OP joined the Forum like 30mins ago, he sounds alot like EurUsd hmmmnnn
I am what Many Dream to be but only a few can achieve, im a part of the 1%
- #21
- Dec 8, 2014 4:32pm Dec 8, 2014 4:32pm
- Joined Sep 2012 | Status: Edge,Phsycology And Money Managemen | 1,011 Posts
- #22
- Edited Dec 9, 2014 2:24pm Dec 8, 2014 5:10pm | Edited Dec 9, 2014 2:24pm
- Joined Dec 2014 | Status: Trader | 412 Posts
wow!!!!!!!! subscribed! where do you get the technology to conduct these tests and modules?
However if you want you can code the same machine learning systems using R but it gets very complicated when you want to perform accurate tests on different forex data sets with different time zones or when you want to change your strategies to live execution. Alternatively you can also code something from scratch that interfaces FANN/Shark/Waffles with MT4 or an API, I personally found it better to pay for a community supported solution rather than go at it myself from zero, but of course a person who is younger than me or has a lot of additional time may choose to go for a completely from-scratch-in-house solution.
OP joined the Forum like 30mins ago, he sounds alot like EurUsd hmmmnnn:thinkin g:
Sorry to say I am not EurUsd
. I have been reading ForexFactory for a while but finally found the time/energy to join and start posting, especially since now I have some free time and would like to exchange ideas with like minded traders. I have found too little online about machine learning, especially about practical true trading applications and therefore I wanted to see if anyone else is working on something along these lines (and try to pass the little I know to those who might happen to be interested).
This model that I mentioned: if TP comes first =+1, if SL comes first =0, could also be modeled using logistic regression, but with what predictor variables ? I personally don't know
We can now take our K-NN model and attempt to add more information to see if it improves results. Inputs can be as varied as you want and the effect of the inputs can be determined by comparing your system's overall trading results. When you add inputs that are irrelevant they make the prediction of values harder because they only add noise or redundancy to your predictions. The following image shows you a comparison of the original system that uses the direction of the past 2 candles as inputs and a system modification that uses the direction of the past 2 candles plus whether the RSI(10) was above 50 (1) or below 50 (0) for those two candles. Adding these two RSI inputs to our ML model makes things worse, therefore they are not useful.
People in academic literature tend to measure predictor value by looking at prediction efficacy (in this case whether we can predict a bearish or bullish bar) but in this case this makes little sense because our profit is mainly associated with the distribution of our prediction rather than their accuracy. Therefore we are more interested in when we are right than how often we are right. As many on this thread have pointed we can create models that target the prediction of variables that are more related with our actual profit goals (trade exits) which we will move into after we continue our exploration of daily candle direction predictions (which I like as a beginning and simple exercise).
People in academic literature tend to measure predictor value by looking at prediction efficacy (in this case whether we can predict a bearish or bullish bar) but in this case this makes little sense because our profit is mainly associated with the distribution of our prediction rather than their accuracy. Therefore we are more interested in when we are right than how often we are right. As many on this thread have pointed we can create models that target the prediction of variables that are more related with our actual profit goals (trade exits) which we will move into after we continue our exploration of daily candle direction predictions (which I like as a beginning and simple exercise).
When you add new information that contributes with predictive power you actually get better results. The graph below shows you our K-NN test using 2 bar directions as input and a second test using 18 bars as inputs. As you can see the addition of 16 new inputs adds enough information to improve results altogether.
It is also interesting to see how the distribution of accuracy in predictions has changed significantly and our two systems take opposite directions through several parts of the simulation. This means that we may be able to obtain an improvement in overall results if we combine different sets of machine learning strategies where the distribution of accuracy is different. Since both systems are expected to have edges in prediction but both are accurate at different times a portfolio trading both systems might yield even better results. Indeed, splitting risk in half and doing a portfolio simulation gives the following result:
The portfolio indeed has a behavior that is now between both systems, performance is smoother than for the two separate strategies as the variance between both of them is compensated.
It is also interesting to see how the distribution of accuracy in predictions has changed significantly and our two systems take opposite directions through several parts of the simulation. This means that we may be able to obtain an improvement in overall results if we combine different sets of machine learning strategies where the distribution of accuracy is different. Since both systems are expected to have edges in prediction but both are accurate at different times a portfolio trading both systems might yield even better results. Indeed, splitting risk in half and doing a portfolio simulation gives the following result:
The portfolio indeed has a behavior that is now between both systems, performance is smoother than for the two separate strategies as the variance between both of them is compensated.
