Thursday, May 17, 2007

Prediction Markets at Google

Bo Cowgill (Stanford Public Policy '03) presented in my E-business and Data Mining class on Monday about how Google use of prediction markets. Here is a summary of his presentation.

Prediction Markets at Google

Origins, Objectives and Implementation

Objective:
  1. Empowering average employee
  2. Improve awareness
  3. Quantify unquantifiable
  4. Improve morale

25 markets per quarter for 2 years.

Markets about:

  • Product launches (number of Gmail users)
  • Office/position openings (will this important position be filled on time)
  • Employee quality of life (when will gym re-open)
  • Competitors’ actions
  • Fun markets (who will win the NBA championship)

Implementation

  • Continuous double auction with limit orders
  • Alternative would be with market makers.
  • People can not see each other’s position.

Growth and Usage

Market stats

  • Number of shares and number of trades has grown over time (due to added liquidity and more participants)
  • Approximately 68,000 trades had occurred by the end of 2006, through a total of 2000 trade accounts
  • Long tail distribution:
  • One participant has made 3100 trades in 1 year 8 months (~5 per day)
  • Approximately 400 participants have made 2 or less trades
  • A little more than 50% of trades are made by engineers (but this reflects the overall structure of the company).
  • People that are higher up (closer to CEO) and have a long tenure at Google are more likely to participate.

Individual performance

  • Geography was a better predictor than job function of success
  • Finance and software engineers have done better (best explained by bots?)
  • "Batting Average" increases slightly with market participation (evidence of learning or do individuals who are not good quit?)
  • Success for certain individuals has been consistent from quarter to quarter

Market set up:

  • Typically 5 “options”
  • E.g., when does a certain project launch: Jan, Feb, Mar, Apr, May or later

Analysis of Predictiveness

  • Average winning price 0 weeks before closing was $55; average losing price was $15
  • Some predicted better (close to $100 the week of closing), whereas others like the Bush v Gore election was still 50/50 on election day

Rewards

  • Money earned through trading is used to buy raffle tickets
  • Prizes are then awarded through a random drawing
  • Financial rewards seem not as strong as ‘Reputation’

Further Research

No comments: