Can Neural Nets Do This? I am trying to predict future sales from past transactional data with a neural network.  My data looks something like this:
Customer, Transaction_Num, Sales, Product, Date, ... About 50 more categorical variables

I would like the network to learn from these transactional rows and predict next week's sales.
The best way I can think of to structure this data is to sum up all data for each week and put it in a column, like so:
Week t Sales, Week t-1 Sales, Week t-2 Sales, etc for all 50+ variables * # of weeks

And for a lot of the variables it wouldn't make sense to sum or average them.  
Is there a better way of doing this?
 A: The data that you have available is temporal data. Because of that there are two choices you can make. 
Either you use a window for your temporal data, and compute statistics over this window, say the mean, std deviation, etc. You can then update your model by shifting the time-window to the end of the period after every week.
A more natural approach however would be to use a recurrent neural network, which is capable of processing arbitrary length time-sequences. At every time-step t you feed in the values of your features for that date, and make a prediction for the sales. 
You can then train the model by back-propagating through time the error between the prediction at time-step t and the actual sales at that time-step. For training I would use a fixed length time sequence of say T=20 (or as long as you deem that the variables will still have impact on future sales). You can obtain multiple training examples by sliding your window of T over your time series, and creating sub-time series of length T.
A: One way to do it is to perform a PCA and reduce the number of variables. 
I am not an expert in neural nets, but picking out around 5 - 10 variables from PCA that explained 80%+ variance can already reduce the computational time by miles.