{quote} Oh yes, you can attempt to train to predict trade exit targets. Which is something very interesting as well (this is somewhat how some of the strategies I trade live actually work). I will continue exploring the case of daily bar predictions to get a little more into the world of machine learning and then we'll move into things along these lines
I use genetic programming , decision trees(random forest etc.) svm and self training NN.My project is still in development phase.But out of sample results with GP is very encouraging.It took me couple of years with GP to find very good results( > 85% winnig rate with low drawdown).So inputs are very important.Dont expect to get very good results just using RSI ,MACD etc.. as inputs.
There are some very good free open source libraries :scikit in python: http://scikit-learn.org/stable/ lib c++ : http://dlib.net/ml.html you will find some very fast SVM algos in c++.
Regarding KNN ,I read that they suffer from curse of dimensionality in high-dimensional data .A Russian prof. suggest Cosine Similarity method for finding the nearest neighbors to avoid curse of dimensionality.
There are lot of hype surrending the Deep learning nets.I read that they are slow to train and have only little effect on financial time series. My experience is (only halfway through ) ensemble methods are the best methods for financial time series.
- #37
- Dec 9, 2014 8:49am Dec 9, 2014 8:49am
- Joined Sep 2012 | Status: Edge,Phsycology And Money Managemen | 1,011 Posts
I am what Many Dream to be but only a few can achieve, im a part of the 1%
If we continue to explore our K-NN algorithm and attempt to find the optimum amount of information for this model we find that using 4 bars as input gives the best overall result (remember that we are always rebuilding the model on each bar using recent past examples). The strategy has a very linear log(balance) Vs time curve which talks about the stability of the algorithm in successfully predicting larger days Vs smaller days.
It is however immediately noticeable that the strategy faced a kind of regime change in 2011, when the strategy changed from having a drawdown below 20% for more than 20 years to go to a 35% drawdown before starting to recover. Although the strategy seems close to making a new equity high in 2014, it seems obvious that such a long drawdown period would make trading unattractive. It is therefore important for us to improve this strategy further in order to ensure that the 2011-2014 period does not become flat/losing but rather shows the same profit as the balance curve before.
How can we achieve this? The K-NN model definitely cannot classify this region accurately without a serious deterioration of previous results (which makes us doubt the results due to the probability that good results in 2011-2014 are just due to chance). We therefore have a few choices (I am assuming we won't be changing the prediction target for now): 
Using this we can create a portfolio that uses 90% of our K-NN strategy and 10% of our SVM strategy (the optimum weights that minimize variance) and we can then generate a portfolio that reaches a drawdown just slightly above 20% in 2014 but does reach a new equity high later on during the year. We now have a model that gives decent stable performance through 25 years of different market conditions and does not crash during the past two years, although it was difficult for the models to predict daily bar directionality for large bars accurately during this period. I think we can do even better! What do you think ?
It is however immediately noticeable that the strategy faced a kind of regime change in 2011, when the strategy changed from having a drawdown below 20% for more than 20 years to go to a 35% drawdown before starting to recover. Although the strategy seems close to making a new equity high in 2014, it seems obvious that such a long drawdown period would make trading unattractive. It is therefore important for us to improve this strategy further in order to ensure that the 2011-2014 period does not become flat/losing but rather shows the same profit as the balance curve before.
How can we achieve this? The K-NN model definitely cannot classify this region accurately without a serious deterioration of previous results (which makes us doubt the results due to the probability that good results in 2011-2014 are just due to chance). We therefore have a few choices (I am assuming we won't be changing the prediction target for now):
- Explore other inputs (see if we can get more information that allows for a better prediction from other sources)
- Explore other models (see if other models have a distribution of accurate results that compensates that of the K-NN)
We can use other type of classifiers to compliment the K-NN (besides the linear mapper we tested at first). We can attempt to use a Support Vector Machine (SVM) to compliment the trading of our K-NN. An SVM model built with a C=10 and gamma=0.01 gives the result showed below in black. You can see clearly that although the SVM does suffer from the same problem near the end of the test, this model does reach an equity high at a much later date than the K-NN strategy.
Using this we can create a portfolio that uses 90% of our K-NN strategy and 10% of our SVM strategy (the optimum weights that minimize variance) and we can then generate a portfolio that reaches a drawdown just slightly above 20% in 2014 but does reach a new equity high later on during the year. We now have a model that gives decent stable performance through 25 years of different market conditions and does not crash during the past two years, although it was difficult for the models to predict daily bar directionality for large bars accurately during this period. I think we can do even better! What do you think ?
