Machine Learning

[hhiusa]

data snooping and machine learning are unrelated. I do not know what kind of system you are referring to but it is a rather specific concept nit just used in funance. Some have said that it does not work, which leads me to believe they have implemented it incorrectly.

You do not buy pre-made programs as that would be fault #1. It is a complicated system that requires fine tuning based upon the situation.

The numbers of dimensions is extremely important in the training set. Too many dimensions will lead to noisy data and whipsaws. Too few dimensions will leads to unresponsiveness.

Machine learning like naive bayes classification relies upon an initial condition or initial conditions and a training set. In order for it to learn, you must teach it learn. An "objective function" must be created.

There are two main classes, namely, unsupervised learning and supervised learning. It is important to know which one you need to use.

 

[Krzysiaczek99]

Hi Krzysztof...

You earlier asked this board about TSSB...and I stated I did check it out...but I was going to focus on learning Deep Thought instead of TSSB.

I eventually ended up playing with both programs...and was amazed to find that they actually may complement each other...

TSSB is extremely sophisticated with it's solid statistical basis for machine learning... It's a little harder to work with for me...but I'm learning a lot with it.

Deep Thought continues to be the more exciting one for me...

Especially in that it's the only advanced machine learning program I know of that actually has built-in MT4 Expert Advisors...and...actually does real-time trading...

As a confirmed novice with machine learning...I love the fact that Deep Thought also has built-in machine learning algo's that I can watch and learn how they work.

In the meantime...let me turn the question around for you...

What machine learning programs are you currently using...and what do you think of either Deep Thought and/or TSSB...?

Thanks and take care,
Robert

Hi

as 'What machine learning programs are you currently using' I'm an author of this thread

http://www.trade2win.com/boards/tra...ef-nets-restricted-boltzmann-machines-24.html

posting there since like 5 years. I'm using MT4-MATLAB-WEKA setup at the moment but tried like 20 different algos on different financial instruments. You can see in the latest results on latest posts of this thread.

regarding Deep Thought and/or TSSB - no opinion as i didn''t try them but for sure choice algorithm is just a small part of the puzzle.

Krzysztof

 

[cosmicbeing]

Hi

as 'What machine learning programs are you currently using' I'm an author of this thread

http://www.trade2win.com/boards/tra...ef-nets-restricted-boltzmann-machines-24.html

posting there since like 5 years. I'm using MT4-MATLAB-WEKA setup at the moment but tried like 20 different algos on different financial instruments. You can see in the latest results on latest posts of this thread.

regarding Deep Thought and/or TSSB - no opinion as i didn''t try them but for sure choice algorithm is just a small part of the puzzle.

Krzysztof

Hi Krzysztof,

Thanks and I will follow up and post over on your thread to keep this thread focused on Deep Thought machine learning...

In the meantime...maybe in the future when I'm more comfortable running the Deep Thought machine learning algo models...I'd like to try and run and compare algo models on Weka and Deep Thought...just to see how close they come to their solutions...

Thanks and take care,
Robert

 

[Solas0077]

However...the fact that Google and Amazon and many other big boys are successfully using machine learning everyday tells me they either found a solution or a work around...

Thanks,
Robert

They are not using it to develop trading systems. Their use is different. ML works but not for trading systems.

 

[Solas0077]

data snooping and machine learning are unrelated.

In general yes but for trading systems they are related. Data-snooping causes bias. This is good article.

 

[hhiusa]

They are not using it to develop trading systems. Their use is different. ML works but not for trading systems.

I do not need an article. I am not talking about pre-designed programs. There is no such thing as evidence that something does not work, only evidence that is does. Ex. There is no such thing as evidence of no cancer, only no evidence of cancer. This means that there may still be evidence but you have not found it. Evidence something does not work is no evidence. Machine learning is pretty big. Stocks prices are described by stochastic processes. Stochastic calculus is used all the time in finance. The goal is usually optimization in order to find patterns in k-dimensional data provided by features. There are many different algorithms to try.

