Is The Connected World Killing Predictive Modeling?
from the the-speed-of-information-is-faster-than-your-algorithm dept
The example he uses is where Michelle Obama recently wore a J. Crew outfit, driving sales of that outfit through the roof. No predictive model could expect that. In some ways, this actually reminds me of another recent story as well. Comedian David Chappelle was apparently in Portland and told a couple people he was planning to show up at a certain city square to do an impromptu midnight outdoor show with a tiny amp and microphone he had just purchased. But in the course of about seven hours, the news spread rapidly via Twitter and Facebook. Chappelle had expected perhaps 200 people, and estimates in the end put the number at many thousands of people (the small amp apparently couldn't amplify his voice enough, and the large crowd became such a concern that he basically sent everyone home after a few attempts to get a bigger amp). While it didn't involve an algorithmic predictive model, this case involved Chappelle's human predictive model that over the course of a few hours you could maybe expect 200 people who would (a) find out about it (b) be nearby and (c) be willing and able to come out at midnight. But the ease of communication changed that equation and made it a lot more difficult to predict the outcome.
Now, of course, the issue with both of these are that they may be outliers. Most other clothing at J. Crew probably followed a typical predictive sales path. And a less famous comedian would probably struggle to get anywhere near 200 people to show up. So, I'd say I'm not at all convinced (as the article posits) that Twitter "confounds" predictive modeling. I think it still requires other random events to occur -- and those have always happened in the past. Perhaps the interconnected nature of the world today can massively amplify a sudden flash fad, but that doesn't necessarily mean you toss predictive modeling out the window. Considering it was such an odd claim to use a few random outliers to damn an entire useful tool, it seemed worth digging a bit deeper... and (surprise!) it turns out that the guy who wrote the article runs a company selling a tool that competes with predictive modeling software focusing on "behavioral analytics." So, his argument is basically a strawman against predictive modeling that uses some outlier data without any evidence of a real trend. I'm certainly interested in the ability of social media to amplify a fad, but I don't think it has a really serious impact on most predictive modeling.