Why rule induction instead of something else? Why rules discovered by genetic algorithm, specifically?
Mechanical vs Neural trading systems
- Thread starter pandabear
- Start date
Why rule induction instead of something else? Why rules discovered by genetic algorithm, specifically?
Nonsense! Properly constructed neural networks do not "go haywire". Overfitting is a beginner's mistake. If training is controlled, overfitting will not occur.
By far, most artificial neural networks are used as statistical regressions. An understanding of inferential statistics, especially various regression procedures (nonlinear regression, kernel smoothers, etc.) would help.
You can find a good, intuitive explanation of backpropagation (the most common neural training procedure), including pseudocode, in "Neural Networks for Statistical Modeling", by Murray Smith (ISBN-13: 978-1850328421). This book also explains two methods of avoiding over-fitting.
In my opinion, the most important aspect of empirical modeling is model testing. "Computer Systems That Learn" by Weiss and Kulikowski provides an excellent introduction to several rigorous test procedures.
This premise is false.
The first problem with all mechanical systems...
Is they fail under the weight of 4 or 5 SD events...
And you take BIG losses.
By definition...
No AI "adaptive/learning system" can handle rare events...
Like the crisis of last 6 months...
Which happens once-in-a-generation.
The second problem...
Is that a human expert will always beat a mechanical system over the medium term...
Whether it's sports betting or poker or trading.
So you are building simple mechanical systems...
In order to compete with expert traders ....
That are also running automated systems with 100 times your tech budget.
It's a hopeless scenario...
You may have marginal success... but so what.
Smart people who just BUY the automation as a "holy grail" line...
Will waste 10 years of their life before they "get" what I am saying.
SOLUTION
The way to go is to become an expert trader FIRST...
And then gradually automate what you do.
But that is just too much work for most.
Murray Ruggiero
Sponsor
I totally agree. I did not become successful using machine learning in the mid 1990's until I could develop robust systems without any advance technologies.
Most people try to make these advance technologies do too much. When using these technology , Advanced components need to be system fault tolerant. If an advance model fails, the system performance degrades but does not collapse.
Why should we think that? There are certainly plenty of examples of machines out-performing humans at information-processing tasks in other fields. Why shouldn't they do the same in this field?
