Scientific Consultant Services, Inc.

ORDERING INFORMATION

TO ORDER DIRECTLY:
Email  (donnamccormick@scientific-consultants.com) telephone or fax (631-696-3333) us the following information:

Full Name
Address
Telephone Number
Email address
Visa, Mastercard, or American Express account number and expiration date

Please Note:  We are obligated to charge those of you who live in New York State NYS Sales Tax.  Therefore, please also inform us of your county and sales tax rate when you order.

Questions?  Please do not hesitate to contact us either by email (donnamccormick@scientific-consultants.com) or by telephone (631-696-3333) with any questions.  Be sure to call for info about our historical data, as well as special purpose libraries and even complete trading systems.  We look forward to hearing from you.  Meanwhile, thanks for your interest in our work!

C-Trader Professional^R See the new C-Trader Professional page

High-res Tick Data

N-TRAIN^R Neural Network Development System

The product has had fabulous reviews in Technical Analysis of Stocks & Commodities and in Futures (a comparative review by Mark Jurik). Used by professional traders, hedge funds and CTAs, educational institutions, the government, insurance companies and others. It was employed in the book, The Encyclopedia of Trading Strategies (Jeffrey Owen Katz & Donna McCormick; McGraw Hill, 2000), to develop a "universal forecaster", and in Computerized Trading (Ed. Mark Jurik; Prentice Hall, 1999) for pattern recognition. We have been told that N-Train is by far faster and more reliable than many packages selling for hundreds or even thousands of dollars.

The product comes with extensive documentation as well as a detailed example of how to easily use N-Train from within Excel (includes well-commented Excel VBA macro code). The TradeNet link module that connects N-Train to TradeStation in real-time is available seperately. Many traders find this is the way to go. It should be noted, however, that N-Train is a stand-alone system, in that neither Excel nor TradeStation is required to use it. N-Train can read and write standard text files, and is compatible with all software capable of writing and reading such files (Fortran, C++, Basic, Java, SPSS, Matlab, you name it).  The product requires a Pentium or higher processor and will use all available memory (depending on the problem size) up to 2GB, if necessary.

DETAILED PRODUCT DESCRIPTION
Neural network technology can be found all around us, from the electronic equipment we use to the cars we drive.  Neural nets help solve a wide range of problems, such as trading the financial markets, assessing credit risk, data mining, detecting and analyzing signals, controlling processes in manufacturing, filtering out noise in electronic communication systems, and much more.  Neural nets hold great promise for any application that involves decision making, pattern recognition, classification, or forecasting.  Our company has been involved in this technology since the mid-1980s, when we used neural networks to develop a system for processing data from a 64-channel cardiac monitor we designed and built.  And, in 1989, we were the first to market pretrained neural network forecasting systems for traders!

In 1992, N-Train was released as the first 32-bit neural network development tool.  It was designed by a mathematician for the ultimate in speed, capacity, and stability.  N-Train received the Readers Choice Award in Technical Analysis of Stocks and Commodities and, over the years, has repeatedly proven itself to be the best neural network tool available.  In his review, John Sweeney (former Editor of Technical Analysis of Stocks and Commodities) said the following about N-Train:

This is the real thing.  With N-Train there is no waiting.  The program is very, very fast.  N-Train converged quickly in all the examples I gave it.

I have either used or tried four other commercially available products.  The product I have found most useful for my work is called N-Train.  This software allows you to write data directly into files that can then be scaled and trained.  The trained network can then be utilized in a trading system from within TradeStation.

N-Train will help you to achieve your goals easily and consistently.  Thanks to 80-bit floating point computations carried out in pentium-optimized assembly language, N-Train exhibits extremely good numerical stability. As a result, it trains reliably and quickly on data sets that other products simply cannot handle. Works great on noisy data (such as that from the financial markets), as well as on extremely large data sets (think many millions of facts). Several training models including standard backprop and genetic evolution may be specified, and sophisticated analysis of the importance of each input variable to the model can be performed. Scripts may be written to automate complex training and testing procedures. Lots more.

The product had fabulous reviews in Technical Analysis of Stocks & Commodities and in Futures (a comparative review by Mark Jurik). Used by professional traders and CTAs, educational institutions, the government, insurance companies and others. It was used to develop a "Universal Forecaster" for the book, The Encyclopedia of Trading Strategies (Jeffrey Owen Katz, Ph.D., and Donna L. McCormick; McGraw-Hill, 2000), and for pattern recognition in Computerized Trading (Ed. Mark Jurik). This is the neural net software that we use for our own options and futures trading.  We have been told that N-Train is by far faster and more reliable than many packages selling for hundreds or even thousands of dollars.

