In this issue:
I. New Releases of AI Trilogy – Windows Vista compatible
II. Editorial Commentary by CEO Steve Ward
III. From Brain Scans to Trading: It’s All About the Patterns
IV. Legacy NeuroShell 2 updated for Windows Vista
V. Complete list of changes in the new AI Trilogy
VI. One way to stop this newsletter.
*******************************************************
I. New Releases of AI Trilogy – Windows Vista compatible
The new releases of NeuroShell Predictor and Classifier break the 16,000 row limit on training data, letting the computer’s operating system allocate space for the data. In addition, changes in the genetic training method allow the user to select swarm optimization and evolution strategy in addition to GeneHunter as different methods for smoothing input values. A complete listing of the changes is in article V (because it is long).
The update prices for the new releases are as follows:
AI Trilogy – $400
NeuroShell Predictor – $250
NeuroShell Classifier – $250
NeuroShell Run-Time Server – $250
GeneHunter – $250
If you purchased or upgraded during 2007, the updates are free if you order before July 1, 2007. If you purchased or upgraded during 2006, you are entitled to a 30% discount if you order before July 1, 2007. If you order on or after July 1, the cost of the update will be $400 for the AI Trilogy and $250 for all other components.
Simply enter an online order at www.wardsystems.com for the appropriate update. In the comments box, request the appropriate discount and include your product serial number.
All updates are delivered via Internet download. You may call us at 301 662 7950 if you have any questions or to place an order.
*******************************************************
II. Editorial Commentary by CEO Steve Ward
1. Seminar. Marge Sherald and I consider our recent NeuroShell Trader Advanced Seminar to be the best we have ever had. In our opinion, the attendees had a great time, and so did we. As in the past it appears that attendees got as much out of the conference from each other as they did from us – probably more. I know I learned a lot of things I didn’t know about trading from the attendees, both in the sessions and in the informal discussions around the pool. Among the attendees were three “real traders” from our website, one that has just become a “real trader” (see the next article), and five or six more that probably should be. We also had a world class neural network expert, and users from as far away as Sweden and Singapore. There were several who had previously attended our seminars, some more than once. Of course, there were many new users as well. The bottom line is that it was so much fun for everyone that we want to do another one in the fall, probably closer to the West Coast this time.
2. The new AI Trilogy. I thought I would comment on a couple of my favorite things about the new NeuroShell Predictor and Classifier, part of the new AI Trilogy. I use it mostly for building trading models.
This may seem minor, but I really like the Predictor’s ability to minimize the % same sign. This means that when predicting, the objective function in the genetic method will try to maximize the number of training exemplars where both the actual and predicted are positive numbers, or both are negative numbers. This is useful especially for building trading models. If the output is percent change in price, for example, you aren’t as much interested in getting the actual amount of change right as you are in getting the direction of change right. Hence use of maximize % same sign.
The genetic method now lets you specify the number of individuals in the genetic breeding pool in both Classifier and Predictor. We were sort of forced to make this user specifiable when we added the ability to have more than 16K training rows. The neural method with TurboProp2 is extremely fast, no matter how many rows you have. The genetic method, however, trains much more slowly as the size of the training set increases. Once I trained with about 40,000 rows and I had to set the genetic population down to about 5 or 10 to have any chance of training in less than a few hours. Fortunately I could do this, and also fortunately the genetic method has very few free variables that need adjusting during training, so you don’t need a great number of generations to get a good result.
We also now have three methods of optimization in the genetic method – GeneHunter, Evolution Strategy, and Swarm Optimization. The swarm name comes after the inspiration for the method – the swarming of insects like bees. Which one is best may depend on the problem, but generally I use Evolution Strategy, but Swarm sometimes provides a more robust solution.
We have had some discussions internally about selling only the AI Trilogy bundle, and not the component parts (Predictor, Classifier, GeneHunter, RunTime Server) individually. If you wanted any of these parts without buying the whole bundle, you might want to purchase soon in case we do decide to sell only the bundle at some future point.
*******************************************************
III. From Brain Scans to Trading: It’s All About the Patterns
Donn Fishbein started developing mechanized trading systems because of an interest in the scientific method and systems that could be tested. “The idea of letting the computer do the thinking appealed to me.”
