Re: [amibroker] Artificial intelligence

Coolman 17.43

Greetings --

Most of the comments confirm my caution that neural networks tend to overfit. Careful attention to modeling and simulation technique, including rigorous protection and limited use of out-of-sample data, is very important. Neural networks typically have very high in-sample accuracy and considerable degradation out-of-sample.

I recommend beginning with a system that uses a single output node and predicts a single binary target -- beLong (1) or beFlat (0). For a couple of reasons:

1. Multiple output nodes add considerable complexity to the model. This increases training time, decreases the accuracy of the prediction of each of the outputs, and increases the likelihood of overfitting.

2. Bottoms are different than, and easier to detect than, tops. The indicators / patterns that identify a bottom and a good time to be long are different than the indicators / patterns that identify a top and a good time to be short.

If you have a data point for every day, which is typical, you have a signal for every day. You are working with state signals, not impulse signals. (That is a good thing.) Your forecast horizon is one day. While it is tempting to forecast several days ahead, overlapping holding periods and overlapping signals complicates both prediction and trading management considerably. After you get a model that accurately predicts beLong / beFlat one day ahead, you might consider more complex models. Try to achieve a minimum of 65% accuracy out-of-sample for one-day ahead long/flat prediction.

Best,

Howard

On Wed, Feb 3, 2016 at 9:13 AM, Pankaj Sharma sharma.pankaj.kumar@gmail.com [amibroker] <amibroker@yahoogroups.com> wrote:

Ron,

I have used NN with Amibroker, via R.

Generally, i have tried a three tier prediction classification - BUY, SHORT, DO NOTHING. If you are looking to just square off via NN, you could train the NN on SELL,COVER,DO NOTHING. You need to define the output condition that will cause a cover or sell trade. It could be a very specific condition- if the market moves 1% higher, i will COVER and if it moves 1% lower, i will SELL. I have generally found a more ambiguous condition to be easier to model - eg. if NN predicts an upbar, i will SELL and if NN predicts a downbar, i will COVER. The conditions to define the output variable are many, it is probably a function of your strategy... I have generally never achieved accuracies of >70% in prediction, which result in a 40%-45% win-ratio in real life.

Rgds,

Pankaj

On 30-Jan-16 1:26 AM, 'Ron J' ronj@san.rr.com [amibroker] wrote:

My programmer is going to apply neural networks to our systems. What I have directed him to do is to use our entries and let the AI Software come up with an exit strategy.

So far when he tried to build a trading model with AI the only outcome is his learning about them.

Has anyone tried this approach?............Since a lot more effort goes into entries than exits I figure is might be worth a try.

Regards

Ron

RonJaenisch.com

"Success is almost totally dependent upon drive and persistence. The extra energy required to make another effort or try another approach is the secret of winning."

— Denis Waitley: American motivational speaker

All of my emails are for Educational or Entertainment purposes only. If you need professional advice please contact an appropriate professional, none is given via my emails.

__._,_.___

Posted by: Howard B <howardbandy@gmail.com>

Reply via web post

•

Reply to sender

•

Reply to group

•

Start a New Topic

•

Messages in this topic (9)

**** IMPORTANT PLEASE READ ****
This group is for the discussion between users only.
This is *NOT* technical support channel.

TO GET TECHNICAL SUPPORT send an e-mail directly to
SUPPORT {at} amibroker.com

TO SUBMIT SUGGESTIONS please use FEEDBACK CENTER at
http://www.amibroker.com/feedback/
(submissions sent via other channels won't be considered)

For NEW RELEASE ANNOUNCEMENTS and other news always check DEVLOG:
http://www.amibroker.com/devlog/

Visit Your Group