Choose a Capability
The following is an overview of our software products by specific capabilities:
Data classification means deciding whether a set of data belongs to one of several categories. It is distinct from prediction, which simply attempts to forecast some data stream into the future, like how NeuroShell Trader predicts prices or indicators. Categories could be concepts like "Pass/Fail", "Excellent/Acceptable/Unacceptable", "Bull market/Bear market", "Buy/Hold/Sell", "Doctor/Lawyer/Banker/Trader/Clerical/Bus driver", or pretty much anything you can think of.
The NeuroShell Classifier is our flagship in this area:
However, NeuroShell2 (see item 2 below) can also make classification models:
If you are a NeuroShell Trader user, the Adaptive Net Indicators add-on can make two category classifications (e.g. Buy/Sell or Up/Down):
In fact, ChaosHunter also can build classification formulas, where the classes are distinguished by whether the formula produces positive or negative values (e.g., tomorrow's price is rising or falling):
Study of classic neural networks
Our oldest general purpose neural network product that we still sell is called NeuroShell 2, first built in 1993. We sold thousands of these, many of which are still being used. The big appeal of NeuroShell 2 was that it is a neural network experimenter's kit, creating neural network interest and learning opportunities by providing a wide range of neural net paradigms, all very "tweakable". We sold so many to colleges and universities that many of the neural network experts of today got their start with NeuroShell 2 in the classroom. Although we still sell NeuroShell 2, we try to make sure we sell to the right people - those who want to wade deeply into a classic neural net education as opposed to someone whose main interest is to solve a problem or make money.
Fully evolutionary neural nets
Except for the neural networks in ChaosHunter, most neural net implementations are just a static or predictably incrementing neural network structure combined with an internal optimization technique that finds appropriate weights. ChaosHunter actually evolves the neural net structure as well as the weights. In fact, the weights themselves can even be structures of some type in a totally flexible neural network framework:
Basically we are using techniques similar to genetic programming to let evolution invent its own neural network algorithm.
The old NeuroShell 2 actually has a predecessor evolutionary technique involving regressions that is called Group Method of Data Handling (GMDH) or Polynomial Nets. Use this link and scroll down to GMDH Architecture:
Fundamental stock picking
If you think about picking stocks from a huge basket based primarily on fundamentals (like Warren Buffett stock picking), you should come to the conclusion that you are simply classifying winners and losers. Why not use a classification neural net with fundamental inputs? Here is but one way using the NeuroShell Classifier:
We wish we could tell you about some very well known investment companies that use that precise technique, but
we are non-disclosed with regard to naming names because we took consulting fees to help them do the work, although it isn't rocket science as you can see.
General purpose, portfolio, and rule optimization
Genetic algorithms (GA) like GeneHunter can perform those tasks and many more, including scheduling, fuzzy logic models, neural network training, classification, prediction, and trading rule creation and selection. This list is really as long as your imagination. What's the catch? The catch is that YOU have to figure out how to formulate your task as the optimization problem it really is. However, the examples here should give you some ideas:
Those who are interested in portfolio optimization realize that most all of the techniques in the literature involve, well, optimization. So if you can code your favorite portfolio optimization technique in Microsoft Excel, you should be able use GeneHunter to optimize it to complete the process.
