I am generally of the same opinion -- specific rules that you can come up with usually perform much better than relying on a machine learning algorithm to learn the patterns. Usually humans do a much better job at perceiving relevant patterns but testing is absolutely critical because humans also often attach too much meaning to insignificant patterns as well.
A very good point indeed! Invention is still out of reach for machines, however, proving the invention being right or wrong is best done with their help!
Oh, 100% agreed. I was thinking of using them in the context of validating rules and finding the right parameters around rules, as opposed to finding scenarios in the first place. I am not a believer in the finding-rules use for NNs. Like say I believe when X happens, Y happens. But what are the conditions around when X happens that Y happens? This is the question that I am interested in the use of NNs for.
Certainly, I actually wrote a book in 1987 on this subject; it's too bad that it is not in English (I never got to translating it). In short: When you go to the library, for example, there are shelves with books placed according to their subject, alphabet order, ISBN number etc. The groups of similar books on the shelf are taxonomical classes; the entire system of shelves with the books on them is the library's taxonomy (a set of book classes). Now, if you pull out a catalog of books that only has book descriptions (also arranged according to some system) then the classes of book names is their meronomy. In order to create the new class of books on the shelves (for example when a brand new book subject emerges that did not exist before) the new entry in the meronomy system is needed first to proceed with the new taxonomical class of books and their book shelf space. Meronomy, therefore, is a type of hierarchy that deals with part-whole relationships; the relationships that could evolve and accept new members without restructuring the entire classification system. I hope it was of some help. Cheers, MAESTRO P.S. Thank you for your reply, you are too kind!
In this case you do n't need NN. All you need is Bayesian tools and straight forward "if - then" type of logic. Have you tried Prolog as a tool?
This is the case according to my limited experience (just 25 years): Step 1: Data mining: find rules (here you have data mining bias problems) Step 2: NNs: Combine rules (here you have fitting problems) Step 3: Genetic programming and walk forward optimization: get the final system with risk management, etc. (also fitting problems but also selection bias). Actually, NNs are now redundant because GPs can also do step 2. Examples: in the following link, the secong program (PAL) finds the if-then rules. The first program (TSL) finds the final system. There are cheaper solutions for the first program. One is the Grail Genetic Optimizer, I think part of TS now. http://www.elitetrader.com/vb/showthread.php?s=&postid=2968698#post2968698 This is how some hedge funds are bulding systems at a high throughput. One other subject we haven't discussed in ET is the need to change systems constantly so to avoid risk of ruin. I was going to get to this in the thread I started about Trading Math.
I am using bayesian tools currently and am trying to expand my toolset, but having some trouble pushing my signal integrity past a certain point. The current trade is profitable but I'm not happy about the margins I am getting and want a smoother equity curve.
It is really a rule-induction problem and there are algorithms and tools for that. For example, one can use GA to evolve an expression tree. Or one can use a tool like RapidMiner to induce a set of rules over some data. In the end, it is all about what you want the model to learn.