The influence of the scientific method seems natural to someone with his background. He earned a medical degree from Jefferson Medical College in Philadelphia in 1981. He also earned a Ph. D. in neurophysiology at George Washington University in Washington, D.C., in 1986.
While working on his Ph. D. in Washington, he was affiliated with the National Institutes of Health from 1984 to 1986. One of his projects involved working with positron emission tomography (PET) scans to identify specific regions in the brain. He began using neural networks for automated image analysis. Because there were no commercial products available (the first NeuroShell didn’t appear until 1988), he wrote his own neural network for the project based on the backpropagation algorithm. The program ran on the mini-computers of the time and was moderately successful.
During this period he read about using neural networks to predict time series data, i.e., stock market data. He decided to use a neural network to predict the S&P 500 as an example for a talk he gave at NIH on the PET scan project. He noted that while the neural network for the image recognition portion of the talk generated a few questions, the S&P prediction section produced a lot of questions. “I decided that if intelligent people thought there was something to it, I should pursue work in that area,” said Fishbein.
He began experimenting with some general purpose neural net tools. Seven years ago he purchased the NeuroShell Trader. “I thought it was a lot of money at the time, but it was worth every penny in time savings. This tool was designed for traders,” according to Fishbein.
While working as a pathologist, he designed some trading systems for Exchange Traded Funds (ETF) such as the QQQQ, SPY and IJR, the S&P 600 small cap fund. In 2000, that led to a talk at the Trenton Computer Festival, which was started by friends of his. The talk described how neural networks and genetic algorithms can be used in trading. He showed some sample systems, and people asked him how to get signals from the systems on a regular basis. He built a web site called www.NeuroQuant.com, and uploaded the signals on the weekends. He had an increasing number of requests to receive the signals on a daily basis, so he started a subscription service.
Currently, his original QQQQ system is showing a 36.1% annual return compared to 8.1% for buy and hold. A second QQQQ system shows a 25.5% return. His SPY system shows an 18.4% annual return compared to a buy and hold of 8.1%, and the IJR system has an annual return of 33.4% compared to 10.6% for buy and hold.
He said he’s had good success with the GRNN nets from the Trader’s Adaptive Net indicators add-on. He uses inputs from the Trader’s Price Momentum Indicator category as well as linear regression indicators.
He recently attended the NeuroShell Trader Advanced seminar in Florida. “I had a great time at the conference. The only con is that I didn’t ID granny. I just got the Adaptive Turboprop 2 add on; there’s a lot of promise there.”
His next projects include developing sector switching systems and intraday signals. “I work with things a while and test before going public. I haven’t had to take a system back yet because it didn’t work.” He is also in the process of launching a hedge fund based on neural network and genetic algorithm trading systems.
*******************************************************
IV. Legacy NeuroShell 2 updated for Windows Vista
We have updated the legacy NeuroShell 2 product so that it is now compatible with Windows Vista. This new release 4.2 also has a few other changes as outlined below. The update price to release 4.2 for old users of release 4.0 or 4.1 of NeuroShell 2 is $125. That fee will be waived for any user who purchased NeuroShell 2 in 2007 as long as they update on or before July 1, 2007.
For owners of releases below 4.0 the update price is $450.
All updates are delivered via Internet download. You may call us at 301 662 7950 if you have any questions or to place an order.
Other changes in release 4.2
1. Example Excel worksheets are saved in the Excel 97-2003 format to provide better compatibility with Excel 2007 and future Excel releases (at some point Microsoft may decide to stop reading older Excel formats).
2. The Predict macro add-in for MS Excel (to call .def files in Excel) has been modified to work in Excel 2007.
3. Run-time programming examples have been updated with examples from: Delphi 6, VB6, and
VC6.
4. A new additional run-time example for VB.NET has been included.
*******************************************************
V. Complete list of changes in the new AI Trilogy
Here are the AI Trilogy changes between release 2.0 and 3.0 (grouped by product):
——————————————————————————
NeuroShell Predictor
New Graphics
The graphics have been updated and improved.
Row Limits
An artificial limit on the number of data rows (16000) is removed. The new limit is defined by the operating system only. It is not, however, how much RAM is installed or the size of your page file. The operating system puts limits on how much memory each application is allowed to use for all of its resources (graphic objects, data arrays, etc).