SPEED.  Unless you have vast amounts of spare time on your hands, when developing neural networks you want a system that makes the process as fast as possible.  Neural net tools vary greatly on this feature.  Fortunately, N-Train has been demonstrated to be among the fastest (if not THE fastest).  In fact, N-Train is hundreds of times faster than some systems and only takes minutes to train nets that other systems spend hours on!

NO PROBLEM SIZE LIMIT.  Network and problem size are characteristics that vary greatly between products.  N-Train places no limit on the number of neurons, interconnections, layers, or facts.  The only size limitation you may experience will be due to the amount of available memory and the size of your hard disk.

USER CONTROL.  N-Train allows you to have as much or as little control over the behavior of the training process as you desire.  Beginners can rely on built-in default settings:  All the parameters and network characteristics default to optimal values.  Advanced users can modify the defaults and experiment to their heart's content.  You can set learning rates (the degree of learning you want the network to do during each step of the training process), which may be specified globally or on a per-layer basis, and may be set to automatically adapt to optimal values during training for faster performance.  Transfer functions (define how the neuron's output changes relative to the degree of input activation, important for learning and generalization) may also be specified on a per-layer basis.  A variety of error measurements (which define how statistical errors are measured and, therefore, represent factors being minimized during training) may be selected, including asymmetric error functions, unique to this product.

DEPENDABLE.  When other systems fail to train, N-Train succeeds.  Why?  Because N-Train is the most numerically stable system available.  Features that help make N-Train the system you can depend upon include:  checkpointing (nets are automatically saved to disk every few minutes so you won't lose much if power is interrupted); data consistency and validity checking (it automatically checks that the nets are training on the data you think you're training them on); thorough fact shuffling (advanced pseudo-randomization technique eliminates any chance of serial dependence interfering with training).

SAFEGUARDS TO IMPROVE SUCCESS.  If a network is overtrained, it could just memorize the data you feed it without learning anything.  Without learning, nets won't generalize properly to out-of-sample data.  We have safeguarded against this possibility by adding OptiTrain, a feature which lets you save the best nets generated in a training series to improve your rate of successful network development.

HIGHLY COMPATIBLE.  N-Train reads files prepared with almost any spreadsheet or database package, text editor, or programming language, and on everything from mainframes to PCs.  If used with TradeNet (see below), the nets you develop with N-Train can be run real-time in TradeStation.  In addition, various N-Train processes may be "spawned" as tasks from within other programs and the system can even run as a time-sliced background task in a DOS box under Windows.  What does this mean?  N-Train gives you maximal flexibility in the design of your neural network systems.  It should be noted that N-Train is a stand-alone system, in other words neither Excel nor TradeStation is required to use it. N-Train can read and write standard text files, and is compatible with all software capable of writing and reading such files (Fortran, C++, Basic, Java, SPSS, Matlab, you name it).

CAN EASILY EMBED NETS IN ANY APPLICATION.  C and C++ programmers (or those who have access to them) will be pleased to know that N-Train is bundled (at no additional cost) with a C Run-Time Library.  This library allows you to easily embed your neural nets in any application developed with C or C++.  For example, you can interconnect networks using pointers to "brain" structures (like file pointers using fopen) to define and access a multiplicity of nets within a single program.

POWERFUL COMMAND LANGUAGE.  Our clients have repeatedly told us that N-Train has exactly the right kind of interface for the job.  Its very easy-to-use scripting language makes developing nets efficient and painless.  You can interact with the system quickly and easily (often with one keystroke), as well as only attend to what you regard as important.  It also facilitates the integration of the system (and its resultant nets) with other software, lets you perform certain kinds of complex procedures that would be difficult, if not impossible, to otherwise accomplish (e.g., walk-forward testing).  As you can see for yourself, only a few simple English words are needed:  SCALE causes the automatic construction of scaling information that is used to scale your files; GETFACTS instructs the system to actually scale and then load testing or training facts; SHUFFLE shuffles the training fact data to remove serial dependence; NEWNET allows you to create a new network; SETPARMS lets you set or alter any of the network's parameters; TRAIN initiates the training or testing procedure; SAVENET saves the trained network and scaling data to a file; LOADNET loads a file that contains a previously saved network; and RUNNET runs new facts through a previously trained network.  N-Train also permits scripting (the creation of a routine that executes a series of processes which perform complex functions, like walk-forward testing, and allows the advanced user to fully automate development protocols), monitoring (pertinent data is always on display), and provides an interrupt function (you can interrupt training at any time, then later return to the system to pick up from where you left off).

