Unicast

I'm now sitting at the T-Mobile hotspot workspaces at Heathrow Terminal 3 on my way home from the Mini-Conference on Information and Prediction Markets organized by Marco Ottaviani at London Business School. Here are some quick notes on the papers for those who actually like my distorting abstracts. I might write more when I get a chance to actually read the papers, perhaps after my econometrics exam.

Paul Tetlock of UT Austin, the first person I've ever heard use the word "epsilon" in a normal sentence (remarking that his prediction markets book was "epsilon away from publication"), presented a paper, written by himself, Robert Hahn and Donald Lien, on decision markets, which he defines as an information or prediction market where a decision maker will make some kind of decision based on the outcome of the market. In the model the decision maker is the market maker and commits to a market making rule ex ante and subsequently makes a decision based on equilibrium prices based on an ex post optimal decision rule. There is a manipulator and an expert with an inperfect signal about the fundamental. The decision maker's profit (or loss) comes both from the use of the information and from trading in the market itself. The authors show that it can be optimal for the decision maker to participate in the market to subsidize liquidity, in particular if the market is illiquid or if the decision is particularly important for profits.

Robin Hanson presented his paper on logarithmic market scoring rules for combinational information aggregation. Only logarithmic market scoring rules preserve certain important conditional independence relations. I think I've heard or read all this before.

After the coffee break there were three papers on horcerace betting that I had a hard time following. Leighton Vaughan-Williams talked about efficiency in exchange betting markets and some of his experiences advising the UK government on taxation of betting. Olivier Gergaud presented an empirical paper on forecasting in rank-order tournaments using data from forecasts of French horse races that is published in newspapers on the day before a race. Frederik Koessler presented an experimental paper with different belief elicitation treatments.

Over the lunch break John Delaney of TradeSports, who sponsored the lunch, talked about some of the things that TradeSports is concerned about, chief of which is regulation. They recently gained eBOT status in the United States. Robin Hanson wondered about the pricing model that TradeSports and other exchanges use — in a market with high fixed cost (technology, regulation, etc.) and low marginal cost, he expected to see more bundling and two-part tariffs.

James Adam from ICAP, one of the world's largest inter-broker dealers, talked about the Economic Derivatives from Goldman Sachs and the Chicago Mercantile Exchange. He could not share volume information, but it was my impression that several hundred million dollars of risk is traded at the monthly non-farm payroll auctions; volume in most of their other markets is apparently quite small. The format is a 1-hour auction where it is possible to revise and pull orders, which are limit orders. The market structure is a parimutuel market, but because of the use of limit orders, participants would not be forced to take unfavorable odds, which is the case in real parimutuel markets. He mentioned one incident where a participant placed a very large order at the beginning of the auction, only to pull it just before the auction ended, frustrating a lot of people. The participant was warned not to do this again or risk being banned from the market.

Robert Wilburn, now at Rimdex but formerly of NewsFutures, discussed corporate prediction markets. He gave the strong impression that NewsFutures has had some very important corporate clients that they're not allowed to name, including pharmaceutical companies. Sales forecasting still accounts for around 80% of corporate prediction markets. Many corporate deployments are relatively small affairs with budgets on the order of $25k, where the sponsor is usually a mid-level manager, and these markets are often not used for prediction events that are truly important for the company, which means that they are often not as valuable as they could be. It is difficult to generate sufficient participation and liquidity for these small markets. Robert also told us that some of the examples of corporate prediction markets that are quite hyped in the press are actually pretty small deployments, but that other more important NewsFutures cases are confidential.

Justin Wolfers talks very fast and does lot leave a lot of time for reading his slides. His paper is to some extent a response to Manski's result that prediction market prices do not have a clear interpretation, and in particular they can not generally be interpreted as probabilities or a weighted average of beliefs. Justin relaxes some of the assumptions, most importantly Manski's assumption that all participants bet the same amount regardless of the odds, and demonstrates that for a wide variety of utility functions and belief distributions prediction market prices can be interpreted as a wealth-weighted average of beliefs.

Peter Norman Sørensen, a professor at my department, presented a paper about aggregation of information and beliefs. In his setup prices end up being a risk tolerance-weighted average of beliefs. He shows that for general utility functions in a fully revealing rational expectations equilibrium, it is not possible to disentangle information aggregation and belief aggregation, although it is possible for CARA utility (see equation 2 in the paper). Then there was a lot of talk about likelihood rations and priors; I will have to study my notes more carefully before I'll understand what that was about. I think the main argument between Justin, Eric Zitzewitz, Robin Hanson and others was about whether Peter's concept of priors was actually what we're interested in eliciting using prediction markets.

Erik Snowberg of Stanford GSB presented a paper on partisan impacts of the economy. Traditional literature in the field has not been able to effectively control for reverse causality when analyzing the effect of partisanship on economic indicators. Erik, Justin and Eric Zitzewitz conducted some event studies during the 2000 and 2004 presidential elections where they regressed certain economic indicators that were traded throughout election day, namely oil and index futures, against prediction market prices. The data set for the 2004 election started around 18:00 Eastern time, slightly earlier for the 2000 election, when most polls had closed and therefore economic indicators was thought to have little impact on voting intent. Any new information influencing prediction market prices only concerned the election itself. One of the interesting results was that a Bush victory was expected to raise both the S&P 500 index and bond yields significantly.

Koleman Strumpf presented a paper on manipulation in the political stock markets that ran in the US from the Civil War through World War 2 as well as the results of a field experiment in the Iowa Electronic Market.

Eric Zitzewitz presented a pre-written paper (with Justin Wolfers) on full distribution event studies. They use option prices to deconcolve a probability distribution for a security and for example showed the distributions for Merck on the day before and after the Vioxx announcement. There was a mean shift but also more risk (more probability mass in the tails).

Tags: predictionmarkets