Machine Learning with algoTraderJo | Page 13 | 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

  • #241
  • Jan 9, 2015 2:22pm Jan 9, 2015 2:22pm
For those of you using the Asirikuy framework you can display your graphs in R instead by using this simple loading script (you need to have the PerformanceAnalytics library installed):

Inserted Code
library(PerformanceAnalytics)
temp<- read.zoo("C:/WhereyouHaveYourtest/results.txt", sep = ",",format="%d/%m/%Y %H:%M", header=TRUE,index.column=1,colClasses=c("NULL", "NULL", "NULL", "character", "NULL", "NULL", "NULL", "NULL", "NULL", "NULL", "numeric", "NULL", "NULL", "NULL"))
backtest<- as.xts(temp)                              
daily <-dailyReturn(backtest)
charts.PerformanceSummary(daily, ylog=TRUE)

This uses the output txt from the Asirikuy tester and generates graphs like the ones I post on this thread. I find this very useful because the graphs in R have the drawdown and return already there, makes it a hell lot easier to interpret.
  • #242
  • Jan 9, 2015 3:07pm Jan 9, 2015 3:07pm
I'm moving now into some more complex and insightful experiments about machine learning. Picture all the experiments we have done up until now, we have been using a model that retrains every day but we have been choosing the number of examples and the number of inputs in a quite arbitrary manner. We could think that the same number of examples/inputs/targets that is best changes as the test progresses, by choosing an arbitrary combination we are inevitably increasing data mining bias and we are missing an opportunity to "adapt" the building of our model to changing market conditions.

A way to solve this problem is to on each candle get the total profit and total loss in pips that the model would have had during the past 200 days (little less than one trading year) using a variety of different examples/inputs/targets and then choose the one that has had the highest value for a given performance statistic. The easiest one to calculate is the profit factor (total pips won/total pips lost). You can picture how this is computationally quite expensive because you need to:

  1. Test the performance of N models during the past 200 days
  2. For each test of model performance each model must be retrained for each one of those 200 days

This implies that we must perform 200*N model training cycles on each candle, instead of simply 1 (as we were doing before). This means that a single back-test can go from taking a bit less than 5 minutes, to taking well over 2 days.

The advantage is mainly that we have a model that is very dynamic, not only from the retraining point of view but from the actual model building point of view. The inputs, targets and examples that determine the training of the model can also change dynamically as the market evolves. You'll see that this process also gives us a large amount of information about how market conditions change with time and how our global prediction capacity increases/decreases on different markets.

I will be sharing some of the results of doing this soon.

  • #243
  • Jan 9, 2015 4:56pm Jan 9, 2015 4:56pm
Quoting algoTraderJo
For those of you using the Asirikuy framework you can display your graphs in R instead by using this simple loading script (you need to have the PerformanceAnalytics library installed): library(PerformanceAnalytics) temp<- read.zoo("C:/WhereyouHaveYourtest/results.txt", sep = ",",format="%d/%m/%Y %H:%M", header=TRUE,index.column=1,colClasses=c("NULL", "NULL", "NULL", "character", "NULL", "NULL", "NULL", "NULL", "NULL", "NULL", "numeric", "NULL", "NULL", "NULL")) backtest<- as.xts(temp) daily <-dailyReturn(backtest) charts.PerformanceSummary(daily,...
Althrough I get a bit different results. However I've got the tools ready for some deeper exploration. This is very interesting topic I have to say.

