SI 779 Aggregation and Prediction Markets

Instructor: Rahul Sami (Office hours: Mon 5-6pm at #3246E SI-N/ Tue 1:30-2:30pm at 417BWH)

1.5 Credit, 7-week course module
First half of Fall 2008
Tuesday, Thursday 9:00-10:30am, 409WH
SI779 additional lecture: Mon 1-2pm (3224 SI-North) most weeks

This is a doctoral course that partially shadows SI679.

Course Goals:

Understand the prediction market design space, and how prediction markets compare to alternative information aggregation methods.
Receive a broad introduction to the research topics, methodologies, and results in this area
Recognize the formal and conceptual connections between prediction markets, learning, and information theory

Overview:

In many settings, the wealth of information on a particular subject is distributed among many entities, with no single source having all the relevant information. In this course, we will study approaches to elicit and combine this information in order to come up with a forecast or estimate that reflects the combined information of all sources. This idea of aggregating information from multiple sources is an essential ingredient of many applications, including weather forecasting, predicting election outcomes, market research on tastes, and assigning betting odds. Recently, prediction markets have been deployed to aggregate opinions and come up with forecasts on election outcomes, scientific advances, product delivery dates, Academy Award outcomes, and many other events. We will study theoretical and practical aspects of several aggregation tools, including opinion polls, machine-learning techniques to combine or select experts, scoring rules, and prediction markets.

Prerequisites

Doctoral standing or permission of instructor.
At least one course in basic probability or statistics, covering concepts such as probabilities, independent events, random variables, and expectations.
In addition, some familiarity with basic game theory, and the concept of Nash equilibrium, will be assumed.

Please contact me if you have any questions about this.

Course Schedule (tentative)

Week 1:
Common lectures: Introduction; prediction market design space; classifications of forecasting processes
Week 2:
Common lectures: Online learning; Sequential predictions and cascades
SI779 additional lecture: Connections between learning algorithms and prediction markets.
Week 3:
Common lectures: Evaluating forecast sequences; proper scoring rules and forecaster incentives
SI779 additional lecture: Scoring rules and entropy
Week 4:
Common lectures: Theoretical models of prediction markets; rational expectations equilibria; no-trade results
SI779 additional lecture: Price formation processes and critiques of rational expectations
Week 5:
Common lectures: Market forms: double auctions, parimutuel markets, and market scoring rules; implementation of market scoring rules; manipulation in prediction markets
SI779 additional lecture: Theoretical results on manipulation, and connections to information theory
Week 6:
Common lectures: Legal and policy issues; real money vs. play money; case study: market research
SI779 additional lecture:Inflation in play money markets
Week 7:
Common lectures: Case study: election markets; course review

Course Work and Assessment

Participation in class discussions 10%
Homework Assignments 30% Assignments will include problem-solving and short questions based on the techniques studied in class and the readings.
Final paper 60%
The final paper for the SI779 students will be a critical survey of the literature on some topic related to prediction markets and online learning. As a guideline, papers are generally expected to be about 10 pages long. Sample topics are (i) A survey of literature on manipulation in information markets; (ii) A survey and critique of comparative accuracy results for election markets and polls; (iii) Online learning algorithms for reactive experts.