Ever wanted someone to walk you through getting started in TradingSolutions? Well here is your chance! TradingSolutions Live is an introductory online course in which a standard web-browser is used to view the instructor’s computer screen.

More information and to sign up for a course, visit: http://www.tradingsolutions.com/resources/tslive.html

In the previous TradingSolutions newsletter, we discussed data normalization and how TradingSolutions automatically takes care of normalizing your data. Does this mean you don’t have to worry about it? Yes and no. You don’t need to do anything to make data normalization happen. However, having an understanding of what is going on behind the scenes can help with effective selection of inputs to your predictions.

Why do neural networks normalize?
In the previous newsletter, we discussed how non-linear models are different from linear models. The inputs to non-linear models can have different effects on the output depending on the value of other inputs. In other words, a one percent change in one input doesn’t always have the same effect on the output.

To model this relationship, neural networks use non-linear functions, like the hyperbolic tangent (tanh) function, which performs best with values in a certain range, such as between -1 and 1. Therefore, to use these functions, the inputs to them are normalized (converted) to values in this range. The amount by which the inputs are normalized is determined by the values available during the training phase.

The Effects of Saturation
Since these values are determined during the training phase, it is important for the training data to include values from the entire range of potential values for each input. Any new values that are significantly larger or smaller than the training data will have the potential of saturating the non-linearities, causing these values to all look the same to the neural network. This is similar to the human eye generalizing areas with little or no light as “darkness”.

While this might seem like a flaw, it is actually a positive benefit when creating a prediction. In the case of unusually high or low values, the neural network is essentially dealing with input cases it has never seen before. Saturation enables the neural network to generalize these values into categories of “very high” or “very low” and process them based on the values it encountered during training.

Using Prices as Prediction Inputs
The effects of normalization are the reason you would typically want to use the percent change in the price of a stock as an input to a neural network rather than the price itself. Stock prices do not have upper or lower bounds, so it is very possible that new prices may be above or below the values seen during training, causing the neural network to potentially view them all as having the same effect on the output. On the other hand, the percent change in the price will typically stay in a fixed range. In cases where saturation does occur, very large values only look like other very large values and very small values only look like other very small values.

The Solution Service is a free service that provides daily trading signals and commentary for a select group of stocks. These models can be used directly as part of your trading strategy or used as examples of how to generate your own winning signals. To access it, simply press the "Update Solution Service" button in the TradingSolutions toolbar.

Since the last newsletter in mid-April, the NASDAQ and the S&P 500 have gained about 1%. During this time, three Solution Service models have gained over 10%! Overall, the Solution Service models combined for a small loss of around half a percent due to some models encountering longer-than-normal trends. These momentary losses are typically recovered as soon as the individual stocks resume normal trading.

Thank you again for your support of NeuroDimension products!