And the F4 rulez BTW.
Attached Image
  • #244
  • Jan 9, 2015 5:54pm Jan 9, 2015 5:54pm
Quoting th0th
{quote} Althrough I get a bit different results. However I've got the tools ready for some deeper exploration. This is very interesting topic I have to say. And the F4 rulez BTW. {image}
Cool Some differences are probably due to the feed you're using which is the Asirikuy data (I presume), it is difficult to compare when using different time periods as well (difficult graphically I mean). However make sure you use C=10 and gamma=0.001 as well as a changing gamma and C can also cause differences. In any case it seems that you're now ready to explore and post your results as well! Great to see that

The coolest thing with F4 is that the thing you just backtested is ready-to-use in live trading. Development-to-live trading is instantaneous while with R it was a nightmare!
  • #245
  • Jan 9, 2015 6:13pm Jan 9, 2015 6:13pm
Going back to the idea I exposed on my previous post, I have ran a test using the following idea. The SVM always uses the direction of return of the next 5 days as target and the number of examples and number of inputs (directions of previous bars (0,1)) used for training are selected from the model that has given the best results in terms of profit factor (PF defined as pips won/ pips lost), if it had been traded by retraining on each bar and taking a trade in the direction of prediction within the past 200 days. We test example numbers from 50 to 250 (20 example steps) and number of inputs from 2 to 20 (2 input steps). This gives us a total of 100 models that need to be tested across the past 200 bars. Each bar we then do the following:

 

  1. Evaluate each model by doing things as if it was a normal test (which means, retrain on everyone of the past 200 bars using the chosen number of inputs and examples and then use the prediction for the next bar to determine profit/loss).
  2. Select the model with the best PF. If the best PF is below 1 do not trade (we wouldn't want to trade if we are losing).
  3. Use that model's number of examples and inputs to retrain on the current bar and trade.

Here are the test results we obtain when doing this. These are the results of doing the above on all bars through the test (the power of the F4 framework!).

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


See that the results are worse than the best SVM model we had developed before (the 5 day model with 145 examples and 20 inputs used for training on each bar). Performance under the most recent market conditions is especially noticeable. Nonetheless I would like to draw attention to other information we can take from this run, which is how the optimum values for the number of examples, the number of inputs and the best model's profit factor change with time. You can see these graphs below:

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_svm_adapts1.png Size: 10 KB

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_svm_adapts2.png Size: 9 KB

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_svm_adapts3.png Size: 26 KB


