Google Prediction Markets Case Study

PREDICTION MARKETS AT GOOGLE

CASE STUDY

-Hrishabh Ashok Sobti

Google was founded by Larry Page and Sergey Brin in 1998 as an Internet search engine company, and quickly grew popular as it displayed most relevant search results compared to other search engines - so much so that by March of 2007, “to Google” had become a verb meaning “to search for information on the Internet”. In 2007, five googlers(google’s employees) – Cowgill, Kirnos, Banks, Friedman and Na1 were working on an internal prediction market. Originially Cowgill’s idea, he posted the idea on a message board that was shared and public across Google. There on, he was able to get enough volunteers to dedicate their 20% time2 to this project that would be called “Google Market Prediction”(GPM). Cowgill also consulted with Hal Varian3 on how to design and implement the market. After extensive reasoning, the team of googlers agreed to stick with the IEM4 Model for their GPM implementation, along with other guidelines:

1. The IEM model relied on the use mutually exclusive securities that covered all possible outcomes for an event. If an event E had possible outcomes – A, B and C, and a participant wanted to invest in outcome C, he/she could buy that security using ‘Goobles’5.  And if that C was trading at a price of 0.25, the market “believed” there was a 25% chance that C would occur by the end of that particular quarter.

1. The team decided to use Gooble. At the start of each quarter, traders’ accounts were reset, and all traders were given 10,000 Goobles.

1. They also realized that a direct monetary award could tempt some people to behave counterproductively, yet wanted to reward successful traders. To that cause, they defined that each trader’s final balance would entitle them to the same number of lottery tickets. The more balance at the end of a quarter, the more chances for winning a $1000 lottery. Since this would encourage them to maximize the number of tickets, they would trade as intelligently as possible. Other rewards included T-shirts to top performers.

While Cowgill and his team had the righ idea, and an thought-out implementation, the concern raised by Dolores Haze6 cannot be overlooked. Despite the accuracy of the market – where the most expensive outcomes actually occur, GPM can not practically replace other planning or forecasting tools for the following reasons:

• Out of the 14 functions as google, how do you guarantee a fair highlight of each function’s topics of forecast? Rather – How do you make sure that no function feels neglected.

• From the “Total Accounts Over Time”infographics, only ~2000 accounts are created, out of a total of 45000+ employees. That’s less than 4% participation – only a fraction of which participate.
• Although the total shares traded over time have grown steeply, it can be attributed to a few users trading much more frequently than other.
• Given Vladimir Humber’s account on GPM particpation, it is clear that users are also misleading the market.
• The accuracy of GPM could be highly contextual - the “Trading Base by Function” infographic clearly shows that most participants tend to make more trades relevant to their own function. Trading over a project’s timely completion can accurately be predicted by the members of the relevant function – naturally boosting the accuracy of prediction.

However, it is most interesting how GPM works in complete harmony with the work culture – imitating the function of a stock market, yet reinforcing values like recognition over monetary gain.