In this issue:
I. Watch for short sales
II. The “one hold out” feature in the AI Trilogy
III. Free Indicators Display Bull versus Bear Strength
IV. Make sure you continue to get the news
*******************************************************
I. Watch for short sales
Each week during the month of October we’re having one week Oktoberfest sales. If the sale is for NeuroShell Trader or addons, you’ll find it on www.neuroshell.com. If the sale is on business and scientific products, we’ll put it on www.wardsystems.com. Watch for them – you could save some money! There’s one now on NeuroShell.com
*******************************************************
II. The “one hold out” feature in the AI Trilogy
Here’s something many users of our AI Trilogy (or NeuroShell Predictor and Classifier) may not realize or understand. These products have a unique feature in the genetic method. For the rest of this article we will use the word “result” to mean either the output of the NeuroShell Predictor (a prediction) or the output of the NeuroShell Classifier (classification). We will use the word “actual answer” to describe what the result should be.
Let’s say your training set contains 100 training patterns. Almost all other modeling algorithms (including regression analysis) try to fit a line or curve through the entire 100 actual answers in the training set at the same time. Most algorithms make one or more passes at the data “learning” all 100, and after each pass they give you results for those 100 in the form of training statistics. The statistics compare the result the model produces after that pass with the actual answer. Of course the result is pretty good because the method saw the actual answers as it learned them.
After training when you ask for a result from input data that was included as one of those 100 in the training set, you are getting an “in-sample” result. In other words, the actual answer was already known to the algorithm during training, so it is not surprising that often the result produced from such inputs is pretty good – after all the algorithm “learned” or even “memorized” the actual answers because it saw the actual answers at each pass.
The genetic method does something quite different. It never tries to fit a curve through all 100 patterns at once. In 100 passes of the genetic method, a different pattern is “left out”, and the model is only built by looking at the other 99 actual answers. In other words, the actual answer isn’t “learned” or memorized at all. Instead the model tries to produce a result for the held out pattern by inspecting the actual answers for the other 99 patterns only! Consequently, the statistics comparing results to actual answers often won’t be as good as with the other methods.
So why would you want to use the genetic method if it doesn’t “learn” as well? Because how well it learns the training set is largely irrelevant compared to how well it produces results on new “out-of-sample” data that it has never seen before, wouldn’t you agree? And the fact of the matter is that it quite often (but not always) produces excellent results (in other words it generalizes better).
But here’s the best part. The genetic method is really appropriate when you don’t have many samples. If you are using one of the other modeling techniques, you are going to need another 50 or 100 patterns not from the training set to test the model “out-of-sample”. With the genetic method, you arguably don’t need these other 50 or 100 patterns, because the model is always “out-of-sample” for all practical purposes. Of course if you have more data, there’s no issue. As a matter of fact, 100 patterns is usually way too few for a training set with the other methods, but probably sufficient for the genetic method.
*******************************************************
III. Free Indicators Display Bull versus Bear Strength
Be sure to check out www.ward.net’s Traders Tips from Stocks and Commodities in the NeuroShell Trader section. The October 2003 tip describes how to build Vadim Gimelfarb’s Bull and Bear Balance Indicators in the NeuroShell Trader. Gimelfarb offers a new twist on indicators that graph the power shift between the bulls and the bears. His indicators take into account the opening and closing prices as well as the price dynamics of the previous day. Our site www.ward.net includes the tip from WSG as well as a chart with Gimelfarb’s indicators.
Unfortunately, we can’t give you the text of the original article because the magazine has a copyright on it.
*******************************************************
IV. Make sure you continue to get the news
We are still surprised that, several months after we sent out our June newsletter announcing release 4.0, users are still calling or emailing to ask when release 4.0 will be available. Seems they all changed their email addresses and never notified us of that. They could have been making money with the new features all of this time. There are many others, no doubt, who still don’t know about release 4.0.
After all these years we are also still getting tech support calls from NeuroShell 2 users who don’t know about the AI Trilogy and its components. So they’re stuck trying to impress their bosses with older technology.
Don’t be left behind – tell us when you change email addresses.
*******************************************************