For me it is very interesting to see that we often have the performance of the entire ensemble drop; the best model is a very poor performer regardless of the number of examples and inputs, which means that all possible variations of the model are under-performing under some given market conditions. We have many times drops below a PF of 1 (although they don't show on the graph as I filtered the log because I put it after the PF<1 filter (my mistake!)) but it is evident that the whole ensemble deteriorates under some conditions. From this we can tell many different things and the information gives some obvious ways to improve the results. I challenge you to give me your analysis on the results and how they tell us we can improve things further. Replies awaited!
  • #246
  • Jan 10, 2015 2:50am Jan 10, 2015 2:50am
Can you increase the PF threshold to say PF>1.2? If this indeed improves the results you can make the threshold value adaptative. Something simple like PF>SMA100(PF)
No greed. No fear. Just maths.
  • #247
  • Jan 10, 2015 5:36am Jan 10, 2015 5:36am
Quoting PipMeUp
Can you increase the PF threshold to say PF>1.2? If this indeed improves the results you can make the threshold value adaptative. Something simple like PF>SMA100(PF)
This is indeed a good point. A PF of 1.2 is the median of the PF distribution. I will upload this when I resume posting on Monday. Have a nice week-end!
  • #248
  • Jan 10, 2015 7:07am Jan 10, 2015 7:07am
Quoting algoTraderJo
... I challenge you to give me your analysis on the results and how they tell us we can improve things further. Replies awaited!
I'd go with applying SVM on mutliple timeframes. E.G. while there are some results on daily, I'd try to predict the direction of weekly and trade only when the direction of both timeframes are in the same direction. Or just increase the risk while there is consanancy on both timeframes.
  • #249
  • Edited 8:03am Jan 10, 2015 7:53am | Edited 8:03am
Quoting th0th
{quote} I'd go with applying SVM on mutliple timeframes. E.G. while there are some results on daily, I'd try to predict the direction of weekly and trade only when the direction of both timeframes are in the same direction. Or just increase the risk while there is consanancy on both timeframes.
The challenge is to come up with ways to improve the last results based (solely) on the information provided on the last post (I don't see how that information would lead you to consider that multiple timeframes could improve results). The exercise here is to analyse the results and come up with ways that the evidence shows could provide improvements (like the PF filter increase suggested by PipMeUp). See what conclusions you can come up with by analyzing information from only my last post !
  • #250
  • Jan 10, 2015 1:58pm Jan 10, 2015 1:58pm
Quoting algoTraderJo
We have many times drops below a PF of 1 (although they don't show on the graph as I filtered the log because I put it after the PF<1 filter (my mistake!))
I do not have access to Asirikuy framework. Would you please post a graph of comulative return vs. date such that the vetical axis is not log-scaled?
Trading is the hardest way to make easy money...
  • #251
  • Edited 2:34pm Jan 10, 2015 2:05pm | Edited 2:34pm
Quoting algoTraderJo
Going back to the idea I exposed on my previous post, I have ran a test using the following idea. The SVM always uses the direction of return of the next 5 days as target and the number of examples and number of inputs (directions of previous bars (0,1)) used for training are selected from the model that has given the best results in terms of profit factor (PF defined as pips won/ pips lost), if it had been traded by retraining on each bar and taking a trade in the direction...
something likes my master-degree graphs (more than 20 years ago, my one is about correlation and is horizontal-mirrored).
I hope my supervisor is not in this forums.
  • #252
  • Jan 12, 2015 6:23am Jan 12, 2015 6:23am
These are the results when using a PF filter of 1.2 instead of 1. As it would be expected the trading frequency changes significantly (becomes much lower) which makes compounding less effective and hence total profits lower. The daily Sharpe ratios for the systems showed are 0.088 (regular SVM), 0.069 (PF>1) and 0.098 (PF>1.2). The reduction in information is in fact compensated by an increase in the risk adjusted return (the information ratio increases in this case). Therefore this filtering is something that improves trading results. I also include the PF evolution graph (now showing all PF values).

Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_svm_adapts4.png Size: 37 KB


Attached Image (click to enlarge)
Click to Enlarge Name: machine_learning_svm_adapts5.png Size: 32 KB


Remember that on this results the best model used for the next day's prediction is chosen based on model performance during the past 200 days, it is interesting to wonder what will happen if this threshold is moved towards larger or smaller numbers (I will be posting about this next).
  • #253
  • Jan 18, 2015 7:32am Jan 18, 2015 7:32am
It seems that going into the use of a deeper learning procedure was a bit off-putting (no one posted in a week!). So I will instead go into some simpler yet extremely interesting stuff related with the use of neural network models. So far we haven't looked at neural network models and we have only looked superficially at random forest models. The reason is that these two model types are non-deterministic (can give different results on each simulation due to the random nature of their initialization, this can be very problematic) so their use is a bit more complicated than a more deterministic model, such as a linear regression or an SVM.

I want to first explain a bit about Neural Networks (NN from now on). They work in the same manner as the models we have explained before, you have some inputs and want to predict some targets. In a Neural network your inputs first go through an input neuron layer and then go through hidden neuron layers until they reach the output neurons, which give you the target predictions. Each neuron is simply a function that does something with the data that comes into it and spits something out the other side. Neurons are functions that process data and the model simply adjusts the way in which each neuron contributed to the overall model (what data gets in, how much it is weight, etc). In an NN the weights assigned to neurons are optimized in an iterative process to reduce some user defined error metric in the output.

