Discussion of article "Neural Networks: From Theory to Practice" - page 2

 
Thank you, this is really a great introduction to NN.
 

The way of optimizing the network is "SMART", even when he's not using the regular "BACKPROPAGATION" algorithm.

Backpropagation adjust the weights automatically in order to reduce the output "ERROR" (or loss) to a truth value.

What is SMART about this Feed Forward Network application ?

- Instead of solving the Optimization problem of "Backpropagation", the optimization problem solved is likely to be named as "Feature Fitting"

or "Feature Extraction Problem". Following this, the solution obtained is... "Given the Trading Rules which are the Features that Perform Better, or maximize the trading output"

in this sense the problem is more about "The proper way to represent the trading situation instead of fitting a function (the regular use of Deep Learning)".

What are Features in Deep Learning ?

- The interesting thing in Deep Learning is that each network layer works as an Adjustable Linear Transformation, this allows to change the space in which the data is projected,

and that creates a new representation (new features).

 

only opens 1 trade in backtest



 
thank you for your articles.
 

An important question:

The RSI N past values is set to 14(N=14).

The number of inputs is 10 (past values).

Is there a problem? Seems that the ML results might be unstable?

Please answer....


Thanks