Machine Learning with algoTraderJo | Page 43 | 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

  • #841
  • Feb 25, 2016 7:39am Feb 25, 2016 7:39am
Quoting olsen-yersen
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
I keep searching how to use HMM for market conditions estimation. The only resources I'm able to find are only about the volatility.
For example this one: http://countingpips.com/2015/10/iden...hine-learning/
When you take the picture of his results and add a few lines onto it you clearly see that the HMM simply thresholds the ATR. No need of a HMM to draw 2 horizontal lines on my ATR oscillator.
Attached Image


A good modelling of the various regimes is required to use HMM. Just "trend slope + returns variance" doesn't sound enough to me. With the same trend slope and the same variance of the returns you can get a nice waving trend or an ugly choppy market.
No greed. No fear. Just maths.
  • #842
  • Feb 25, 2016 1:54pm Feb 25, 2016 1:54pm
Yes I have seen that article ,thanks.In the following article the author claims to doubling his return per trade from 2.1 pips to 4.2 pips: https://inovancetech.com/ml-strategy-techniques.html But the thing is why one needs hidden markov model for thresholding the ATR?I have a hidden markov model which is not better than some certain type of moving average.Ernie Chan confess similar things with HMM's :''I have never been able to create a HMM model that outperforms simple technical indicators'' http://epchan.blogspot.com.tr/2012/0...ied-to-fx.html
  • #843
  • Mar 8, 2016 3:18pm Mar 8, 2016 3:18pm
Some of you might find this (fairly detailed) blog post interesting:
Machine learning for financial prediction
  • #844
  • Apr 6, 2016 12:53pm Apr 6, 2016 12:53pm
Quoting FXEZ
Some of you might find this (fairly detailed) blog post interesting: Machine learning for financial prediction
Good stuff in here re number of variables. Thanks.
  • #845
  • Apr 8, 2016 3:57pm Apr 8, 2016 3:57pm
Quoting PipMeUp
For trading purpose we have very few samples compared to the universe of the possibilities. Here is ELM with 30 neurons vs Gaussian Processes. There are now only 50 samples and no mis-labels. I much prefer what the GP finds. {image}{image}
I was surfing the net about application of Extreme Learning Machines on financial time series.Then I read and enjoyed Yann LeCun's'' What's so great about Extreme Learning Machines"?in Reddit:https://www.reddit.com/r/MachineLear...treme_learning
Especially the Chinese Prof. was funny, enjoy
Inserted Video
  • #846
  • Apr 9, 2016 4:28am Apr 9, 2016 4:28am
Yann LeCun seems to lead the detractors of this method. I think with reason. I read his article several weeks ago and some others. I had also watched the video you posted (which brings nothing). Clearly the ELM are mere overfitters with poor to no generalisation ability. They work well when the number of neurons in the hidden layer is big. Actually when they tend to be an infinity. Come on give me a 200 order polynomial I can fit whatever you want! They would be interesting if they could work with a very low number of neurons. I tried to see if it was possible to use them with more than one hidden layer. This exists and is called hierarchical-ELM. They overfit much better . I also tried to find hierarchical Gaussian Processes. I found only one paper but the maths went quickly beyond me when they introduced the latent gaussian processes. It's already not easy to understand what a GP is. Anyway I think the computation time must explode (GP are already very slow).
No greed. No fear. Just maths.
  • #847
  • Apr 9, 2016 4:34am Apr 9, 2016 4:34am
A very interesting article about how the NN see what they want to see:
http://www.cv-foundation.org/openacc...CVPR_paper.pdf

That was translated into a funny video presentation anybody can understand:
Inserted Video
No greed. No fear. Just maths.
  • #848
  • Apr 9, 2016 11:39am Apr 9, 2016 11:39am
Quoting PipMeUp
They work well when the number of neurons in the hidden layer is big. Actually when they tend to be an infinity.
In theory this is true - accuracy on the training data should go up the larger the number used. But you will also get overfitting this way no matter what tool you use (GP, NN, regression, ELM). So basically this is more a critique on a flawed model building process rather than on this specific tool.

In practice with ELM it appeared that there was a sweet spot range of neurons for accuracy and it wasn't that large. I was creating a model then measuring the outcome and repeating this process walking forward. If too large a number was used in my test, then it started overfitting the training data and the accuracy on out of sample data went down.
  • #849
  • Apr 10, 2016 5:08am Apr 10, 2016 5:08am
Quoting FXEZ
In theory this is true - accuracy on the training data should go up the larger the number used.
Sorry I didn't explain myself clearly enough. ELM can be interpreted as an approximation to a single layer feed forward network with infinite number of hidden units. The approximation is better when the number of neurons actually tends to infinity. A SLFN with an infinity of neuron can obviously fit anything by "learning by hart" (store) the dataset.
No greed. No fear. Just maths.
  • #850
  • Apr 18, 2016 8:10am Apr 18, 2016 8:10am
Anybody using scikit-learn or TensorFlow?

I'm playing a bit with both, and I'm trying to figure out how to map time-series to features.

What I'm thinking at the moment is something like taking for example the last 5 prices at an 1 hour sample rate (so basically the Close of previous 5 H1 candles), fixing the last price as 0 and treating the previous ones as a pip delta.

So [1.1310, 1.1315, 1.1298, 1.1294, 1.1301] would become [9, 14, -3, -7, 0] after fixing 1.1301 as 0.