EXCELLENT MANUAL.  Our highly comprehensible, step-by-step instructions hold your hand through every phase of using N-Train.  Also included is information on the art of developing successful neural networks, and detailed examples of trained networks with information on how they were trained and even the source code used in the data preparation phase.  Included is a detailed example of how to easily use it from within Excel (includes well-commented Excel VBA macro code).

GENETIC TRAINING MODULE.  The basic N-Train system uses feed-forward, back-propagation to train neural networks.  This module adds four other training paradigms.

Genetic Training Component.  This component allows an entire population of nets to be trained using genetic algorithms.  In the back-propagation paradigm, the weights of a single neural net are modified.  In the genetic paradigm, the system breeds a population of neural networks and, in much the same way as in biological evolution, it "evolves" this population through the use of such processes as mutation, cross-over and selection, in order to produce the best possible population of nets for the problem.  The training process is vastly less susceptible to the problem of local optima and has all the other advantages associated with the use of genetic algorithms.

Partially Connected Neural Networks Component.  Like the basic N-Train system, this component uses back-propagation to train a feed-forward neural net; however, it also allows you to specify the connection scheme between the neurons.  This permits the development and training of neural nets in which only some neurons in each layer are connected to only some in the next layer.  Some problems are easier to solve using partially connected nets because the reduction in the number of connections (without the loss of model fit) results in the conservation of degrees of freedom, thus improving generalization.  Certain problems that cannot be solved using a fully connected network can be solved quite easily and well using a partially connected network.  The partial connection approach will work for a network with any number of layers.

Regression Component.  Regression is nothing more than a "linear" neural model, and a neural network is nothing more than a regression model that permits non-linear relationships between variables.  If the relationships between the variables can be reasonably approximated with a linear model, regression will provide a solution that is almost as good as the solution provided by a neural model and will do so much more quickly and with fewer degrees of freedom (consequently, yielding improved generalization).  We suggest that before computing more elaborate neural models, a regression is run in order to obtain a baseline against which to compare the other models; the regression itself may provide a reasonable solution.  As well as a trained "net" (representing the multiple regression), statistical tables are also generated showing correlations, regression coefficients, means, standard deviations, and other relevant information for all of the variables in the model. With regression, it is also easy to examine the importance of each input variable to the model.

Probabilistic Neural Networks Component.  This component allows you to create a probabilistic neural net solution using two fact files: a training or "base" fact file used as the knowledge base from which probability estimates for the classifications are made; and a "running" fact file containing the facts for which the probabilities are to be calculated.  This is useful when you wish o estimate the probable classifications of facts into one of several categories, and can be used whether your data is naturally categorical or is continuous.  As the number of cases/facts get larger, the probabilistic neural net classification model approaches that of a theoretically optimal Bayesian statistical classifier.

Variables Assessment Component (RelCon).  This component uses a unique and powerful approach to determine how much each independent variable contributes to your model.  When search for good variables, the prime consideration is whether they contain information that will help the net generate predictions about the matter you are trying to forecast.  The component extracts a variable from the model and reassesses the correlation of the net's output with the target.  If the correlation declines substantially from when the variable was still included in the model, then that variable was probably very important and the information it contributes cannot be reconstructed from the other variables.  This component helps you remove useless variables that could overcomplexify your model and cause curve-fitting.

Values Relationship Component (ROX).  Uses the output file to calculate and display statistics (histograms) concerning the relationship between the net output(s) and the actual target(s).  You can examine different kinds of predictions:  how each independent variable individually relates o the target(s), how well any individual net's output (and different ranges of outputs) relate to the target, or how a group of nets collectively predict the target ("majority vote" technique).  It is especially useful for screening variables and testing network models.  Traders can obtain information on the profitability of taking a trade for each range of the network's output:  you would know how much the market would move on average and the standard deviation.  We find ourselves using this component constantly.

Variable Correlation Component.  Quickly computes a correlation matrix showing the Pearson Product-Moment Correlation between every input variable and every output or target variable, as well as other additional statistics (e.g., means and standard deviations).  It is useful in screening certain data characteristics.

Decile Statistics Component.  Computes "decile statistics," which show the relationship between a given neural net or regression output and it corresponding target.  The model or net output is examined and all the outputs are divided over all the facts and distributed into 10 "bins" (ranges), each of which contains approximately 10% of the total number of cases:  The first bin contains the lowest 10% of the model outputs, the next bin contains the next 10%, etc. until the last bin, which contains the top 10% of the model outputs.  The percentage and number of "hits" are then calculated for the facts in each bin and are then written to the screen.  This is especially useful for direct marketers who, e.g., are attempting to predict response to mailing campaigns.

