Visit neurosolutions.com | Download Evaluation Volume 7, No.4 This is a Neurosolutions Newsletter, which you are receiving because you requested to stay informed about new developments at NeuroDimension. You can also access current and past issues of Neurosolutions Newsletters online. Using the L2TemporalCriterion for Time-Series Prediction Problems Back in the early stages of development of our TradingSolutions product, we enhanced NeuroSolutions by adding a new Error Criteria component called the L2TemporalCriterion (L2T). This component allows you to give more weight to certain types of errors which pertain to time-series prediction problems such as financial forecasting. Building a Simple Neural Network Before going into the details of the options available with this component, let's cover the steps involved in swapping out the default L2Criterion with the L2T: Open your existing temporal network, or create a new "Prediction" network with the NeuralExpert or a new "Time-Lag Recurrent" or "Recurrent" network with the NeuralBuilder. If you do not have the ErrorCriteria toolbar visible, open the Customize dialog (from the Tools menu) and check the "ErrorCriteria Components" checkbox (click OK to close). Left-click on the "L2 Temp" toolbar button so that the cursor changes to a stamp icon. Hold the cursor over the L2 component (the rightmost Axon) and left-click. The L2 component should now be replaced with an L2T component. Right-click on the L2T icon and select "Properties" to display the inspector for the component. Below is a description of each option, taken from the NeuroSolutions Help file: Recency of Observation This criterion weights the error of the exemplars at the end of the data series more heavily than those at the beginning of the data series. This is useful when predicting time series in which the conditions may be changing over time, such as predictions based on long-term historical data. The amount of weighting is determined by the Discount Rate. Direction of Change This criterion weights the error of the exemplars whose output is the opposite sign of the desired output more heavily than those whose signs match. This is useful for applications such as trading, when being on the right side of the trade is the most important. There is one weighting specified for exemplars with outputs in the Right Direction and another weighting for exemplars with outputs in the Wrong Direction. Magnitude of Change This criterion weights the error of the exemplars whose output is far from the desired output more heavily than those with an output and desired output that are close. This is useful for learning infrequent data that has a desired value that is not the same as most of the desired values. However, this may result in less accuracy overall. There is one weighting specified for exemplars with outputs with a Large Change from the desired output and another weighting for exemplars with outputs with a Small Change from the desired output. Data Pre-Differenced Check this box if the output data is a measurement of the difference (or percentage difference) between the current exemplar and the previous exemplar. Discount Rate This parameter is used when the Recency of Observation box is checked. The higher this value, the more weight is given to the errors produced by the recent data (the data at the end of the series). Right Direction This parameter is used when the Direction of Change box is checked. The higher this value, the more weight is given to the errors produced by the output having the same sign as the desired output. Wrong Direction This parameter is used when the Direction of Change box is checked. The higher this value, the more weight is given to the errors produced by the output having the opposite sign as the desired output. Large Change This parameter is used when the Magnitude of Change box is checked. The higher this value, the more weight is given to the errors produced by exemplars in which the difference between the output and the desired output is large. Small Change This parameter is used when the Magnitude of Change box is checked. The higher this value, the more weight is given to the errors produced by exemplars in which the difference between the output and the desired output is small.