We are frequently asked if our software can provide confidence factors for our neural network outputs. In fact, the genetic method in the NeuroShell Classifier DOES provide confidence factors. . So do the classification Adaptive Net Indicators available for the NeuroShell Trader. They are the probabilities which are the outputs of these nets. They are in fact probabilities that the pattern under consideration is in the respective categories.

If you have a continuous output value like price, temperature, concentration, etc., you can probably formulate your problem as a classification problem. In the financial world, you can use buy, sell, and hold categories. For temperature or concentration, you may be able to make ranges that are of interest. These become categories. For example, the category of LOW is temperature 0 to 40 degrees centigrade. The category MEDIUM is 40 to 80, and HIGH is above 80. Then you can use the NeuroShell Classifier.

If you own the NeuroShell Trader, the Professional version can call the NeuroShell Classifier nets if you also own the NeuroShell Runtime Server. Or you can just use Adaptive Net Indicators addon with the Trader.

But there is a danger in relying upon confidence factors. DON’T BE FOOLED BY CONFIDENCE FACTORS PROVIDED BY US OR ANYONE ELSE. You may be tempted to make false conclusions from these confidence factors. Those people predicting financial markets are particularly susceptible to making these false conclusions. For example, if you receive a probability of 90% that the current market is in the BUY category, you will want to take that literally, and call your broker right away. Unfortunately, the correct way to look at the output is that the probability is 90% that the current market matches the patterns with which you trained the BUY category. If you did a poor job of building your net (poor inputs, unrepresentative training patterns, overfitting, …) or even if the market has just changed, then the confidence factor cannot really be used the way you want it to be used.

The paragraph above just means that confidence factors, especially in non-linear models, are probably not as useful as testing your nets on lots of out-of-sample patterns, and deriving your confidence from that testing.

Update 4/24/01: Now, in addition to the Adaptive Net Indicators addon, the following addons for the NeuroShell Trader will also produce a confidence factor of the prediction:

Neural Indicators
Cluster Indicators
Fuzzy Pattern Recognizer