Helder Borges on Thu, 18 Apr 2013 09:40:13

Hi guys.

I´m testing what algorithm have better performance at predicting sales from generic businesses.

After i settle the NN algorithm, i´m doing prediction with the following DMX query:

  Predict([ATLANTIDA].[Total Value Mes]),
  PredictProbability([ATLANTIDANULLLR].[Total Value Mes])
  OPENQUERY([Business Analytics DW],
      ... Variables ...
      (SELECT        ... variables ...
FROM            vwSalesForecastTS AS t
WHERE        (CompanyNK = ''ATLANTIDA'')) [ATLANTIDAL] Order By Date ASC
    ') AS t
 ... variables ...

Being "... variables ..." all variables i need and TotalValueMes the total sales of the month.

The predictions are working fine.But i´m only able to make prediction for months wish i already know at least the external values.

What do i need to do to predict future values where i don't know any kind of variable? 

Best Regard,
Helder Borges



MarkTab on Thu, 16 May 2013 18:54:28

Hi Helder

Use time series/forecasting instead of neural networks.  There is a way to reformulate the problem using neural networks, but that approach is comparatively harder.  Time series is hard enough. 

Use time series with caution:  I recommend "Practical Time Series Forecasting" by Galit Shmueli.

Helder Borges on Thu, 16 May 2013 20:02:09

Hi Mark.
I Already solved the problem.
I applied both Time Series and Neural Networks.
Now, i'm looking for an SVM approach, but unfortunately Analysis Service don't support SVM (i cant use the codplex plugin because of software versions :))
Thanks for the tip anyway :). And i will take a look at the book. :)                                                         Best Regards                                                                                                                                         Helder Borges