further investigation.

Besides, I don't think that the results from one experiment are indicative of what can be obtained by building examples in this way in general. It may be the case that this is a near-optimal system when using this method, but it may be the case that when using a different search space, different pair, different inputs, different outputs, different ML models etc we may be able to improve our results tremendously. How would we know without testing?

I also wonder how DMB will be affected - all the more reason to continue to testing in my opinion. But don't worry KaBo, my "euphoria" won't be spoiled quite so easily
  • #825
  • Jan 15, 2016 5:40pm Jan 15, 2016 5:40pm
read read and read http://papers.ssrn.com/sol3/papers.c...act_id=2326253
  • #826
  • Jan 21, 2016 4:46pm Jan 21, 2016 4:46pm
Quoting surfeur
read read and read http://papers.ssrn.com/sol3/papers.c...act_id=2326253
Measure the probability of over fitting at different amounts of data used using the method on the paper. For machine learning systems that always retrain (this is key) you will see that this value becomes smaller and smaller as you use more and more data. The best strategy is to use all available data to create the systems, this makes perfect sense: the more data you use the more market conditions you can face.
  • #827
  • Jan 21, 2016 4:53pm Jan 21, 2016 4:53pm
Quoting KaBo
I don't want to spoil the euphoria about having a new approach, but aren't the net results factor 5 worse than even the simplest ML experiments before? By which metric do you value the success of the experiment? And at what point or value do you continue researching the approach?
I am not using this thread to show you the "best possible system", I am using it to expose new ideas that can help you expand your number of potential ML systems. The pattern based inputs I showed are simply a quick demonstration of the viability of using patterns like these as inputs. They are very interesting because you do not have to filter by hour which is new on the thread. They can yield much better systems as well (this is also not the best possible system, better performance is still possible using pattern filters):

Attached Image


When I post never assume that what I post is the "best" that you can achieve with that method, I often have little time so most often I just post quick examples to show viability. As I mentioned within the beginning of the thread I do not intend to share my best setups here I just want to demonstrate concepts and think about new ideas with whomever might be interested.
  • #828
  • Jan 21, 2016 5:01pm Jan 21, 2016 5:01pm
Quoting beachboy44
{quote} Wow, this opens up a whole new avenue for 1H ML systems! I suppose the first question we should ask is how much more computationally demanding this method is than building sequential examples at specific hours. We no longer have the trading hour parameter but we now have to (potentially) retrain on every hour. There must also be some overhead from the pattern finding. Couldn't we store the bar index for each valid example from within the is_viable_examplesByStructure function? Then we don't need to find the pattern again in the dataset creation...
Combining the pattern filters with the hour filters is actually what I was going to continue with. You can do this and you can get good results and one very interesting consequence is that you reduce the number of examples needed per retraining significantly. This is one of the first things showing that having "good examples" can be more important than having "many examples". Sample searches are also not extremely costly, especially if you're just doing prototyping and want to find what avenue is really worth pursuing. When you move to DMB measurements you can use some tricks but given that the ending example range is small it's not that much of a problem either way.
  • #829
  • Jan 21, 2016 6:55pm Jan 21, 2016 6:55pm
You can find an example of using pattern filters + hour filtering below. This is the code I used for the creation of the examples:

Inserted Code
RegressionDataset regression_i_simpleReturn_o_tradeOutcome_examplesByStructure(int period, int barsUsed, int hourFilter, double initial_SL, int frontier, double minStop){
   Data<RealVector> inputs(period,RealVector(barsUsed));
   Data<RealVector> labels(period,RealVector(2));
    double openPrice;
    double SL_long, SL_short, new_SL;
    double atr;
    int tl = frontier;
    int hour;
    double open1, open2, open3, open4;
    
    open1 = iOpen(0,1);
    open2 = iOpen(0,2);
    open3 = iOpen(0,3);
    open4 = iOpen(0,4);
    
   int i = 0, j = 0, n=0, m = 0;
   while(m<period){
      
      hour = getHourShift(PRIMARY_RATES, i+tl) ;
      
      if (
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+2)) == (open1 > open2)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+3)) == (open1 > open3)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+4)) == (open1 > open4)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+3)) == (open2 > open3)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+4)) == (open2 > open4)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+4)) == (open3 > open4)) &&
           hour == hourFilter
           ){    
 
        atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl+1);
        openPrice = cOpen(i+tl);
        SL_long = openPrice+spread()-atr*initial_SL;
        SL_short = openPrice+atr*initial_SL;
        
        n = 0;
        
        while ((SL_long < low(i+tl-n)) && (n<tl)){
            n += 1;
            atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl-n+1);
                
            new_SL  = cOpen(i+tl-n)+spread() - (initial_SL*atr);
            if(new_SL <= SL_long){
                new_SL = cOpen(i+tl-n);
            }
            
          if (cOpen(i+tl-n)-new_SL > minStop){
              SL_long = new_SL;
            }
        }
        
        labels.element(m)[0] = SL_long-(openPrice+spread());
        
        n = 0;
        
        while ((SL_short > high(i+tl-n)+spread()) && (n<tl)){
            n += 1;
            atr = iAtrWholeDaysSimpleShift(PRIMARY_RATES, 20, i+tl-n+1);
            new_SL  = cOpen(i+tl-n) + (initial_SL*atr);
            if(new_SL >= SL_short){
                new_SL = cOpen(i+tl-n);
            }
          if (new_SL-cOpen(i+tl-n) > minStop){
              SL_short = new_SL;
            }
        }
              
        labels.element(m)[1] = openPrice-SL_short;
        
       for(j=0;j<barsUsed;j++){
// remember to test Daniel's heiken ashi suggestion
       inputs.element(m)[j] = (iOpen(0,tl+j+i)-iOpen(0,1+tl+j+i))/iOpen(0,1+tl+j+i);
        }
       m++;
      }
      i++;
   }
  
   RegressionDataset dataset(inputs,labels);
   return dataset;
}

I also modified the function that checks the number of valid examples, adding the hour filter as well:

Inserted Code
int is_viable_examplesByStructure(int period, int hourFilter, int frontier){
    int tl = frontier;
    double open1, open2, open3, open4;
    int hour;
    
    open1 = iOpen(0,1);
    open2 = iOpen(0,2);
    open3 = iOpen(0,3);
    open4 = iOpen(0,4);
   int i = 0, j = 0, n=0, m = 0;
   while(m<period && i+tl+500<iBars(0)){
    
       hour = getHourShift(PRIMARY_RATES, i+tl) ;
        
      if (
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+2)) == (open1 > open2)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+3)) == (open1 > open3)) &&
           ((iOpen(0,i+tl+1) > iOpen(0,i+tl+4)) == (open1 > open4)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+3)) == (open2 > open3)) &&
           ((iOpen(0,i+tl+2) > iOpen(0,i+tl+4)) == (open2 > open4)) &&
           ((iOpen(0,i+tl+3) > iOpen(0,i+tl+4)) == (open3 > open4)) &&
           hour == hourFilter
           )
           {      
       m++;
      }
      i++;
   }
  
   if (m<period){ return(0) ;} else { return(1); }
}

If you see I have changed the pattern detection to a less complex setup (checking only the past 4 bars instead of 7) because the hour filtering + a 7 bar patter based filter would be too demanding. As on every post the ML algo retrains on every filtered hour, before making the trade decision. You can also change it to the value of C as beachboy44 suggested and you'll see this can lead to better results when searching spaces. The input/output structure is as described on the last post. The setup is as follows:

Inserted Code
SL 150% ATR20
B = 32 (examples used per training)
C = 3 (bar return inputs used per example)
D = 250  (frontier for the trade output)

Attached Image


As I said to beachboy44 before I want you to realize that the number of examples used is much smaller than for my previous algos. Far fewer are needed to obtain highly linear results. This means that we can also be successful if we consider only similar examples instead of retraining blindly with everything that might be available in the past. Establish the concept of training using only examples that are in some manner similar to what you currently have and you will improve.
  • #830
  • Jan 21, 2016 8:14pm Jan 21, 2016 8:14pm
Quoting algoTraderJo
You can find an example of using pattern filters + hour filtering below. This is the code I used for the creation of the examples: RegressionDataset regression_i_simpleReturn_o_tradeOutcome_examplesByStructure(int period, int barsUsed, int hourFilter, double initial_SL, int frontier, double minStop){ Data<RealVector> inputs(period,RealVector(barsUsed)); Data<RealVector> labels(period,RealVector(2)); double openPrice; double SL_long, SL_short, new_SL; double atr; int tl = frontier; int hour; double open1, open2, open3, open4; open1 = iOpen(0,1);...
It will be interesting to see if building examples in this way can reduce the number required when using a multiple trade outcome mechanism. For larger variation spaces, most of the highly linear systems I have found have a learning period of between 350-400 which is not ideal - especially for non-linear models. I'll let you know how it goes

Thanks again for posting this stuff!
  • #831
  • Jan 22, 2016 5:59am Jan 22, 2016 5:59am
Quoting algoTraderJo
{quote} They can yield much better systems as well (this is also not the best possible system, better performance is still possible using pattern filters):
Could you please poste the setup / code to the post 827? Thanks!
  • #832
  • Jan 22, 2016 3:16pm Jan 22, 2016 3:16pm
Quoting algoTraderJo
Take a look at this example on the EURUSD 30M timeframe. I am using the last described trailing stop that involves no additional parameters. The machine learning algo is a linear regression which is trained after every trading decision, as I explained in past posts. This one uses the same A,B,C,D parameter meaning I used in the past too. To the advantage of those who are Asirikuy members this trailing stop has now been implemented in the official ML system release in F4, so I also attach the configuration and parameters I used to run this test as...
Hi algoTraderJo,

I derived the ATR20 values from your results, but it seems that this time it varies much greatly compared to the ATR20 derived from the previous results you posted. Could it be possible that you were using a different timeframe (instead of daily) for the computation this time?

