Menu
Sister Sites

Options

Bookmark Thread

First Page First Unread Last Page Last Post

Print Thread

Similar Threads

Machine Learning with algoTraderJo 939 replies

Algorithmic Quant Trading (Machine Learning + Stat-Arb) 25 replies

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: Ideas for Data & applications of Machine Learning with Azure ML
Exit Attachments

Ideas for Data & applications of Machine Learning with Azure ML

  • Post #1
  • Quote
  • First Post: Edited at 2:56pm Jun 11, 2018 6:38am | Edited at 2:56pm
So I've managed to create an EA that communicates with Azure ML and the Web Service API Provided, I use prices volumes and every indicator in the book as input. I tried predicting pretty much everything in every way i thought reasonable and useful. In the Azure ML platform what i found it most accurately predicted was an assigned column as label to be predicted as output of levels of how many PIPs the market moves and use Multi-class Models rather than a regression model. As in (1,2,3,4) for bullish moves and (-1,-2,-3,-4) for bearish move and (0) for little or no movement, or of same idea with just 2 or 3 class (1)UP (-1)DOWN.

I've been trying different data bases and ways of how i can make use of the predictions, but i ran out of functioning ideas or i am not using the results the correct way, i am new to this and still learning so i wanted to open this thread to get some help and some ideas of applications for Machine Learning in Forex and the use of Azure ML. I wanted to get insights, views and ideas from people with more experience on how to actually use forex data and what could potentially work better or some idea of how to use ML for forecasting forex, or at least get some statistical edge with it. Like i said i am still learning and, i think i lack the perspective of a forex veteran even if it's not in machine learning in a programming and coding level but on how to actually make use of forex data and how the market actually works and share some wisdom and give me some light and share an idea of how it could potentially be applied or of what could work in any level.

I am also interested in the insights of people that might not have any knowledge in ML or coding skills or in the subject but have a perspective of a possible application. I also am willing to help and share with people that are too interested in such applications and Machine Learning for forex and the building of data bases for such predictions. As long as we're all willing to share and help each other and grow together with our ideas.

So back to my results: I got results above 90 percent accuracy in Azure ML in almost all my tries after some input switches, i also got good results in predicting ZigZag Points.
But i can't seem to make the EA work as perfectly as the results i get in Azure ML either with Neural Networks or Any other Machine Learning algorithms. i don't know if i am missing something or if the data base is probably not convergent or good enough or something else. So i would like some thought and help with building a better database for Forex forecast with Azure ML.

So can someone please try to explain to me and help me figure out why that is that i get nearly perfect results in Azure ML but in Real Time Forward testing and back-testing i get nowhere near as good results.

What is more interesting is that if i test it with known data, it works perfectly. it doesn't get anything wrong, so the problem is not in the input or output, I actually thought it was perfect for a second until i realized it was known data i was back-testing with, but as soon as i test with unknown data it's as good as random. So please.. does anyone have any perspective on it and can help me figure out what is wrong with it. and or what would be the best and most reliable data for input. I've tried everything i could think of, but it didn't live up to the expectations so far.
So i wanted help and insights on it, any perspective on it would be of great help and or if someone has had better results than mine with this.

Thanks in advance.

-Image 1 is a dataset i created for predicting about the next 10 candles in the M15 Time-frame with 2 Class UP(1) and DOWN(-1), Image 2 is the results i got.
Attached Image (click to enlarge)
Click to Enlarge Name: dataset azureml.jpg Size: 78 KB
Attached Image
Very good effort my friend.

1. Have you taken the base Tick data as the input or have you taken processed data- M5, M15 or so on.

2. ML to work effectively, we need to have training data. - am assuming when you mention Indicator data, you are using indicator data as training data.

3. Initial ML we can just look at a pattern- Fibo Levels as an example. and then we can get a success on the prediction. or simple UP/Down.

