Once the human is satisfied that there is a high probability of the distinct movement following certain patterns, he or she can then watch for those patterns in the future. When the patterns appear, the human expects one of the distinct movements to follow and places orders appropriately.

The example above is the inspiration behind the enhancements to Adaptive Net Indicators (ANI) release 2.0. ANI always did pattern matching, but now in release 2.0 new functions not only include the current bar in the inputs but a number of lags of the input as well.

Technical Details

Adaptive Net Indicators are special versions of GRNN and PNN neural nets formulated to do pattern matching, both predicting and classifying. They retrain (quickly) on every new bar, so they are never more than 1 bar behind. You can set the contribution factors yourself so that the net uses your specification of how important the inputs are, instead of the other way around. Of course, you could also let the genetic algorithm find them if you own the NeuroShell Trader Professional or the DayTrader Professional. As a matter of fact, you can also let the GA find the optimal number of bars ahead to predict and the optimal training set size!! If that's not enough exclamation points, how about this: you can even let the GA find the best thing to predict by optimizing the parameters (including whether open, high, low, or close) of your output indicator!!! Adaptive Net Indicators is a package with unprecedented flexibility and capabilities!!!!

Many of you have expressed a desire to have confidence factors for your nets. The classification series of Adaptive Net Indicators will provide you with confidence. In addition, Adaptive Net Indicators will make no prediction at all if they feel they have no basis on which to do so.

Our Adaptive Net Indicators do pattern matching by comparing each new pattern encountered with a number of immediately previous known patterns. They do not use weights like most neural nets. The Net "output" (i.e., the output of the indicator) is derived from the outputs of the immediately previous patterns. It is most heavily influenced by the most closely matching of the known patterns, and so the output of any new pattern encountered will be much like similar known patterns. You the user can set the number of immediately previous patterns which the Net compares to the new pattern.

The Net indicator has "inputs" in which the pattern is stored, just like other indicators and neural nets. If a Net has 3 inputs, you can feed in the current values of the RSI, a CCI, and the Momentum to form the pattern, for example. Or you could feed the Net today's close, yesterday's close (lag 1 of close), and the close the day before (lag 2 of close). You can feed Net outputs into a Net, just as you could do with any other indicator. With some of our indicators, you can also input many "lags" of the primary inputs as well.

Nets also have another type of input, called the "Actual" value. This is where you show the Net what you want the outputs to be like. In other words, you "train" the Nets to produce values like the Actual value whenever the corresponding inputs are closely matched. You want the Net to predict for you the Actual value X bars in advance. The output of the Net is the prediction signal of the Actual value X bars ahead. You get to pick what X is for each Net. This is just like our neural nets in the NeuroShell Trader. Furthermore, you can even optimize the value of X in a Trading Strategy.

There is another big difference between our Adaptive Nets and the neural nets in the NeuroShell Trader. The contribution factors for each input are also inputs. That's right, instead of the net telling you the contribution of each input variable, you get to tell the Net what the contribution should be. The higher the contribution, the more heavily the Net will weigh that input when it does pattern matching. Of course, if you'd rather have the contribution factors figured out for you, the NeuroShell Trader Professional can optimize them.

There are two kinds of Adaptive Nets, depending on the type of output they produce. There are "Prediction Nets" whose outputs are predicted values (like price change, percent change in price, predicted indicator values, etc.). There are also "Classifier Nets" whose output is a probability of the pattern being of one type or another. Types might be "Buy" and "Hold", for example. Other types can be "Good" and "Bad", or "Up" and "Down", etc.

The Classifier Nets don't actually read or produce the strings like "Buy" and "Hold". You use positive numbers in the Actual for one category like "Buy", and zero or negative numbers for the other category like "Sell". The predicted output will be a number between -1 (strong probability of sell) and 1 (strong probability of buy). Numbers close to zero could be considered "Hold" (YOU would decide how close to zero a prediction should be to be considered a "Hold")

Example

One of the interesting things you can do with Adaptive Net Indicators is build adaptive moving averages. You can adjust how tight or how loose the adaptive moving average is. The chart below shows an adaptive net indicator configured as an adaptive moving average.

Configurations

There are a total of 18 Adaptive Nets, nine Predictor Nets and nine Classifier Nets. There are 9 of each because each of the nine takes a different number of inputs as follows:

Predict2 - Prediction Net which takes 2 inputs

Predict3 - Prediction Net which takes 3 inputs

Predict4 - Prediction Net which takes 4 inputs

Predict5 - Prediction Net which takes 5 inputs

Predict6 - Prediction Net which takes 6 inputs

LagPredict1 - Prediction Net which takes many inputs: 1 primary input and any number of "lags" of that primary input.

LagPredict2 - Prediction Net which takes many inputs: 2 primary inputs and any number of "lags" of those primary inputs.

LagPredict3 - Prediction Net which takes many inputs: 3 primary inputs and any number of "lags" of those primary inputs.

LagPredict4 - Prediction Net which takes many inputs: 4 primary inputs and any number of "lags" of those primary inputs.

Classify2 - Classifier Net which takes 2 inputs

Classify3 - Classifier Net which takes 3 inputs

Classify4 - Classifier Net which takes 4 inputs

Classify5 - Classifier Net which takes 5 inputs

Classify6 - Classifier Net which takes 6 inputs

LagClassify1 - Classifier Net which takes many inputs: 1 primary input and any number of "lags" of that primary input.

LagClassify2 - Classifier Net which takes many inputs: 2 primary inputs and any number of "lags" of those primary inputs.

LagClassify3 - Classifier Net which takes many inputs: 3 primary inputs and any number of "lags" of those primary inputs.

LagClassify4 - Classifier Net which takes many inputs: 4 primary inputs and any number of "lags" of those primary inputs.

For more details, please view the product manual for this add-on