Prediction Markets at Google: A Guest Post
In my last post, I promised to say a bit more about prediction markets at Google. Google has been running internal prediction markets for a couple of years, and Eric Zitzewitz and I were fortunate enough to team up with Google whiz Bo Cowgill to analyze these markets.
Ask any economist about the “theory of the firm” (what firms do and don’t do), and they will emphasize the importance of information. So we decided to move beyond asking, “Do prediction markets work?” and instead use them as a tool for better understanding how information flows within a (very cool) corporation. Information is terrific to theorize about, but hard to measure. That’s where the prediction market is useful: if you and I trade similarly on a market, then we can infer that you share similar information.
Our research uses this insight to try to understand which employees trade in correlation with which other employees, and, hence, to measure how information flows within Google. We came up with some pretty interesting findings:
Sitting within a few feet of a workmate has a big effect. (Our data includes the exact GPS coordinates of each person’s desk, as well as their previous desks.)
Beyond this, sitting on the same floor as someone barely has any effect.
There is some evidence that organizational proximity (such as working for the same managers) also plays a role.
Professional relationships are quite important, although having been assigned to the same cross-departmental project as another person didn’t have any effect.
Shared interests (such as being on similar e-mail lists) also yield similar trading behavior.
Yet when Googlers were asked who their friends were, the relationships they reported did not explain trading behavior at all.
Demographic similarity was surprisingly unimportant, except where two colleagues shared a native non-English language.
We also have some interesting insights about biases in prediction markets. If there’s interest, I’ll return to this in a future post.
The overwhelming importance of “micro-geography” was quite striking, particularly as this is the sort of organization in which Instant Messaging and e-mail (plus blogs and wikis) might have otherwise suggested the death of distance. Certainly this research changed my mind about the importance of open-plan seating. This isn’t a lesson lost on Google either, as cube-mates are kept in close proximity, and Googlers are asked to move desks approximately once every three months. Interestingly, personal relationships persist once these moves have occurred, and people tend to trade in a way correlated with that of their cube-mate from three months ago; although, reassuringly, they do not trade in a way correlated with their future cube-mate. (I say “reassuringly” because this is a useful way of testing whether our results reflect Google seating people with similar opinions near each other, rather than people near each other influencing the opinions of others.)
I don’t know about your firm, but we academics are too self-important to ever sit in cubicles. Our research suggests that this may be unfortunate, and perhaps many of the best ideas in economics never occur, because the idea is waiting for us at a water cooler conversation at which we never arrive. I would love to see my colleagues brainstorm more often and more freely. If we can’t tear down the physical walls between our offices, how can we all change our workplaces to encourage the free flow of information and ideas? Comments are open.
Related links: click here for the full research paper; a companion post at the Google blog; a recent write-up in the Times. Other commentary: Marginal Revolution, Andrew Gelman, Zubin Jelveh at Portfolio, Justin Lahart at the Wall Street Journal, BloggingStocks.