Can Braincel find the best set of inputs ?

Yes. Braincel has a BESTNET option that directs the program to repeatedly eliminate the least useful inputs until all is left are the most useful. Braincel knows that the only way to do this correctly is to have two data sets, one for training the net and one for testing.

You can also specify a minimum number of inputs that the neural net must have.

Can Braincel find the best neural net size and shape ?

Yes. Braincel has a BESTNET option that the user can direct to find the best number of layers and number of processing modules (neurons) at each layer.

What's the best application of a neural net in a trading system ?

Numerous users of neural net technology have reported that using a neural net to predict the future TRADING RANGE of a price time series is likely to produce better results than most other kinds of forecasts, including future percent-price-change, future price, and future volatility.

To forecast future price range, you need to create TWO neural nets. One to forecast the HIGHEST price of the next N days (minutes, ticks) and one to forecast the LOWEST price of the next N days.

Does Braincel automatically normalize input data?

Yes. Braincel will normalize each input column separately. In general, normalization is a 2-step process that shifts and scales the data in each column into z-scores. That is, normalization evaluates the mean and standard deviation of all values in a column, then applies the following formula to each value X in the column:

Z=(X - mean) / Stdev

The normalization process in Braincel is a little more sophisticated than the above formula, producing similar results that are less influenced by outliers.

Can Braincel operate on data in real-time?

Yes. Braincel can process data in real-time in Microsoft Excel and TradeStation.

How often do I have to stop to retrain my neural net?

A properly trained and tested neural net does not need to be retrained every day. You may want to consider retraining on a weekly or monthly basis. It all depends on how fast long term market behavior changes in a month.

What's the best way to increase a NN's accuracy?

The training session automatically sequences through data records in random order. As a consequence, different training sessions will likely produce NN models with different weightings. You can get stable and accurate results by creating 5 NN models (using the same data and training parameters for each model) and using the "Olympic Scoring" method: after running a data record through all 5 models, eliminate the highest and lowest result, then average the remaining three values.

What programming language is needed?

VBA in Excel or Easy Language in TradeStation. For each platform, sample code is provided to help get you started.

Which platforms allow creating NN models?

You can create NN models in Excel. TradeStation users would need to first create a Braincel NN in Excel, then use Easy Language to apply the NN to real-time or historical data.