I have other non-price-data features I will use, but the price-data gives me most troubles.

Any thoughts or papers on this?
  • #851
  • Apr 29, 2016 10:16pm Apr 29, 2016 10:16pm
Quoting 000
Anybody using scikit-learn or TensorFlow? I'm playing a bit with both, and I'm trying to figure out how to map time-series to features. What I'm thinking at the moment is something like taking for example the last 5 prices at an 1 hour sample rate (so basically the Close of previous 5 H1 candles), fixing the last price as 0 and treating the previous ones as a pip delta. So [1.1310, 1.1315, 1.1298, 1.1294, 1.1301] would become [9, 14, -3, -7, 0] after fixing 1.1301 as 0. I have other non-price-data features I will use, but the price-data gives me...
Just started diving into scikit-learn, I also want to try looking into hidden Markov models. Here is a paper you might find useful: https://www.cs.nyu.edu/media/publica...TR2013-953.pdf
  • #852
  • May 1, 2016 4:06am May 1, 2016 4:06am
Quoting Domopips
{quote} Just started diving into scikit-learn, I also want to try looking into hidden Markov models. Here is a paper you might find useful: https://www.cs.nyu.edu/media/publica...TR2013-953.pdf
The performance in this paper isn't really good ...
  • #853
  • May 1, 2016 3:34pm May 1, 2016 3:34pm
Quoting surfeur
{quote} The performance in this paper isn't really good ...
Just wanted to highlight the methodology. Google around and you'll see profitable and not so profitable applications. Understanding the ones that are not profitable may just help your own research
  • #854
  • Jun 22, 2016 1:26pm Jun 22, 2016 1:26pm
I'm currently performing 10 years backtest on AUDUSD 1H employing neural network in R and blotter. Though it's running terribly slow (one year per 4-5hours) it seems promissing .

What I'd like to highlight here is that the main idea behind my model is available in this thread. Thanks to algoTraderJo I've finally found something reasonable for further working on!
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot1.jpeg Size: 373 KB
  • #855
  • Edited Jun 30, 2016 12:46am Jun 29, 2016 12:31pm | Edited Jun 30, 2016 12:46am
And this is final result of almost 11 years backtest.
Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot1.jpeg Size: 517 KB
  • #856
  • Jul 19, 2016 2:15pm Jul 19, 2016 2:15pm
Quoting mcquak
And this is final result of almost 11 years backtest. {image}
did your model experience any negative years?
I am what Many Dream to be but only a few can achieve, im a part of the 1%
  • #857
  • Oct 6, 2016 9:23am Oct 6, 2016 9:23am
Quoting mcquak
And this is final result of almost 11 years backtest. {image}
Can you provide any information on the topology of neural network that you used?
  • #858
  • Mar 11, 2017 4:02pm Mar 11, 2017 4:02pm
At last I have some time! Having kids often means no time for fun on forums.

Time to get back to where we were and make a few additional contributions here, for all those machine learning enthusiasts out there.

I have been exploring the use of renko charts in machine learning. I will be posting a few things about that soon.

Cheers!
  • #859
  • Mar 11, 2017 4:13pm Mar 11, 2017 4:13pm
Renko bars can be complicated. If you have ever backtested renko bars using something like MT4 you'll notice that you never seem to get accurate results. Yes, even if you use tick data. You write a pretty EA, get very good results and then you see that they never work equally well when live trading. The problem comes from the time variability in the renko chart creation which makes renko bars vulnerable to a few damn awful problems. Any use of renko bars must for the sake of reproducibility and accuracy fulfill these requirements:

  1. It must not matter when your data starts. If your data starts in 2000 or in 1995, the renko charts generated must be the same. If they are not the same thing then you're done.
  2. It must not matter when your test starts. If your backtest starts in 2011 or in 2015 it must be able to always give the same results for the time period in common. So a backtest started in 2011 and a backtest started in 2015 should give the same results in 2016.

Why do these things matter? Because if you cannot get things to be constant then you will most likely not be able to create an equivalent renko chart in live trading, your broker will not have the same amount of data or starting points as your backtesting data. This is why renko backtesting never works in MT4. Yes, even if you use tick data. Because the renko chart creation procedures that are out there do not fulfill these requirements. If you have a renko creation procedure that fulfills the above requirements it won't even matter if you create your renko from say ticks or 5M charts.

Any guess how me might do that? I bet some of you do.

  • #860
  • Mar 11, 2017 5:49pm Mar 11, 2017 5:49pm
The problem can be better understood when you think about how renko bars are made. For the sake of example let's say you want a 2 point renko chart and the market moves in the following way through 7 ticks:

1. 10
2. 11
3. 12
4. 11
5. 10
6. 11
7. 10

If you start from tick 1 then your renko sequence will look like this:

  1. Bar up (open=10, high=12, low=10, close=12)
  2. Bar down (open=12, high=12, low=10, close=10)
  3. Open bar (open=10, high=11, low=10, close=10)

If you start from tick 2 then your renko sequence will look like this:

  1. Open bar (open=11, high=12, low=10, close=10)

It's not difficult to see how the place where you start can lead to very different renko charts. Even one tick can make two renko charts look completely different if you build them like most people do. In this case a 1 tick offset leads to having 2 fewer bars :o

1 trader viewing now