Hi,
I am creating a neural network in which the output (a stock return or rank) is dependent in a non-linear way on the inputs. In general, the idea is to output a rank function where the highest ranking stocks are those with the highest return. (I don't care about forecasting the specific return of each stock as much as getting the best stocks).
Here is my issue. From what I have read on neural nets, neurons typically sum the weighted inputs and push it through an activation function. However, I believe the best stocks are going to be the ones where multiple criteria are present. Addition doesn't accomplish this for me.
Here is a numerical example. Say I have three inputs and they each range from 0 to 1. It may turn out that I must have all three present to a significant extent in order to have a great stock. Thus, ideally instead of a summation at the neuron I would like to say (and(X1>.9, X2>.9, X3>.9)) as the output, or something like it. What I originally thought it to have a generalized function of the inputs that is the PRODUCT of the inputs: X1 * X2 * X3. I could create create a layer of linear neurons prior to this multiplication operation that pre-process the data.
Is there a way to do this in matlab? Perhaps what I need to do is stick a sort of "midstream" processing of some neural outputs before they go into other neurons. There doesn't have to be any optimization for the multiplication process, I think.
Please let me know if this is possible - thank you so much.
Best, Mike
