S&P Trading Model - Quite Technical - Meet the Competition

[UrmaBlume]

http://forumserver.twoplustwo.com/sh...age=0&vc=1

[UrmaBlume]

NDA's are worthless in this arena.

I have certain trading relationships and some of them are p*ssed that I post at all.

[crazy canuck]

I am an expert in Genetic Algorithms. I could publish in the field since I used it for years. You're right that it can provide an edge, but the code you posted has nothing to do with GA. You just posted a system that uses a lot of inputs.
Actually, it is kind of contradictory with you using GA, because with that many inputs you'd need years of data to avoid overfitting. But like you said a fitted system loses effectiveness over time...

[UrmaBlume]

GA can be applied in many areas. While this particular model doesn't reflect GA, survival of the fittest algorithms were used when selecting the inputs.

To avoid overfitting is modeling 101. As to the amount of data, it is not the amount of time that is required, it is the quantity of records/facts for training/development that is relavant. If you create one record/fact per day you would indeed need many years of data. If you create 500/1000 records per trading session then you would need considerably less than 1 year.

I use a ratio of 100-300 records per input, not including test data, in the developement process.

Inputs for these types of operations require normalization both historically and as to time of day.

For all those years of experience you seem to have missed the most effective uses of GA in this arena.

Also, I didn't post any system. I only posted the algorithm for one model that is but a part of a system.

[mrbaseball]

[ QUOTE ]
I believe my posts present information not generally available to the public. My purpose here is to discuss/share and in some cases mentor.


[/ QUOTE ]

I haven't seen much information? Just vague references to some sort of (possibly) automated trading system using methods you aren't willing to discuss which doesn't really lend itself to discussion.

[eastbay]

[ QUOTE ]
GA can be applied in many areas.

[/ QUOTE ]

Yes, but it is always just an optimizer, and often not even the best choice of one. GA is a totally uninteresting topic for a trading board, because the choice of optimizer is mundane and beside the point. What are you optimizing, is the question. Let me guess: you aren't saying.

Snooze.

eastbay

[UrmaBlume]

Relative to trading you can use GA to optimize: inputs to neural networks, settings to neural networks, time frames, stttings to MA's & ADA's & other oscillators, stop losses, take profits, holding periods, exit times, legs in & legs out, options spreads, ratio spreads and that's just to name a few.

What is the real Snooze is a total lack of vision and imagination as regards powerful algorithms and how they are deployed.

[eastbay]

[ QUOTE ]
Relative to trading you can use GA to optimize: inputs to neural networks, settings to neural networks, time frames, stttings to MA's & ADA's & other oscillators, stop losses, take profits, holding periods, exit times, legs in & legs out, options spreads, ratio spreads and that's just to name a few.


[/ QUOTE ]

Actually you can use GA to optimize anything. The question is: what works.

[ QUOTE ]

What is the real Snooze is a total lack of vision and imagination as regards powerful algorithms and how they are deployed.

[/ QUOTE ]

Assumptions, assumptions.

eastbay

[crazy canuck]

[ QUOTE ]
GA can be applied in many areas. While this particular model doesn't reflect GA, survival of the fittest algorithms were used when selecting the inputs.

To avoid overfitting is modeling 101. As to the amount of data, it is not the amount of time that is required, it is the quantity of records/facts for training/development that is relavant. If you create one record/fact per day you would indeed need many years of data. If you create 500/1000 records per trading session then you would need considerably less than 1 year.

[/ QUOTE ]


GA is useful, however in most cases it is far from an ideal method for input selection (as eastbay said). It is much better to do a principal component analysis to make sure the inputs are not correlated.

And it is terrible for input selection for neural networks. NNs tend to overfit even without optimizing for input. And cross-validation using "training/test/test" samples is not very effective. Might work, but again it is not ideal.

Also, I was trying to point out that I don't get the purpose of your post. It didn't give any useful info for somebody trying to implement a system.

[hapaboii]

[ QUOTE ]
This particular model uses calculations from advances/declines, UpVol/DnVol, cash indexes and other sources. The model is run every 60 seconds and produces a fresh prediction every 60 seconds. As trading is different from hour to hour over the 24 hour trading day different models are developed for different time periods.

[/ QUOTE ]

[ QUOTE ]
EVERYTHING I do in the market is based on 2 basic priciples:

1 - Information = Equity. Everytime a trade is entered a certain equity is assumed. If the information that is the basis for the trade is good then the equity will increase.

2. Buying and Selling Pressure. It is buying and selling pressure that drives price. If you can precisely measure the net buying and selling pressure in any time frame then you can predict the price.

[/ QUOTE ]

For those looking to filter Urma's posts, I think the quotes above can lead to an interesting discussion. Basically, I think this gives an idea of information that he believes can be used to provide a statistical edge in automated trading. Everything else(the "sample" code) is pretty useless, perhaps meant to make his posts/knowledge/systems look complex while throwing around words like GA, AI, NNs a bit too much.