At SAS, we use terms like “machine learning,” “predictive modeling” and, of course, “analytics” quite a bit in our day-to-day business. Not surprising, given that we're the largest analytical software vendor out there. But have you noticed that these terms are popping up more frequently in news articles and blogs?
I have. Maybe because that’s the kind of stuff I end up reading, but I’m not so sure. I feel like these terms are popping up more and more; not only that, but when these terms are mentioned, it’s quite often in passing, or in a matter-of-fact way.
I like this. I like the fact that people talk about analytics now, and not only that, they're talking about sophisticated analytics. That’s great. Really great. It’s talked about in a way that makes it sounds easy and achievable. Which it is. At SAS, we've done a lot to advance the use of analytics and to make the development of analytical assets easier and easier. We've even made the deployment of those assets as easy as possible.
But, and you knew this was coming, there's a problem. It’s still not easy to get the kind of analytics that the trade magazines and we at SAS talk about into a production process. Yes, we as a software company have done pretty much everything we can do to make it easy, but there's still a problem.
There, I said it. Sorry to be so blunt, but it needed to be said.
It’s ok, it’s not your fault. In fact, most people have the same huge and costly problem. Once an analytical asset has been created, it will sit on a shelf while IT teams work hard at what are often manual tasks, aligning the stars so that the new analytical asset can be dropped into a production system. After which it can finally start churning out its analytical magic.
This all takes a lot of time – often months. And those months of your analytical asset sitting on the shelf quickly add up to substantial dollar amounts. You see, the moment an analytical asset is created it's new, shiny and the best it's going to be. Analytical assets are built using historical data. Yes, yes, I know you know that. But what we often forget is that as time ticks by, that data gets older.
Which means that the day we finish our model it's the closest it will be to our historical data, and therefore the best it can be. From then on, as each day that passes, our model deteriorates – it becomes stale.
Why is deploying models such a manual process? Sadly, it’s because that’s the way it's always been done. It’s time for a change.
SAS has software that can help, and we have a best practice framework for you too. In my next blog post I’ll talk about one of the reasons companies like yours find it hard to put models into production.
In the meantime, for more about deployment challenges and how to fix them, take a look at this white paper: Breaking Down the Barriers Between Model Development and Deployment.
Fav point (sadly true, Adam):
"Which means that the day we finish our model it's the closest it will be to our historical data, and therefore the best it can be. From then on, as each day that passes, our model deteriorates – it becomes stale."
Nice post, Adam.
Thanks for the comment Lonnie!