Contrary to rumor, unsupervised neural nets are not neural nets on spring break.  However, in spite of the wild sounding name they can improve your trading systems.  Unlike the supervised network type in the NeuroShell Trader prediction wizard, unsupervised nets do not predict anything; there’s no period ahead to predict.  Neural Indicators are just giving “intelligent” signals based on what they know about the past.  The net is only trained to give buy/sell signals.

This chart for USG displays the trading signals and equity curve for the out-of-sample period.

We built this chart with one of the net types you can find in our Neural Indicators (NI) add-on. NI provide signals from -1 to 1 which may be interpreted as probabilities for buy/sell decisions. NI learn how to give their signals based on evolutionary pressure. The genetic algorithm in the NeuroShell Trader “evolves” NI that gives better and better signals. Survival of the fittest controls the evolutionary process, where fitness is determined by how much money the Neural Indicators make.

In this example for USG Corporation, the NI version produced a 202.1% return compared to 184.8% for buy and hold during the training period. In the out-of-sample period NI returned 21.4% compared to 13.8% for buy and hold.

The Indicator Wizard lists the parameters you can control in a Ward net with 4 inputs.

The Trading Strategy used in this model used the same Ward4C indicator for Long Entry, Long Exit, Short Entry, and Short Exit conditions.  However, the indicator was allowed to individually evolve for each part of the trading rule.

Scale refers to the number of previous patterns or bars included in the model that are used to normalize the input values.  The scaling is recomputed as each new bar is added to the chart, so the nets keep up with current market conditions.  H1 and H2 refer to the activation functions used each of the two hidden neurons in the model.   You can choose either the hyperbolic tangent or a Gaussian activation function, or let the optimizer choose for you.   Next, the inputs are listed.  In this case we simply used Close, Open, High, and Low.  The following parameters that begin with “w” are the network weights and are best set by the optimizer.  So even if you aren’t a neural network expert, you can effectively use the Neural Indicators add-on.

The “C” in the Ward4C indicator name means that this is a conditional version of a Ward4 indicator.  The C versions of the indicators produce a value of either true or false, therefore you don’t have to insert the Ward4 indicator in a relational indicator such as A > B in order to create a trading rule.

There are several different types of Neural Indicators, including Ward Nets (a type of backpropagation), Jump Nets (which cross connections between network layers), Recurrent Nets (which include a portion of previous training patterns), and Sparse Nets (which are loosely connected so they generalize better).