To make things simple, these are the things you should worry about when building an NN model:

  1. What is the topology of the NN ? How many inputs, how many ouputs, how many hidden layers (layers between input and output) and how many neurons per hidden layer.
  2. How to initialize neurons? What values will you use for neuron initialization, this is quite important as it can make convergence easier or harder. This also affects the none deterministic component significantly.
  3. What are the functions used in the NN? remember that neurons are functions. They can be linear, logistic, sigmoid, tangential, etc. Every function you can think about can be coded into a neuron. Which ones you use determines the level of complexity you can get (along with the topology). Functions should match the problem you're trying to solve. For classification problems use some types, for regressions use others.
  4. What optimizer to use? How are weights varied in the training process to reduce the error function.
  5. What error function to use? How will you measure the error of your neural network when doing training.
  6. How long to optimize for? What criteria will you use for error convergence.

The free and open source Shark library (also implemented in the Asirikuy F4 framework as I have mentioned before) allows us to build neural networks with very fine control of all the above variables. Here is a first example where I used two different neural networks that differ only in the function used for the input neurons. One uses a logistic neuron function while the other uses a fast sigmoid neuron function (see Shark website for more info on this). These are the results compared with our previously best SVM model:

Attached Image (click to enlarge)
Click to Enlarge Name: machinelearning_nn_1.png Size: 37 KB


These neural networks used the next day's return as target and used the return of the past 3 days as input. The models were retrained on each bar using the past 100 examples. A stop loss of 60% of the ATR(20) was used. The neural network weights are improved for 5000 cycles using the IRpropPlus optimizer, using the SquareLoss function to measure error in the ouput. (more info on this on the Shark library page)

You can already see from this that NN can be more powerful than SVM models. They are extremely flexible and can lead to very important improvements in results. You can see that the sigmoid function NN gets the best results (because the sigmoid function is much more flexible than a logistic function). Nonetheless the logistic function achieves great results under recent market conditions (so it works better sometimes). The world of NN is very complex and very rewarding, you will see how very good systems can come out of NN models. Using NN ensembles is a keystone of my trading system building procedures.

Note: I will be calling regression neural networks rNN and classification ones cNN.
  • #254
  • Jan 18, 2015 11:11am Jan 18, 2015 11:11am
Quoting algoTraderJo
Let us get down to business now. A real practical example using machine learning. Let's suppose we want to build a very simple model using a very simple set of inputs/targets. For this experiment these are the answers to the questions: What to predict (what target) -> The direction of the next day (bullish or bearish) What to predict it with (which inputs) -> The direction of the previous 2 days How to relate the target and inputs (what model) -> A linear map classifier This model will attempt to predict the directionality of the next daily bar....
Well your thoughts here are somewhat akin to some thoughts i have had myself.

How about concentrating on the break of the high or low of the day before? Actually, in my experience thats more important then if the day ends as a bear/bull candle. See the thing is, time is a continous flow right? Its not divided into frames. So if the day comes to and end as a bear or bull candle really dont matter so much. But if the day breaks the high and/or low of the day before, THAT seem to matter a lot. If lets say the low of the preceding candle is broken, but not the high, then to me thats a DOWN day, no matter if it ends bear or bull. See what i mean?
My signature is: "Classified".
  • #255
  • Edited 4:48pm Jan 18, 2015 3:13pm | Edited 4:48pm
Quoting algoTraderJo
It seems that going into the use of a deeper learning procedure was a bit off-putting (no one posted in a week!). So I will instead go into some simpler yet extremely interesting stuff related with the use of neural network models. So far we haven't looked at neural network models and we have only looked superficially at random forest models. The reason is that these two model types are non-deterministic (can give different results on each simulation due to the random nature of their initialization, this can be very problematic) so their use is...
good overview of NNs. I see you say "deeper learning" but just to clear for everyone, this is not related to deep learning in NN sense (you just meant more advanced methods). The main difference in deep learning would be the selection of features itself - here you manually craft features (last 3 days returns).

anyhow look forward to further discussion. i think you are progressing the thread very nicely.


