Neural Indicators (NI) are indicators based upon neural networks. They require the NeuroShell Trader Professional or DayTrader Professional. Here are the salient features:

1. Signals. NI provide signals from -1 to 1. The general interpretation of these signals will be some sort of "binary" classification, like "buy" (>0) or "don't buy" (< or = 0). Usually, we simply insert them in a trading strategy as a rule (condition) to buy and sell.

2. Probabilities. NI give probabilities of the signals they produce, because if they are providing a sell signal, for example, the closer to 1 the signal is, the stronger the probability of sell. The closer to -1, the stronger the probability of "don’t' sell".

3. Unsupervised. NI are "unsupervised" neural networks, meaning that they do not need to be trained by showing them the correct answers, like most neural networks (i.e., "supervised" neural networks). You do not teach them by providing any kind of actual output which they learn to reproduce, as with most neural nets.

4. Evolutionary. NI "learn" how to give their signals based upon evolutionary pressure. The genetic algorithm (GA) in the Trader Professional or DayTrader Professional "evolves" NI that give better and better signals. Survival of the fittest controls the evolutionary process as usual, where fitness is determined by how much money the NI make, or how good they work as inputs to other nets or indicators.

5. Architectures. Although anyone can use NI, neural network aficionados will love them because there are several highly technical neural network "architectures" included from which you can choose. The following technical outlines are for neural network experts. If you’re a neural network novice, refer to Neural Network Architectures .

A. Ward Nets. This architecture has two different "activation functions" in the hidden neurons. These are called "Ward Nets" since Ward Systems Group invented them many years ago (they first appeared in our classic product NeuroShell 2). The genetic algorithm will find out how to pick the activation functions for you.

B.  Jump Nets. This architecture has connections directly from inputs to outputs as well as the usual hidden neuron connections. This architecture also features connections from one hidden neuron to the next, like Turboprop 2 has.

C. Recurrent Nets. This architecture analyzes not only the current bar of information to produce its signal, but it also reviews a condensed summary of the most recent bars as well. More recent bars receive more weighting than older ones.

D. Sparse Nets. These are nets which are not fully connected between the input and hidden neural layers. This means that more inputs can be fed to them without increasing the number of weights drastically. The fact that less information can be stored in sparse weighting connections is compensated for by the fact that less weights allow better optimization.

6. Generalization. These nets generalize very well, meaning they do not have a strong tendency to "overfit" or "curvefit" like backpropagation neural nets do.

7. Limitations.

a. In order to provide good generalization, and to keep low the number of weights the GA needs to find, NI will allow from two to six inputs, which can be any indicators.  Sparse nets will take up to 12 inputs, however.

b. The number of hidden neurons is always two for the same reasons (except with Sparse Nets, which can have 3).  We have not found these limits to be restricting; on the contrary, real neural net experts will recognize them as good neural net practice.

c. Of course you can use multiple NI together or feed them into one another to bypass the input and hidden neuron limitations, but is it not recommended because of the large number of weights which would need to be evolved. Generally, you will be using either 2 or 4 NI in a trading strategy anyway (one for buy signals, one for sell, one for short, and one for short covers).

d. Training time is not nearly as fast as with our usual Turboprop 2 neural nets, since the nets must be "evolved" by the GA. However, our results have shown that the increased time is well worth it.