Long ago we realized that the NeuroShell Predictor, NeuroShell Classifier, and GeneHunter working together are a powerful combination that solves a very common problem: what mixture of product components makes the best final product? Our inspiration was a real farmer who wanted to figure out how to mix the ingredients in his cow feed so as to maximize milk production. First you need a neural net to tell you how much milk each different mixture produces. Then you need GeneHunter to discover a mixture that optimizes milk production. We combined several existing products to make the AI Trilogy to solve the AI Trilogy problem:
The NeuroShell Trader Professional programmers did not specifically have in mind the classical market neutral pair hedge trading technique when the product was developed. However, it can be done and a tip on www.ward.net shows how it can be accomplished. It does require a very good understanding of how NeuroShell works. Later we developed a method of using a neural net to help find the best qualified pairs, and added some special indicators in Advanced Indicator Set 3 to make the process a little easier:
Portfolio "basket" trading
Portfolio trading is very similar to pair trading because you can trade hedged pairs, and do it a lot easier even than when using Advanced Indicator Set 3. You can also let NeuroShell select the hedged pairs from a basket. Instead of hedged pairs, you can have hedged quadruples, where the software goes long two and short two, with the four being picked from the basket. Instead of quadruples you can use 6 or 8 or more. The difference between this and classic pair trading is that you are selecting stocks and trades by comparing independent indicators applied separately to each stock in the basket, instead of using spreads between specific stocks. Some of you are no doubt saying "Why didn't I know about this sooner?" I don't know because it is taught right there in NeuroShell examples 18, 19 and 22. There is another example on ward.net under the section "NeuroShell Trader Tips from Active Traders Magazine and Other Sources", which is called "International ETF Trading System (March 2011)."
Some years ago a company formed by one of our users began selling a product that made stock selections using fuzzy logic applied to indicators. Some people say our software is expensive, but this thing sold for $25,000 way back when. So in a few weeks' time I built the Fuzzy Sets Add-on ($299) so our users could do much the same thing. The idea is that you want to use rules like "Indicator A is high and Indicator B is very high and Indicator C is low". Then you want to optimize and backtest the rules:
Before fuzzy Sets, the first fuzzy logic add-on we made came about because there was a time when practically every couple of days we'd get a call from some prospective user who would describe the pattern he wanted to recognize, and ask if our neural nets could find that pattern for him. The pattern was usually described in terms like "The prices rise sharply, then drop sharply, then drop again". We got tired of saying neural nets find their own patterns, so we built the Fuzzy Pattern Recognizer to let the user enter those descriptions in a simple coding scheme:
which is not to be confused with something very similar called the Pattern Matcher:
CLICK HERE which actually lets you point out on the chart a representative pattern you want to be recognized in the future, among other things.
Everybody knows what this is; it is predicting tomorrow's rainfall, the change in the DOW next week, or the company sales for next month. You use regression or neural nets. The NeuroShell Trader Professional does this with neural nets, and ChaosHunter does it with either neural nets or symbolic regression using evolutionary or swarm techniques. But before those flagship products, we had the NeuroShell Predictor, which is still very popular because it is less expensive and oriented towards business and science predictions:
It can even do 3D graphics sensitivity analysis.
NeuroShell2 (the oldest product) also does neural net predictions of course:
The NeuroShell Trader add-ons Adaptive Net Indicators and Adaptive Turboprop2 do rolling predictions with automatic continuous retraining.
Panel of experts
You use a panel of expert modeling techniques like you would use a group of experts on your staff to make a decision. Usually you look for a large majority, but sometimes a simple majority will do to convince you to make the right decision.
The NeuroShell Trader Professional can make panel decisions, and example 20 shows how it is done with neural nets. (Trading strategies are much harder).
We have taken the process to a whole new level with the ChaosHunter Trader:
This system makes it very easy to trade a large number of models you have built, looking for consensus. No rules are necessary; all the consensus logic is built-in and automatic. There can be multiple consensus systems.
Program your own neural network software
GeneHunter can be programmed and has examples of building neural nets trained via optimization:
However, if you really want to program your own neural net software using the algorithms in NeuroShell Trader Prediction Wizard, NeuroShell Classifier, and NeuroShell Predictor, you need the NeuroShell Engine:
You must first purchase one of the NeuroShell products that uses the Engine. Prices vary by application, but if you have to ask how much it costs, you probably cannot afford it.
One of the oldest un-supervised clustering techniques is called the Kohonen Self Organizing Map, a simple version of which can be found in NeuroShell 2:
It is great for education on self-organization, but I found that doing unsupervised clustering with GeneHunter is more effective. Two examples are distributed with GeneHunter:
NeuroShell Trader Professional users have had some good results with supervised clustering using the Cluster Indicators add-on:
Basically, you want to find out where a set of indicators tend to cluster before major market moves occur.