PS: BTW, one reason for no response in a week could be many people handling the fallout of CHF moves. as you saw, many brokers went out of business or having troubles.
Think : Trade : Live Life
  • #256
  • Jan 18, 2015 5:31pm Jan 18, 2015 5:31pm
Quoting yaed
{quote} Well your thoughts here are somewhat akin to some thoughts i have had myself. How about concentrating on the break of the high or low of the day before? Actually, in my experience thats more important then if the day ends as a bear/bull candle. See the thing is, time is a continous flow right? Its not divided into frames. So if the day comes to and end as a bear or bull candle really dont matter so much. But if the day breaks the high and/or low of the day before, THAT seem to matter a lot. If lets say the low of the preceding candle is...
I was waiting for someone to suggest something along these lines. We haven't used the High/Low values for anything yet. What input/target structures would you suggest to use them? They do provide additional information but is this information relevant to the making of predictions or do the high and low just provide noise? If you want me to test something I will be happy to post the result so that you can take a look.
  • #257
  • Jan 18, 2015 5:35pm Jan 18, 2015 5:35pm
Quoting stt
{quote} good overview of NNs. I see you say "deeper learning" but just to clear for everyone, this is not related to deep learning in NN sense (you just meant more advanced methods). The main difference in deep learning would be the selection of features itself - here you manually craft features (last 3 days returns). anyhow look forward to further discussion. i think you are progressing the thread very nicely. PS: BTW, one reason for no response in a week could be many people handling the fallout of CHF moves. as you saw, many brokers went out...
Yes, you're correct, this is "deeper" meaning it is more complex, not in the sense of the field of deep learning (as in the machine learning field). Nonetheless see that I am assigning inputs dynamically on this post, looking at the input structure that gave the best profit factor across the last 200 days. I am not manually crafting features in that sense, although the possible features are limited to some constraints (look at the input graph to see how the optimum input number varies through the test). TY for your encouragement, I'm happy to hear you are enjoying my posts.
  • #258
  • Jan 18, 2015 5:51pm Jan 18, 2015 5:51pm
An important issue with NN implementations is how much to train them. There are two ways to do this, the first is to train the NN down to a given error limit (for example train until the error is below X threshold) and the second is to train for a given fixed number of optimization cycles. One would think that the first approach would work better because intuitively we are interested in having as small training errors as possible. However it is often the case that when error thresholds are pushed too low on data like financial time series (noisy data) the NN are able to incredibly overfit and start predicting information from the "noise" instead of making a generalization over the examples. We then end up with predictions that tend to be wrong more frequently because we're listening more closely to noise in the market data.

Taking the same best example showed above (fastsigmoid functions on the hidden layer, linear neurons on the output layer) we can see what happens if we decrease training from 5000 cycles to 500 cycles (this network has 3 input neurons (return of the last three bars), 3 hidden neurons (one layer) and one output neuron (the next bar's return). Remember that the model is rebuilt and retrained on every bar using the past 100 examples. This will cause the NN training on each bar to stop much sooner and therefore fail to reach a much tighter convergence than it would be possible on larger training cycles. This means we capture something very general instead of something much more specific. I'm displaying the results here:

Attached Image (click to enlarge)
Click to Enlarge Name: machinelearning_nn_2.png Size: 23 KB


See how the balance evolution is more homogeneous (linearity of log(balance) is higher) and we reach much better results under recent market conditions. This is simply because we're changing how we train the network. If we use different optimizers, different error functions or different ways to stop training we will reach different results. Tuning the appropriate degree of generalization is an important problem you always need to consider when using NN.

Neural networks are very complex, as the degrees of freedom in the creation of NN systems are significant. It is also true that changing anything will also change the behavior of the whole model so it can be time consuming to develop NN implementations as you need to really think all the variables you can choose depending on what input/target structure you're using (you need to really think about the system you are trying to predict). Changing input number and character affects the best topology, the best optimizer, etc. Brute force, blind, optimization of NN characteristics is often an exercise in futility (at the least very sub optimal). This type of model requires you to think what you are doing.

