Greetings from NeuroDimension!

The World Leader in Neural Network Software

 

This issue of the NeuroDimension newsletter highlights the latest developments at NeuroDimension, as well as providing insights on how and when to use neural networks.

 

In this issue you’ll find:

 

What’s New and News?

   *   TradingSolutions Real-Time Beta Available

 

Designing Neural Networks

   *   Avoiding the Potential Pitfalls of Correlation Analysis

 

Note: You are receiving this newsletter because you requested to stay informed concerning new developments at NeuroDimension. If you would like to stop receiving these newsletters, please see the bottom of this newsletter for removal instructions.

 

=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=

What’s New and News?

 

TradingSolutions Real-Time Beta Available

NeuroDimension will soon be releasing TradingSolutions Real-Time and TradingSolutions End-of-Day v2.1. To ensure quality standards, a beta release of both versions is being made available to current licensed users of TradingSolutions in advance of its full release.

 

TradingSolutions v2.1 will be available in both Real-Time and End-of-Day versions and will introduce many new features and enhancements, including: streaming data support, programmable alerts, eSignal data support, and asynchronous processing. TradingSolutions Real-Time will be priced at $1995. Upgrading from TradingSolutions v2.0 to TradingSolutions End-of-Day v2.1 will be free. Upgrading to TradingSolutions Real-Time will be $1000.

 

Use of the beta is free. Availability is currently limited to licensed TradingSolutions users. The beta version includes all of the features being added in v2.1, including those found only in the real time version. For more information about this beta, log into the Licensed User Center on the TradingSolutions web site or contact us at info@tradingsolutions.com

 

=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=

Designing Neural Networks

This Issue: Avoiding the Potential Pitfalls of Correlation Analysis

 

Neural networks are nonlinear models. As such, linear statistical methods such as correlation analysis may not provide a complete picture when analyzing input-output relationships.

 

Correlation Analysis

When trying to determine which inputs will be “good” for predicting an output, in linear methods you would use correlation analysis. The higher the correlation with the output, the better the input predicts the output. Specifically, if an input value moves in some direction, highly correlated output values will move by a proportional amount. Correlation analysis, however, only measures linear relationships.

 

Good Inputs With Low Correlation

Indicators and state-related variables are a good example of inputs that may be nonlinearly related to the output. They can take the form of “if X = true, do this, otherwise do that”, which is a non-linear relationship. This type of input can be very important to the solution and yet may have a very low correlation.

 

Another example is as follows: assume you have a function y = x1 * x2 where both x1 and x2 are variables. This is not a linear function. If x1 varies from -1 to 1 and x2 varies from 0 to 1, then x1 will have a reasonably high correlation with y (high x1 generally means high y, low x1 generally means low y), but x2 will have a very low correlation with y because x1 switches the sign of y independent of the value of x2. Thus, a large x2 sometimes means a large positive y and sometimes means a large negative y. A simple 1 hidden layer MLP can easily solve this nonlinear function, but a linear system could not – and in addition, the correlation analysis would be deceptive as to which variables are important to the solution.

 

Use Correlation Analysis Wisely

In neural networks, we recommend you use correlation analysis only as an initial guess as to which variables may be important. Values with high correlation to the output have a good possibility of being useful inputs. However, values with low correlation may still be very important components in nonlinear relationships.

 

Sensitivity analysis is a preferred way of detecting the significance of non-linear inputs. For more information about sensitivity analysis, see the previous newsletters. A list of previous newsletters by topic is available at: http://www.neurosolutions.com/newsletters_topic.html

 

=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=~=

Comments or Suggestions?

 

We appreciate your feedback! Please send us your comments or suggestions concerning this newsletter, our web site, or part of the NeuroDimension product line. Write to us at: feedback@nd.com

 

Have questions about NeuroDimension products or training services? Send your questions to: info@nd.com

 

Some mail readers may display an attachment at the bottom of this newsletter. This is typically caused when the mail reader is set to display text only. The original, formatted version is sometimes included as an attachment.

 

This issue and previous issues of this newsletter are available on the NeuroDimension web site at: http://www.neurosolutions.com/newsletters.html

 

If you would prefer not to receive these newsletters or subsequent product updates from NeuroDimension, please reply to this letter with the subject heading changed to the word REMOVE.

 

Thank you again for your support of NeuroDimension products!