Machine Learning Algo for Trading

[stepseazy]

Obviously, but the question is what algorithms do they use for pattern recognition and how is the data formatted so the computer can understand it? There are many machine learning algorithms. For instance, in spam detection, they convert the email into a bag of words, which is a list of words and the counts of these words, and they can use a Support Vector Machine. To classify the results into two categories: spam and not spam.

The problem of creating a trading algorithm is similar, it could check to see whether a certain time point fits the condition of a trade, or not a trade. However, it could further place a score on the probability of the trade working based on past decisions to determine trade size, and based on subsequent events decide whether to fully or partially exit the trade.

But for the sake of simplicity, let’s assume that the algorithm simply decides whether or not to place a trade of a fixed size. This then, is a binary classification problem with supervised learning (supervised because past data will show if this trade would have worked in the past).

These are the algorithms wiki lists for binary classification:

• Decision trees
• Random forests
• Bayesian networks
• Support vector machines
• Neural networks
• Logistic regression

So the trick for the simplified algorithm then, is to choose one of these algorithms, then convert market data into a format which can be used in the algorithms, and create a buy or sell entry when conditions are met.

I’m thinking that possibly a boosted tree could work well for this situation. Any pros out there that can verify the algorithm they use? Thanks.

-Rik

 

[conduit]

I use a classification algorithm implemented via convolutional network and it works reasonably well but the focus is only on predicting certain time series dynamics and not to make buy or sell decisions itself.

Obviously, but the question is what algorithms do they use for pattern recognition and how is the data formatted so the computer can understand it? There are many machine learning algorithms. For instance, in spam detection, they convert the email into a bag of words, which is a list of words and the counts of these words, and they can use a Support Vector Machine. To classify the results into two categories: spam and not spam.

The problem of creating a trading algorithm is similar, it could check to see whether a certain time point fits the condition of a trade, or not a trade. However, it could further place a score on the probability of the trade working based on past decisions to determine trade size, and based on subsequent events decide whether to fully or partially exit the trade.

But for the sake of simplicity, let’s assume that the algorithm simply decides whether or not to place a trade of a fixed size. This then, is a binary classification problem with supervised learning (supervised because past data will show if this trade would have worked in the past).

These are the algorithms wiki lists for binary classification:

• Decision trees
• Random forests
• Bayesian networks
• Support vector machines
• Neural networks
• Logistic regression

So the trick for the simplified algorithm then, is to choose one of these algorithms, then convert market data into a format which can be used in the algorithms, and create a buy or sell entry when conditions are met.

I’m thinking that possibly a boosted tree could work well for this situation. Any pros out there that can verify the algorithm they use? Thanks.

-Rik

 

[syswizard]

I've used SVM for logistic regression....with good results....a healthcare application.
Now setting up the data structures to do the analysis for market data....that is quite tricky.
You've got to create covariates based on the conditions and the signals you think is appropriate. Then you must put in the close prices for "x" bars into the future....and the analysis will provide you probabilities of a profitable trade.
Of course it's even more complicated if you are also incorporating a sell signal.

 

[Rationalize]

Random forests / Particle filters
But
There's more efficacy in working out mechanical financing trades with many moving parts, and automating that.

 

[jcl366]

The algorithm is relatively irrelevant. Much more important are the data with which you feed it. Normally a bunch of indicators or other derivatives of the price curve.

When you selected the features, then you can find the best algorithm for them. All of your list are available in R. You can relatively quickly try them out.

I found deep learning networks the most versatile algo for financial prediction so far. There are several R implementations for deep learning.

 

[Rationalize]

I disagree that prediction is worthwhile
Energy better focussed on optimising hedges
Finding reasonable replications
With those t00ls.

 

[victorycountry]

I am testing patterns in 1 min and find many statistically significant patterns that predict future volume and return but when I have tested in lower frequency time frames...they're not statistically significant...I guess that's why high frequency traders use automated system in 1 minute timeframe...

 

[Rationalize]

On a one min time frame
It's likely the bid ask will eat your significance

 

[victorycountry]

On a one min time frame
It's likely the bid ask will eat your significance

Sounds like you don't understand what my result is suggesting...never mind

 

[conduit]

Disagree, since when does the gasoline matter more than the engine in a sports car?

Make sure the gasoline is clean and of decent quality that's all there is to gasoline. The rest is accomplished by the engine.

The algorithm is relatively irrelevant. Much more important are the data with which you feed it. Normally a bunch of indicators or other derivatives of the price curve.

When you selected the features, then you can find the best algorithm for them. All of your list are available in R. You can relatively quickly try them out.

I found deep learning networks the most versatile algo for financial prediction so far. There are several R implementations for deep learning.