Since you think it does not work, you must be fairly versed with partial differential equations and solving systems of PDEs not to mention other fields in mathematics.

If it did not work for you, I can think of several reasons why.
1. Your objective function is not finely tuned
2. Feature bloat i.e., it has too many dimensions
3. Improper initial conditions
4. Not enough or too many restraints
5. Not enough training data in training set or faulty data

There are many more.

 

[Solas0077]

ML does not work for trading systems because it curves fits to past conditions. Everyone knows this but you.

This guy knows better than you

http://eranraviv.com/sample-data-snooping/

 

[hhiusa]

ML does not work for trading systems because it curves fits to past conditions. Everyone knows this but you.

This guy knows better than you

http://eranraviv.com/sample-data-snooping/

The fact that you mention "curve fitting" shows me you have no clue. Why don't you post in your words about it, instead of using a website. Since you think it has to do with "curve fitting" you are probably finding snake oil salesman.

You obviously did not read any of previous posts or were unable to understand as I listed topics, which do not use "curve fitting". Try re-reading and explaining in "your words" how the math does not work!

 

[Krzysiaczek99]

If it did not work for you, I can think of several reasons why.
1. Your objective function is not finely tuned
2. Feature bloat i.e., it has too many dimensions
3. Improper initial conditions
4. Not enough or too many restraints
5. Not enough training data in training set or faulty data

There are many more.

I would add two more

6. Too noisy data for algorithm to pick up the patterns
7. Too much training data in training set which cause loosing of precision

Krzysztof

 

[hhiusa]

I would add two more

6. Too noisy data for algorithm to pick up the patterns
7. Too much training data in training set which cause loosing of precision

Krzysztof

Thanks.

 

[Solas0077]

You obviously did not read any of previous posts or were unable to understand as I listed topics, which do not use "curve fitting". Try re-reading and explaining in "your words" how the math does not work!

I am sorry you do not have the capacity to understand that post by Ravin and instead you troll, immersed in snake oil.

 

[hhiusa]

I am sorry you do not have the capacity to understand that post by Ravin and instead you troll, immersed in snake oil.

You have not explained anything except that you want some other guy to do the talking for you. You have trolled more than I. You continue to reply each time and tell me how wrong I am. If you were truly secure in your views, you would not need to. Others on this thread agree with me. You must have missed @Krzysiaczek99's posts as well.

I think "the lady doth protest too much" with you.

 

[Krzysiaczek99]

article from PAL

In general yes but for trading systems they are related. Data-snooping causes bias. This is good article.

The author of this article sells Price Action Lab software so its is conflict of interest here.

his conclusion is just a manipulative language in my opinion.

Nowadays, methods for dealing with data-mining bias define the nature of the process that can be used to discover trading algos. Therefore, those methods constitute the edge and the data-mining process. For example, the founder of a well-known fund has supported a specific method for measuring data-mining bias and because of that many quants are trying to replicate it. However, if one looks at the performance of that fund it has been flat or negative in the last several years. Thus, it is evident that the fund has no edge.

The conclusion is that machine learning without an underline rationale and philosophy for dealing with data-mining bias that is not based on methods that become part of the loop, as discussed above, is indistinguishable from a random process and gambling.

of course data mining bias exist and of course Machine learning is working - the best example - google search. So the trick is just not to get fooled by data mining bias and properly backtest trading strategies.

Krzysztof

 

[Krzysiaczek99]

And here are the free toolboxes for strategy evaluation in MATLAB and PYTHON

https://quantiacs.com/For-Quants/Tutorials/MatlabToolbox.aspx

and another one for FOREX in PYTHON

https://www.quantstart.com/qsforex
https://github.com/mhallsmoore/qsforex

Krzysztof

 

[Krzysiaczek99]

biases

I made some investigation and here is more info about different biases which occur in machine learning and backtesting. I think is very important to realize with what we are dealing here.

http://www.trade2win.com/boards/tra...ricted-boltzmann-machines-24.html#post2612880

Krzysztof