Training Model

Before the first use of your model in predicting a stock, you have to generate the initial training. Click the LEARN button or select START OR CONTINUE. Then a learning mode wizard will appear, which will help you to set up the learning parameters.

Forecasting technology.  

This parameter sets up how intensive your model will be learned.
The most important parameters which are changed by selecting different forecasting mode, is the number of iterations, and error limit.

The number of iterations sets up how many iterations will be processed in learning, in other words, how many times the data sets will be pushed through the neural network, and, how stronger they will effect on nn structure memory.

Error limit parameter sets up the level when nn structure is considered to be learned already.

When using long historical periods (longer than 0.5 years), and large NN structures, it is recommended to use ‘strong’ and ‘aggressive’ learning mode, which is Associative forecasting mode.

In that mode, the number of iterations is 15.000, and error limit is smallest. This way, the learned model will use associations for forecasting, and based on current input data set, it will try to associate it with all previous data sets which were used in learning process. If there’s no association-the model will not give any output result.

From memory forecasting mode-has 5000 iterations, and middle error limit. The model will try to recognize current input data set, by comparing it with all previously seen, which were used in learning process, and, by generalization of difference between this unknown data set and in-sample data sets.

Intuitive forecasting mode-the number of iterations is 3000, and error limit is in high position. The effect of learning in that mode is not so aggressive for model, and, it learns to give intuitive forecasting on unknown data. The memory effect is not strong, and there’s almost no defined dependencies between in-sample data sets. So, nn structure gives some kind of ‘suggestions’ based on unknown input data set, it tries to classify data set using degraded classification clusters.

So, when using large periods, and complex models, associative forecasting is recommended. For middle range of structures (standard nn structure) you may wish to try ‘from memory’ mode.

For small nn structures, and short historical periods (half a year of historical quotes, for example, and, small/normal NN structure), it may be useful to select ‘intuitive’ forecasting technology, as the number of data sets is not too large, and NN structure is small and can be set up to stable working state without thousands of iterations, which will put NN structure into overlearned state, which means that the NN model will not be able to give any adequate forecasting/results.

Initial Training/Aftertraining option.

The Training process is automatic.

If you start learning for the created model, select FORGET ALL AND START TRAINING FROM THE BEGINNING.

If you wish to aftertrain the model with new updated historical data, select AFTERTRAIN.

If you already have a trained model and just wish to retrain it without any ‘forgetting’ of the previous training, choose ‘Aftertraining’.

If you wish to train the newly created model, choose the training method with forgetting and randomizing the system.