Just trying to understand how I can reproduce the results using the free data available online. Thanks so much for the inspiring posts.
  • #833
  • Jan 22, 2016 6:35pm Jan 22, 2016 6:35pm
I haven't followed this thread for a while but I came across this course at udacity today. Maybe some of you will be interested.

https://www.udacity.com/course/deep-learning--ud730
  • #834
  • Edited 2:45pm Jan 28, 2016 9:56am | Edited 2:45pm
I'm pretty new to this field.

In addition to the deep learning class mentioned by Pip_Monk (I started the class a couple days ago.... I would prefer R, but it is in python TensorFlow..... It assumes you have knowledge of Machine Learning. I don't but it has enough practical examples to keep me intereseted.) there is an introduction to machine learning here:

Intro to Machine Learning

I am doing this one below (you probably need an account to see this link). This class (I believe) is part of a nano-degree path:

https://www.udacity.com/course/viewe...37/m-640579189

Also there is a free book "Introduction to Statistical Learing" which uses R. The data sets, code and book can be found here:

www.StatLearning.com

Stanford lectures with slides and youtube videos for the book can be found here:

In-depth introduction to machine learning in 15 hours of ..

Ok, there are two free books. This is more advanced:

The Elements of Statistical Learning
  • #835
  • Jan 28, 2016 2:11pm Jan 28, 2016 2:11pm
this is fresh off the press : Google AI algorithm masters ancient game of Go
http://www.nature.com/news/google-ai...-of-go-1.19234

I thought members of this thread will be interested in this development. Even though it is not trading related directly, it is in the same domain we are discussing here.
Think : Trade : Live Life
  • #836
  • Jan 30, 2016 11:34am Jan 30, 2016 11:34am
Quoting revival_fx
I'm pretty new to this field. In addition to the deep learning class mentioned by Pip_Monk (I started the class a couple days ago.... I would prefer R, but it is in python TensorFlow..... It assumes you have knowledge of Machine Learning. I don't but it has enough practical examples to keep me intereseted.) there is an introduction to machine learning here: Intro to Machine Learning I am doing this one below (you probably need an account to see this link). This class (I believe) is part of a nano-degree...

Even better:

CS 7646: Machine Learning for Trading free at Udacity
  • #837
  • Feb 9, 2016 4:41pm Feb 9, 2016 4:41pm
On the matter of data mining bias, I'd like to draw the board's attention to https://m2.icm.edu.pl/boruta/ and a blog about it here -> http://danielhomola.com/2015/05/08/b...ection-method/. It would seem that this approach would be a valuable tool to have.
  • #838
  • Feb 11, 2016 5:37pm Feb 11, 2016 5:37pm
Quoting algoTraderJo
You can find an example of using pattern filters + hour filtering below. This is the code I used for the creation of the examples: RegressionDataset regression_i_simpleReturn_o_tradeOutcome_examplesByStructure(int period, int barsUsed, int hourFilter, double initial_SL, int frontier, double minStop){ Data<RealVector> inputs(period,RealVector(barsUsed)); Data<RealVector> labels(period,RealVector(2)); double openPrice; double SL_long, SL_short, new_SL; double atr; int tl = frontier; int hour; double open1, open2, open3, open4; open1 = iOpen(0,1);...
Hi algoTraderJo,

I would like to simulate your trade results more accurately, but I just have a small question. For the Christmas and New Year days, how do you handle the hourly data the day before and after these days if they are weekdays? Do you treat these holidays like weekends, where you cut the data after the 5pm the day before, and the data before 4am the day after? I think how you cut the data is crucial to the result of at least some algorithms (e.g. linear regression).

Thanks for your clarification.
  • #839
  • Feb 17, 2016 11:46pm Feb 17, 2016 11:46pm
Quoting algoTraderJo
{quote} I have talked about many ML algos, some using MLE/MSE but many using other outputs. The latest NN algos do not use the MLE/MSE they use the hypothetical result of trades taken with a trailing stop. Read those posts again if you want to get a better grasp about how the inputs/outputs are generated. You can use the MLE/MSE but you will get worse results IMHO. Accuracy is also meaningless because accuracy does not mean profit, if you measure your algos in terms of prediction accuracy you will fail IMHO. Academics love to talk about accuracy...
Actually it is quite the opposite. Like what you did, I used linear regression to fit the hypothetical result of trades using previous returns. I also followed what you did and discard results where SL is not reached. Since there is no take profit in your system, most valid examples have negative returns. Naturally, linear regression would predict most trades as unprofitable. As a result, the number of trades is much less than the previous systems you introduced without using trailing SL. I am actually quite surprised that your GU results on 30M timeframe show so many trades.

Also, I compared my result with yours from a previous result you posted using linear regression to predict MLE and MSE in the following D hours. I think most of our trading decisions are the same (I am using data posted by PipMeUp and the data from 2010 to 2015 is very similar). But the result you posted seem to not trade when the predicted MLE is not different from MSE by some threshold. This caused your number of trades to be somewhat fewer than mine as you are out of the market somewhat longer.

If you could post the predictions made with the trade report, it would be very helpful to compare.
  • #840
  • Feb 22, 2016 2:46am Feb 22, 2016 2:46am
Has anyone tried market switching model for strategy filtering with hidden markov model or any other ML methods? http://blog.revolutionanalytics.com/...ov-models.html
1 trader viewing now