AI-powered Trading System

[MarioTang]

Can someone recommend a proven AI-powered tool? I extensively tried TradingSolutions for a couple of months but haven't managed to create a profitable strategy based on neural nets. I'm thinking about purchasing Alyuda Tradecision Pro (www.tradecision.com) or NeuroShell Trader Pro (www.neuroshell.com). What would be your recommendations? Thank you.

Regards,
Mario Tang
[email protected]

 

[Rognvald]

The crucial words here are "proven" & "tool"

I cannot and I don't think many here can

Read
http://www.trade2win.co.uk/boards/showthread.php?s=&threadid=8886
section 1.2.1 very carefully and then think about the software you are contemplating

Good luck

 

[Hwyman]

Tradecision or Neuroshell

Can someone recommend a proven AI-powered tool? I extensively tried TradingSolutions for a couple of months but haven't managed to create a profitable strategy based on neural nets. I'm thinking about purchasing Alyuda Tradecision Pro (www.tradecision.com) or NeuroShell Trader Pro (www.neuroshell.com). What would be your recommendations? Thank you.

Regards,
Mario Tang
[email protected]

If you have decided to purchas eone of the two and you are trying to decide which one, I would recomment Neuroshell. It has been around for many years and it is built on the basis of a generic neural net program which has been used in many different fields of science so at least you know you are buying a robust product. Tradecision on the other hand was created recently by enterpreneur Russian programmers. I have some experience using Neuroshell and no experience using tradecision. In short, it may take you years of hard work to create a good predictive model using neural networks. While neural networks have been used successfully in other areas of science, medicine, etc. we still have not seen a breakthrough when they are applied to trading. From my not so good understanding of the issue at hand I believe it is due to the nature of the financials "time and sales" type data which suffers from too much noise, expternal influences and interruptions, whicle neural networks sort of assume a continuous stream of values in a system isolated from the external world where patterns are easier to identify and classify. neuroshell alone will offer you so many possibilities for inputs/outputs/training targets, etc that you can get lost looking for a goldfish in the atlantic ocean. The prodict is great, as long as you grow above the expectation that a "artificial intelligence model" will quickly identify proftitable opportunities and make you money.

 

[Directional]

Perhaps check out the SirTrade Safir products - this isnt an endorsement, and I dont use it myself, but I've seen it spoken highly of in the TradeStation forums and it sounds like the sort of thing you are asking for.

www.sirtrade.com

Its a high end product with a matching price tag though - the Pro version costs a hefty €30,000

 

[Hwyman]

Perhaps check out the SirTrade Safir products - this isnt an endorsement, and I dont use it myself, but I've seen it spoken highly of in the TradeStation forums and it sounds like the sort of thing you are asking for.

www.sirtrade.com

Its a high end product with a matching price tag though - the Pro version costs a hefty €30,000

I have a tradestation account also and from time to time read the postings on various topics. I have found tons of helpful information on all sorts of topics. I am a little sceptical about sirtrade. it is not as robust as the dedicated neural net products and it lacks in terms of ability to thoroughly test a model on out of sample data for example. One thing about neural nets is it is very easy to build an "overfitting" model which works perfectly on "in-sample" data but when applied on "out of sample data" performs very poorly. If you look at Trading solutions, the product allows you to separate the data into training set, evaluation set and out of sample, "testing" set. neuroshell goes a step further. it lets you devide the data just like Trading solutions, but then move a step forward and change these periods by let's say, a month into the future, providing the most lifelike "out of sample" peformance report possible. For someone not familiar with neural networks (I don't think I am, so i am trying to educate myself) it is very easy to be mislead into thinkiing that you have built a good model while in fact all you have is a model of the past with little or ability to work on out of sample, future data. In other words, the tools that should come with a neural network program to help you evaluate trading strategies are quite different from the tools you need with "linear" strategies. In the same manner, the traditional backtest report does not mean much. An otherwise worthless model may provide you with a perfect backtest report. I don't believe sirtrade provides the functionality you need, especially at that rediculously high price. I am just now starting to evaluate different neural product such as neuroshell, Biocomp profit, Trading solutions, etc. if other forum members are interested, i would be happy to share my findings and results.

 

[Tuffty]

Hwyman, Any feedback you have on various Neural net software would be much appreciated. Have you found any literature that has been useful? Cheers Tuffty

 

[Hwyman]

interesting archive on using AI tools

Hwyman, Any feedback you have on various Neural net software would be much appreciated. Have you found any literature that has been useful? Cheers Tuffty

This is a link the the letter archives of Jeffrey Mishlove. In each letter he explored the use of various indicators to feed in an AI trading system. His findings looked very interesting. Unfortunately his endeavours in the field started in 2002 and ended in the middle of 2003. I believe he is involved in activities not related to trading these days. For those with some time and patience these 20 or so "letters" contain some very interesting insights and practical experiments:

http://www.forecastingsystems.com/2003/Archive2003.htm

 

[jalsck]

Hi,

I run my market timing models for the SP contract using BioComp Dakota. Previously, I used neural network models. Some of my older work was featured on Jeffrey Mishlove's website.

Dakota is based on particle swarm adaptation. You can implement your own models / indicators using ScriptBots and take advantage of the walk-forward adaptation. Particle swarm adaptation is a big step up from other forms of parameter optimization. I've built my own swarm adaptation library for Dakota and I'm thrilled with the results.

You can check-out my work at AdaptiveTradingSystems.com. I'm very keen on predictive modeling and enjoy a good chat / discussion on the subject.

Kind Regards,

James