The above really makes you think about data mining bias, which is a topic we will start discussing later on this thread.
  • #259
  • Jan 18, 2015 6:45pm Jan 18, 2015 6:45pm
Since we have both NN (100 examples, 3 inputs, 500 cycles per training) and SVM (145 examples, 20 inputs) models with decent results (the first a regression model and the second a classification model). It's cool to attempt to build an ensemble of the two models and see if we can improve our results. This means retraining both models on each bar and simply only acknowledging a long signal when both say "go long" and a short signal when both say "go short" (do nothing if they disagree). We build a model ensemble using both classification and regression, so we expect to get some complimenting behavior from both worlds. The result is good:

Attached Image


This ensemble is already showing the high stability characteristics that I consider necessary to consider live trading (though I usually demand much higher!). We still have a long way to go but I wanted to show you how the road to successful ML tends to involve collaboration between different model types instead of a single model that just "works". This ensemble has an annualized sharp ratio of 1.53 while the SVM model has a Sharpe of 1.39 and the NN model has a sharpe of 0.8538.

Putting models in ensembles is not a "sure way" to improve results, but it is always something good to try. This is especially worth doing if both models have different degrees of generalization (different strengths and weaknesses) and they are both highly linear in log(balance) Vs time.
  • #260
  • Edited 5:09am Jan 19, 2015 4:01am | Edited 5:09am
Quoting algoTraderJo
{quote} I was waiting for someone to suggest something along these lines. We haven't used the High/Low values for anything yet. What input/target structures would you suggest to use them? They do provide additional information but is this information relevant to the making of predictions or do the high and low just provide noise? If you want me to test something I will be happy to post the result so that you can take a look.
Well i think i sort of explained something that seems interesting here:
http://www.forexfactory.com/showthre...82#post7876282
Since May E/U has been going down pretty clearly if you ask me. There has been a bit to suggest differently, but not in a big way.
Of course now that we know what occured between Say May 10 until now, just selling where ever and holding was a apparently a good idea. But lets assume we didnt know THAT much.

So, in a sell trend like we have had on E/U since beginning of may, sell as the high is broken of D1 candles, TP=40% of ATR (low period, say 5). Now in this scenario, apparently we dont have to worry about SL, since if the account would contain a couple of hundred pips of possible drawdown, nothing bad ever happend. But lets assume we didnt know that much. Lets assume this is a carefull MM, with quite low leverage: 100 pip=1% of account. Lets put SL at a total of... 700 pip, no matter the amount of positions, every positions lot size equal. So that would be a catastrophic loss of -7%. A good normal day nets +0.4%, since ATR is roughly 100 pip for this period. Spread in this example will be quite a small bit of total earnings or losses.

Many other kinds of SLs are quite doable:
Total of 300 pip?
Only 1 position, SL= 100% of current ATR when the position is placed?

I think a trader should not be afraid to experiment with multiple positions, since this might be what hedgefunds are doing.

There are many possibilities to create more intelligent ways of take profit. For instance a bit of code that reads the PA on say H1 and only closes a trade if a high is broken on a candle if its a sell position, and vice versa for a buy position.

All this is still pretty primitive thinking. But i think it might be a start. I mean, hedgefunds employ PILES of people to do this kind of stuff. If people like us should have ANY chance of beating them we gotta get REALLY creative. But really creative MIGHT actually mean really SIMPLE. There is the possibility that there are mechanics that they have missed.

I for one are very much concentrating on highs and lows of candles, if a high or a low is broken or not. If it ends as a bull or bear candle, from my experience has little meaning.

So lets have a look at what i mean by that in depth and how i read a chart.

Attached Image (click to enlarge)
Click to Enlarge Name: Thinking about h4 highs and lows.jpg Size: 302 KB


Attached Image (click to enlarge)
Click to Enlarge Name: M15 zoom.jpg Size: 228 KB


The SNB cant defeat me though. I am immortal. So are all of you too. So: ONWARDS!
My signature is: "Classified".
0 traders viewing now