Additional Genetic Training Methods
The Genetic training method has two new optimization options: Evolution Strategy and Swarm. These two new modern optimization techniques complement GeneHunter’s classical genetic algorithm used in previous versions.
Control Over Breeding Pool Size
The Genetic training method now allows manual control over breeding pool size and its random initialization (random number seed).
More Control Over Data Selection
Random selection of training/out-of sample data sets is now allowed in addition to the existing methods. Random selection is done by using either the number of rows or percentage points. When the random rows option is selected and the user applies a trained net, there is an option to choose which rows to apply the net to: out of sample, training or all available rows.
New Optimization Function
Predictor only: The Genetic training method has a new optimization function: Maximize percent same sign of actual and predicted values.
Sound Alert for Genetic Method
The Genetic method makes a sound alert if training has ended because of the “Stop after N generations without improvement” criterion. This is helpful in the case of lengthy genetic training sessions to alert you when training has finished. The alert can be selected from system sound files (.wav files) available from your Windows\Media folder. Use the test alert button to play the sound and decide which one is best suited for your taste and environment.
3-D Graphics Add-On
The 3-Dimensional Graphics Add-On is an additional program for NeuroShell Predictor. It is designed to display a projection on the screen of a 3-dimensional surface (response curve), formed by a trained neural network output. The network output (dimension Z in 3D space) is calculated while varying two selected input variables (dimensions X and Y) and while keeping all other input variables at some fixed values.
Supported Operating Systems
Windows Vista, Windows XP and Windows 2000 with SP4 are the only supported operating systems. The software may still work on Windows 95/98/Me/NT, but its performance is not guaranteed.
——————————————————————————
NeuroShell Classifier
New Graphics
The graphics have been updated and improved.
Row Limits
An artificial limit on the number of data rows (16000) is removed. The new limit is defined by the operating system only. It is not, however, how much RAM is installed or the size of your page file. The operating system puts limits on how much memory each application is allowed to use for all of its resources (graphic objects, data arrays, etc).
Additional Genetic Training Methods
The Genetic training method has two new optimization options: Evolution Strategy and Swarm. These two new modern optimization techniques complement GeneHunter’s classical genetic algorithm used in previous versions.
Control Over Breeding Pool Size
The Genetic training method now allows manual control over breeding pool size and its random initialization (random number seed).
More Control Over Data Selection
Random selections of training/out-of sample data sets is now allowed in addition to the existing methods. Random selection is done by using either the number of rows or percentage points. When the random rows option is selected and the user applies a trained net, there is an option to choose which rows to apply the net to: out of sample, training or all available rows.
Change in Agreement Matrix
The Agreement Matrix displays the Positive Prediction Value statistic which is yet another measure of how well the model makes correct classifications in each category. The higher the Positive Prediction Value the better the model’s predictive power.
Sound Alert for Genetic Method
The Genetic method makes a sound alert if training has ended because of the “Stop after N generations without improvement” criterion. This is helpful in the case of lengthy genetic training sessions to alert you when training has finished. The alert can be selected from system sound files (.wav files) available from your Windows\Media folder. Use the test alert button to play the sound and decide which one is best suited for your taste and environment.
Supported Operating Systems
Windows Vista, Windows XP and Windows 2000 are the only supported operating systems. The software may still work on Windows 95/98/Me/NT, but its performance is not guaranteed.
——————————————————————————
NeuroShell Run-time Server
The NeuroShell Run-Time Server is made compatible with Microsoft Windows Vista and Microsoft Excel 2007.
Three new methods are added to nsrun.dll and the nsrun.ocx control: PredictorStats, ClassifierStatMatrix, and Prob2ClassName. These methods help to compute various kinds of statistics upon actual and predicted data.
——————————————————————————
GeneHunter
1. GeneHunter is made Microsoft Windows Vista and Microsoft Excel 2007 compatible.
2. All Visual Basic 5 examples are updated with Visual Basic 6.
3. The Microsoft Visual C 4.0 examples is updated with Visual C 6.0.
4. All Borland Delphi 3.0 examples are updated with Delphi 6.0.
5. All Microsoft Excel example workbooks are updated with Excel 2003 (they are re-written in Excel 97-Excel 2003 workbook format).
*******************************************************
VI. One way to stop this newsletter.
It is really easy. Just change your email address and don’t tell us.
*******************************************************