SAS Global Forum 2009 Recap: Tuesday

I know this post is late, seeing how the conference ended over two weeks ago, but I figured better late than never. My impressions of the presentations are not as fresh, but I have more time now to write a bigger post about the sessions I attended.

The first session I attended was on time series predictions using neural networks, from Goel. This session was interesting and informative if for no other reason than its novel approach. The author used neural networks and ARIMA models side by side to predict time series data. He used this approach across several different types of data sets. In the end, the neural networks were able to achieve a reduction of error of around 80% to 90%! If these results can be replicated, I would be interested in using neural networks in my work to see if I can get that kind of accuracy.

The second session I attended was on regression assumptions, from Cerrito. She gave a lot of helpful pointers about some common-sense tests you should run on your data before modeling, and some common-sense tips about interpreting your output. Some good takeaways? If n gets too large, the variance will go to 0 regardless of whether or not the model is correct; if you have a large n, take a random sample of your data and fit a model based on that. When you're predicting rare events, your model can have high accuracy but horrible predictive power. For instance, if a disease occurs in 1% of the population, a model that says no one ever gets the disease will be right 99% of the time. Great accuracy, but horrible predictive power!

The next session I went to was on customer retention, from Pruitt. This presentation was not too compelling. The presenter essentially gave one big example about how he had developed customer retention scores using Enterprise Miner. We don't have Enterprise Miner, and he didn't go into any real depth on the theory side, so I left early to go work on some other projects. Several of my colleagues were there, so if I missed anything truly important I can find it out from them.

The next session, on Bayesian modeling using MCMC, was one I had already been to! I saw Fang present this paper before, I'm not sure where, but I had definitely seen the paper before. I left early to go to lunch.

After lunch I went to a couple of short sessions. The first was from Knafl about a macro for adaptive regression modeling. The presentation was about a macro the author had written that used k-fold likelihood cross-validation to determine the best model from a class of models. This macro could be interesting; I will probably check it out from the author's website. The second was from Chou and Steenhard about count data regressions, which are useful to me since most of my data are count data. They just presented a macro they had written that deals with 17 different distributions for count data, instead of SAS's built-in Poisson and negative binomial distributions.

At this point I realized there were no more presentations that looked interesting to me, so I staked out a chair in the convention center and did some work for the next few hours.

Tuesday night a few colleagues and I went out to a wonderful restaurant in Alexandria's Old Town called Las Tapas. The food was good and there was live Flamenco dancing, with one dancer and one guitar player. It had a fun, intimate feeling. If you're ever in the area, I highly recommend it.