Generally, Neural Indicators will be used in Trading Strategies to give buy and sell signals. They do not have to be trained to predict anything. The optimizer causes them to evolve to the point where they give the most profitable signals. Think of using a Neural Indicator in a Trading Strategy as you use other indicators in a Trading Strategy. You put one in as a long entry condition using the relational indicator (A>B), where A is the Neural Indicator, and B is zero. Then you put another one in as a long exit condition, with A<B, again where A is the neural indictors, and B is zero.  

The main thing users of traditional neural nets have to understand about Neural Indicators is that they aren't predicting anything, so there's no period ahead to predict. They are just giving "intelligent" signals based on what they know about the past. The net is only trained to give buy/sell signals.

Each of the two Neural Indicators will have inputs, which are usually some other regular indicator. You might, for example, like to use the RSI, CCI, and DMI indicators as inputs to each of the NI.

Next, you tell the optimizer to find the best combination of parameters for the RSI, CCI, and DMI, as well as the best weights for the networks.

Note for advanced users: actually, you can use A>B for both entry and exit conditions, because the optimizer can make sell signals for A>0 as easily as it can for A<0. It doesn’t really matter. If you want to, you can even let the optimizer set B to something other than zero, but it isn’t necessary. We prefer to fix B at zero, and let the optimizer adjust the output; that way the optimizer has one less parameter to find.