my assumption, is your base data is not appropriate, as you getting good results on known processed data. May be try to change the base one. you might see better output.
Quoting cdsaa
Very good effort my friend. 1. Have you taken the base Tick data as the input or have you taken processed data- M5, M15 or so on. 2. ML to work effectively, we need to have training data. - am assuming when you mention Indicator data, you are using indicator data as training data. 3. Initial ML we can just look at a pattern- Fibo Levels as an example. and then we can get a success on the prediction. or simple UP/Down. my assumption, is your base data is not appropriate, as you getting good results on known processed data. May be try to change the...
Hi, Thank you for your reply. i use processed data the one provided for MetaQuotes in mt4, do you know how i would organize the tick data for input for a relevant output? The problem is i'm not using my own ML algorithms, i use Azure's so i know it needs columns of data related to one label column to be predicted. So i don't know how i would organized the tick data and an output for it , do you have any idea how i would do that? and what other types of data could i use that might be more appropriate?
It's sad that it seems promising but it just doesn't work apparently, if machine learning can't find a working pattern the market must be as random as it can be. It's a game still to be cracked. I hope at some point someone can give me some light. i will still experiment with it. if someone can post ideas or something that might work better here i will try them out.
does anybody know if there is something different about the market this year, because what i am doing is using data from the previous years as training data and leaving 2018 months out for testing in mt4 which returns me terrible results, maybe i get those good results in ML algorithms for the past but the market is just too different somehow in 2018, i don't know, that might make sense maybe. Well, i will let people know that are interested if i have any progress relevant for real account in 2018.
Quoting Excelx
does anybody know if there is something different about the market this year, because what i am doing is using data from the previous years as training data and leaving 2018 months out for testing in mt4 which returns me terrible results, maybe i get those good results in ML algorithms for the past but the market is just too different somehow in 2018, i don't know, that might make sense maybe. Well, i will let people know that are interested if i have any progress relevant for real account in 2018.
Of course market is always different every year.My experience is try to predict the current market and avoid historical examples that do not represent current market behavior. I limit the amount of past history between 300 and 2000 bars so that the neural network will learn data relevant to today’s market.
Quoting herbie88
{quote} Of course market is always different every year.My experience is try to predict the current market and avoid historical examples that do not represent current market behavior. I limit the amount of past history between 300 and 2000 bars so that the neural network will learn data relevant to today’s market.
That makes sense, i also i think i use too much historical data too far back that apparently are not relevant or represent the current market conditions, i thought more data would be better. But that might not be the case here. I tried with 2000 candles back like you said it's and used the last 2 weeks to test it with. The results in Azure ML are not as accurate with less training data ( Image 2 ) but even so i got better results with the EA than before ( Image 1 ). But i it's hard to know if it is reliable enough with just 2 weeks of results. I will keep trying new things see if i can improve it, Thank you for your insight it was very helpful!. I hope someone can share insights so i can develop this further and hope for better results.
Attached Image (click to enlarge)
Click to Enlarge Name: ea zure res.jpg Size: 119 KB
Attached Image
What are your inputs-outputs?What are you predicting?
  • Post #10
  • Quote
  • Edited at 10:38pm Jun 12, 2018 10:21pm | Edited at 10:38pm
Quoting herbie88
What are your inputs-outputs?What are you predicting?
The inputs are all standard mt4 indicator values and price, what i predict is a column related to those values with a value of 1, or -1, telling me if the price went down or up. so i try to predict that value based on the other values.
Quoting Excelx
{quote} The inputs are all standard mt4 indicator values and price, what i predict is a column related to that values with a value of 1, or -1, telling me if the price went down or up. so i try to predict that value based on the other values.
that value is the output, of 1 or -1, telling me if the price will go up or down.
Quoting Excelx
does anybody know if there is something different about the market this year, because what i am doing is using data from the previous years as training data and leaving 2018 months out for testing in mt4 which returns me terrible results, maybe i get those good results in ML algorithms for the past but the market is just too different somehow in 2018, i don't know, that might make sense maybe. Well, i will let people know that are interested if i have any progress relevant for real account in 2018.
Dear Excelx,
I don't know if it will help, but I am also encountering issues of market randomness in my EA development. I get equity curves of perfectly straight gradual increases and then something happens that the system is not designed to deal with. On another thread I have proposed to identify groups of several scenarios and run the EA through those, and it makes scene to me to almost use that for your ML and not the entire history (might be corrupt/incomplete). Anyways that is just my view/ opinion.

This will be my first thread as soon as I have some time to work out how to approach/ present it to this awesome community. So please do look out for my thread soon, and hope it will be useful for some.

Tiny example here of what I mean, perfect ascend over many weeks/ months and then something changes. EA has never encountered such a situation, sh1t hits the fan and drawdowns/ losses occur.
Attached Image

P.S I would avoid using every indicator in the book, not all are developed equally and some are just too derivatized and lagging that they will only slow down your system and ML process by giving contradictory signals. Less is more and bad knowledge is worse than no knowledge at all, IMHO.

