Predictions

Hello all,
I was handed a new project and I'm not quite sure how to
do it or even what technology to use. Now, I think that
SQL Server's Analysis Services could help (or even do the
entire thing) but I'm not sure. So, basically, I'm
looking for recommendations and ressources on the subject.
Here's my scenario:
We have a Data-Warehouse that stores information about
the production of our company. Around this fact table, we
have multiple dimensions including one about the origin
of the production (who) and a time dimension (when).
We are building cubes and linking Excel to them for
future analysis. Everything is done. Now, the all-mighty
supervisors would like to have some predictions built-in
to the system for the rest of the year based on the
previous years. Basically, we would be adding production
values for every possible time (the step is every hour,
so the sum of the production comming from a certain
origin is set as the fact and linked to a time dimension
row for that hour).
As you can see, it's not very difficult and it seems to
me like Analysis Services might already have that
functionnality although I couldn't find it (seems like
predictions can only be made by looking at relations
found by the data mining algorithms and these predictions
need to be made by a human because they are not really
registered as "new" facts or rows but more like new
columns).
So, in short, I'd like to have a tool calculate future
productions for the rest of the year and set a new column
(or dimension) to a certain value indicating that this
fact row has been predicted and is not certain. Of
course, the refreshing of data would not create duplicate
rows for predictions, newer predictions (probably based
on more facts) would replace the last equivalent
prediction.
I hope that's clear.
What do you recommend I use? Is it a case where I will
need to build my own separate program (that updates the
data-warehouse) or is it possible to achieve this using
data mining?
Thanks a lot,
Skip.You could use the datamining prediction task in DTS. I have used this before
and it seems to work well. The accuracy of your model (and suitability of y
our data) can be tested directly by applying the prediction task to the sour
ce data. This task as far a
s I used it is suitable for Y/N type questions - eg given a male living in L
ondon in a certain wage-bracket, would they be likely to join our gym. I can
't see how this type of prediction tool can be used for numerical data apart
from bracketed ranges. If
this IS suitable for you, then it's a simple matter to take the predicted da
ta and load it into the fact table and then do an incremental update of the
cube - eg in DTS.
Regards,
Paul Ibison