Competing Against Luck: The Story of Innovation and Customer Choice

by Clayton M. Christensen

the spring of 2014 Science magazine published the findings of an academic study of the popular Google Flu Trends11—Google’s flu-tracking service, which is supposed to predict flu trends ahead of the traditional Centers for Disease Control and Prevention reports. Google Flu Trends (GFT) was based on an algorithm that matched fifty million search terms against 1,152 data points. In essence, Google hoped to predict flu outbreaks by cross-referencing search terms (symptoms, medical providers, remedies) with relevant objective data. The authors, academics from Northeastern University, Harvard University, and the University of Houston, concluded that Google Flu Trends had wildly overestimated the number of flu cases in the United States for more than two years. The article, “The Parable of Google Flu: Traps in Big Data Analysis,” concluded that the errors were, at least in part, due to the decisions made by GFT engineers about what to include in their models—mistakes the academics dubbed “algorithmic dynamics” and “big data hubris.” Google had admirable goals: maybe early warning on flu trends could prevent the spread of disease and save lives earlier than conventional methods. But as Google engineers found, you have to make choices about what to analyze. Unfortunately, the exact link between specific search terms and Google’s algorithm was far too complicated and subject to myriad human dynamics (maybe a hypochondriac searched the same terms month after month or perhaps Google ...

Google flu terms data anamoly