Kind Regards,
Tzamo
"Only you can Make the Future you will be proud to be a part of..." -Me
  • Post #13
  • Quote
  • Edited at 11:56pm Jun 12, 2018 11:19pm | Edited at 11:56pm
Quoting tzamo
{quote} Dear Excelx, I don't know if it will help, but I am also encountering issues of market randomness in my EA development. I get equity curves of perfectly straight gradual increases and then something happens that the system is not designed to deal with. On another thread I have proposed to identify groups of several scenarios and run the EA through those, and it makes scene to me to almost use that for your ML and not the entire history (might be corrupt/incomplete). Anyways that is just my view/ opinion. This will be my first thread as soon...
Hi, Thank you for your thoughts, I subscribed to you i look forwards to your thread. I am still learning i enjoy seeing new perspectives and ideas on trading and the market, but when you say groups of several scenarios, i don't know how exactly identify those, i'm sorta new to the market, i know i have a lot to learn because there's still things i see people posting about i am yet to know what there are and make use of them, like market conditions, and i have very little knowledge of fundamental analysis, I am still to go down that road. haha That's why i seek ideas. i am trying it hard to figure it out. that's one of why posted this thread looking for someone to help me with a more advanced perspective on it than mine on market behavior.
And what i am doing so far isn't working the best way i would hope so.

And following your perspective, i also think i am using "dirty" data, i am going to try to start cleaning and see what is relevant and what isn't. I am pretty sure the problem is my database, that's also why i am asking for help in this thread with building my data base. I think i am not using data relevant or the correct way for this to work. So a perspective from someone with a Forex background and knowledge of how the market works would help me.

I am still going to keep trying this, i think there's potential in it.

Well, thank you. I will see to it. and i will post for who are interested in the development.

Regards, and best wishes.
Use dukascopy Tick data.

You have tools to download the Tick data from dukascopy.

https://strategyquant.com/tickdownloader/

If the training data is clean, it would give improve on the results.
Quoting Excelx
{quote} The inputs are all standard mt4 indicator values and price, what i predict is a column related to those values with a value of 1, or -1, telling me if the price went down or up. so i try to predict that value based on the other values.
Have you tried to predict price change directly? I am currently experimenting with neural networks and finding out that regression models are working better on the future datasets
Quoting vladi1979
{quote} Have you tried to predict price change directly? I am currently experimenting with neural networks and finding out that regression models are working better on the future datasets
How do you predict the price change(percent ?) ?Use the lag value of price the change as input then use the lead value as output?End up something like this https://hackernoon.com/dont-be-foole...g-bf27e4837151
Quoting patara
{quote} How do you predict the price change(percent ?) ?Use the lag value of price the change as input then use the lead value as output
I am not actually using price change as inputs. I am using indicator value [pattern] as an input to the network and lead value of price change as output to train the model. In theory this should mimic traditional mechanical systems that make use indicator levels as a trading rule. Additionally NN are known to be superior in pattern recognition capabilities. Deep learning model implementation in *h2o* allows both discrete and factor as a label, hence my question about regression models

Method is quite cumbersome to build and am still testing it. Obviously it's not a holy grail as
Quote
'past data is not guarantee for the future results...'

Attached Image (click to enlarge)
Click to Enlarge Name: Screenshot2.png Size: 28 KB
Quoting vladi1979
{quote} I am not actually using price change as inputs. I am using indicator value [pattern] as an input to the network and lead value of price change as output to train the model. In theory this should mimic traditional mechanical systems that make use indicator levels as a trading rule. Additionally NN are known to be superior in pattern recognition capabilities. Deep learning model implementation in *h2o* allows both discrete and factor as a label, hence my question about regression models Method is quite cumbersome to build and am still testing...
When you do clustering and classification models, do you still use price change as output ?Or do you use the price change as output only for regression models?
Quoting patara
{quote} ....use the price change as output only for regression models?
Use indicator pattern as input and price change (shifted) as output. This can do both regression and classification but I saw experimentally that strategy tests are better using regression models

I wrote a blog post about the method with more detailed implementation steps... I am running the entire thing on local machine using h2o framework for deeplearning not on Azure ML Studio though...
You said in your blog that the result of the model can be in the range from -1 to +1.Do you have a choice to select the transfer function(tanh, sigmoid,symmetric logistic, gaussian etc..) for a neural network model with h20?What is the best way to select the best transfer function for a neural network model?
0 traders viewing now