Machine learning in trading strategy?

Question: Does anyone here apply machine learning techinques (such as Neural networks, PCA, local sensitivity hashing) to his trading strategies?

There are two main commercial vendors of AI models for retail traders:
Neuroshell and TradingSolutions.

I think that they are offering a very easy tutorials allowing to grasp the use of those models in trading systems.

On both sites there is a lot of free information about basic models. In trading solution there is even a demo version which gives you several models and gives you the opportunity to make new models.

If I can summarize it in really rough way basically the basis of everything is:

* To select the predictive inputs (mainly technical indicators) and try to predict with them something else output.

* Use the output as a trading signal.

For example:

You have 2 inputs
You use a moving average of period of 20 as input.
You use a second moving average with period 10 as input.

You want to predict the next day close.

So the rule may be if the predicted next day close is higher than the next day open then buy.

If the predicted next day close is lower than the next day open then sell.

You train the model. And you backtest the model over a past unseen data.


So that is the basis and from there start the variations including their use in portfolio.


I have two remarks:

* The problem is that trading with neural networks is not an easy task. My opinion is applying a model over an already viable strategy or idea will allow to squeeze more profit from the idea compared to the rough model.

* Another problem I may say is luck related. Very often you can confuse the optimization with the luck. Very often the results you get are purely due to the luck. Monte Carlo simulations may help a lot.

Please check the book http://www.evidencebasedta.com/ and Monte Carlo
Permutation Evaluation of Trading Systems
http://www.evidencebasedta.com/MonteDoc12.15.06.pdf

I just want to add another remark that by using commercial software within the software is integrated a fool proof.

In Neuroshell for example they are offering to you some default settings that will limit the overfitting. In trading solutions it is the same. If you use your own model or an open source software and you do everything by yourself you need to be carefull what are you doing, for example adding too many weights to the model and too many neurons and hidden neurons will increase the risk to overfit the data and not to learn from the data.