Distribution Statistics Component.  This computes statistics for all the variables in a given fact file.  Specifically, it computes the means, standard deviations, and provides a visual representation of the frequency distribution for each variable.  It is a very good way to quickly screen data.  The information provided allows you to easily spot outliers, asymmetric distributions, and other data characteristics.

Automatic Scaler.  In the basic N-Train system, the scaling control file has to be written in a text editor or spreadsheet like Excel.  This component automatically generates a scaling control file that will tell the system which scaling method to use for each of the variables in the fact file, thus greatly enhancing the ease of development work.

Multiple Vote Utility.  This utility processes the independent variables from a set of facts through any number of previously trained networks that you specify and generates a file consisting of the output(s) from each of the nets and the original target(s) for each of the facts.  It allows you to construe each of the nets as a committee member that has its own opinion (output) as to the value of the target.  By examining the "opinions" of a group of nets and obtaining the results of their collective wisdom, the kind of error that may be made by one individual network can be minimized and, thus, a better model may be obtained.

Multi-file Handler.  This utility allows you to concatenate (integrate) several fact files into one larger fact file.  It is especially useful for developers whose tools (e.g., TradeStation, Excel) are limited in the number of facts that can be prepared in any one operation.  It allows you to get around the limitations of your tools and to develop systems using multiple individual contracts, rather than artificially constructing continuous contracts.

N-TRAIN ORDERING INFORMATION.  Formerly, N-Train and its modules sold for $946.  Now, the entire system is yours for only $299.  Those who intend to run N-Train under UNIX or LINUX should know that a version of N-Train is now available for doing just that.  This version includes source code which can be compiled to run under all of the previous operating systems for which N-Train was originally available, as well as for UNIX/LINUX; the cost is $399.  There are no shipping costs since we send the software and detailed manuals by email in zip format files. Purchase includes telephone support to get you up and running.

TRADENET^TM Module that links N-Train and TradeStation

TRADENET ORDERING INFORMATION.  Formerly, TradeNet sold for $159.  Now, it's yours for only $25.  There are no shipping costs since we send the software by email in a zip format file.

TS-EVOLVE^TM Genetic Module for TradeStation

TS-Evolve provides a more efficient alternative to the selection of good parameters than other approaches (e.g., trail-and-error or "intuitive") and tools (such as TradeStation's built-in optimizer, which does not contain genetic algorithms).  TS-Evolve uses genetic algorithms (or GAs) to search for the best parameters (and combinations thereof) from incredibly large sets of potential solutions.  A genetic algorithm solves a problem using the same processes as biological evolution:  It works by recombination and mutation of gene sequences, which may be sets of number series.  Recombination involves the process of cross-over, whereby fragments of different solutions are combined to form new solutions.  Mutation involves the introduction of random alterations to these fragments to provide additional variation in the sets of solutions being generated.  Finally, the process of selection weeds out the less-fit solutions so that only those solutions that are the most fit can recombine and mutate, yielding another generation that may contain better solutions than the previous one.  This process of recombination, random mutation, and selection has been shown to be very effective in solving complex problems that contain many elements that can be arranged in many ways or parameters with values that must be determined.

We also use TS-Evolve, rather than TradeStation's built-in optimizer, for another important reason: speed.  For example, as discussed in the above-mentioned article, when juggling six parameters (two rules, each having two lookbacks and one threshold) to systematically find all possible solutions, a brute-force optimizer like TradeStation's is very inefficient.  If each of the six parameters had to be stepped through 100 possible values, there would be a quadrillion combinations to try; such an endeavor would keep TradeStation very busy for years!  In contrast, TS-Evolve is many magnitudes faster and the time for it to obtain a solution is measurable in hours instead of years.

TS-Evolve works with all 32-bit versions of TradeStation (2000i, version 6, etc.) and provides seamless genetic algorithm capability.  It consists of a DLL and a computational server that provides "global variables," as well as GAs.  Includes a detailed, well-commented example and free technical support to help get you up and running.

TS-EVOLVE ORDERING INFORMATION.  Formerly, TS-EVOLVE was sold by Ruggiero Associates for $495.  Now, it's yours for only $159.  There are no shipping costs since we send the software by email in a zip format file.

Stock Options Charting and Pricing Spreadsheet

SPREADSHEET ORDERING INFORMATION.  The cost of this invaluable tool is only $49.  There are no shipping costs since we send the software by email in a zip format file.

Simple Structure Factor Rotation in Fortran

FACTOR ANALYSIS PACKAGE ORDERING INFORMATION.  This highly useful research tool is now being offered free of charge under the Lower GPL.  It can be downloaded (as a zip file) using anonymous ftp: Download Now