Warner

At some point this morning it became clear the Mark Warner would not be seeking the Democratic presidential nomination for 2008. I'm not sure exactly when the first public announcement appeared, but here's a Reuters post stamped 9:55 AM.

As you might expect, the price of the Warner contract on Intrade (which would have paid $10.00 if he had been the nominee) collapsed rapidly from $1.50 to practically zero:

The price at 9:40 was $1.50; by 10:41 it was 11 cents. What's interesting about this is that the decline in price should have been matched by comparable increases in the prices of other contracts. Specifically, if the market believed that there was a 15% chance of a Warner nomination prior to the withdrawal, this probability should have been reallocated to other contenders once the announcement became public. This did not begin to happen for quite some time, and the process is still incomplete. The only contract to have risen significantly at this point is that for Edwards:

A sharp increase in price, from $0.81 to $1.28, took place during a ten minute interval starting at 11:33, about an hour after the collapse of the Warner contract. A couple of other contracts (Clinton, Bayh and Obama) have also seen modest increases but the main beneficiary of Warner's withdrawal seems to be Edwards.

Price Inconsistencies

In my last post I provided an example of price inconsistencies across contracts in the Iowa Electronic Markets. These discrepancies are usually pretty minor. But price inconsistencies across different exchanges can be quite large, which makes you wonder about the rationality of traders.

Here's a recent example. The IEM House06 market has two contracts (RH.hold06 and RH.gain06) which, taken together, are equivalent to the HOUSE.GOP.2006 contract on Intrade. In fact, IEM's RH.hold06 contract is dominated by Intrade's HOUSE.GOP.2006 contract, since the former expires at zero if the Republicans gain seats. So the price of RH.hold06 should never exceed that of HOUSE.GOP.2006.

But take a look at the prices of HOUSE.GOP.2006 on Intrade between 10/3 and 10/6:

At no point did it cost more that $4.90 for a contract that pays $10 if the Republicans keep the House. The average price was $4.52 on the 4th and $4.38 on the 5th, with almost $120,000 worth of contracts traded over those two days combined. Now take a look at the IEM prices on October 4 and 5:

Date	Contract	Units	$Volume	AvgPrice
10/04/06	RH.gain06	128	3.348	0.026
10/04/06	RH.hold06	275	140.369	0.510
10/05/06	RH.gain06	249	5.812	0.023
10/05/06	RH.hold06	673	336.133	0.499

To buy a set of contracts that are equivalent to HOUSE.GOP.2006 cost $5.36 on average on the 4th and $5.22 on the 5th. That's a price difference of more than 18%.

Why did traders on IEM pay so much more for contracts that they could have purchased on Intrade? I'm not really sure about this, but it's clear that the traders on these two markets are receiving different information, or processing it in different ways. Meanwhile, the stakes on IEM are too small to make it worthwhile for someone to engage in arbitrage across the two exchanges.

Arbitrage in Prediction Markets

One of the interesting things about the IEM Congessional Control Markets is that there are lots of different ways in which to take a position on an event. So, for example, if you want to bet on the event that Democrats take control of the House, there are four ways to do it: (i) buy RH.lose06 in the House06 market, (ii) buy a fixed price bundle and then sell both RH.gain06 and RH.hold06 in the House06 market, (iii) buy both NH_RS06 and NH_NS06 in the Congress06 market, or (iv) buy a fixed price bundle and then sell both RH_RS06 and RH_NS06 in the Congress06 market. In a thick market with heavy trading and optimal behavior on the part of traders, these four different ways of entering the same position should cost the same. But these are thin markets with varying degrees of sophistication among participants, so prices across contracts and markets are sometimes not mutually consistent.

In fact, sometimes prices get so misaligned that there appears an opportunity for aribrtage: one can take a combination of positions that yields a sure profit (regardless of what happens in the election). Consider for instance the prices as of 7:30 CST this morning:

07:30:00 CST, Thursday, October 05, 2006.

Symbol	Bid	Ask	Last
RH_RS06	0.445	0.460	0.430
RH_NS06	0.031	0.045	0.031
NH_RS06	0.300	0.315	0.320
NH_NS06	0.208	0.235	0.207

Symbol	Bid	Ask	Last
RH.gain06	0.022	0.039	0.022
RH.hold06	0.516	0.534	0.520
RH.lose06	0.455	0.515	0.462

At these prices, a trader could bet on Republican loss of the house in the House06 market and simultaneously bet on Republicans maintaining control of the house in Congress06 for a sure profit. Selling RH.gain06 and RH.hold06 in the House06 market, and selling NH_RS06 and NH_NS06 in the Congress market has a total net cost of $0.954 per contract. Whathever happens on November 7, the positions in one market will expire at $1 while positions in the other will expire worthless. So you're buying $1 for a $0.954 on each set of contracts traded in this way.

Nobody is going to get rich looking for opportunities such as this, which is why it's so easy to find them. But this example is a nice way to see the logic underlying arbitrage based pricing relationships such as the put-call parity theorem.

Out of Hibernation

I think it's time to revive this blog after almost three years of hibernation. With the elections fast approaching, there's a lot of interesting activity on the prediction markets. I've been looking at Intrade and the Iowa Electronic Markets (Intrade is the same exchange as Tradesports, about which I've posted before).

What interests me about these markets is the manner in which their design affects the accuracy of their predictions. The prices in one market can be inconsistent with those in another market, sometimes even on the same exchange. This sets up opportunities for arbitrage, although the gains are typically pretty small. It also means that the markets sometimes make conflicting predictions, which raises the question of which market is more accurate on average. This is where market design and participation matter.

So what's been going on recently? Well, Intrade's SENATE.GOP.2006 contract (which pays if Republicans end up with 50 or more seats) has fallen from above 80 to around 75, which means that the Dems have about a one-in-four chance of taking the senate. The decline seems to be driven by Harold Ford's momentum in Tennessee: the TN.SENATE06.DEM contact now gives Ford an even chance of winning.

Meanwhile, thanks perhaps to this, the HOUSE.GOP.2006 contract (which pays if Republicans maintain control) has tumbled too, and now gives the Dems a slightly better than even chance of taking the House.

More on Policy Analysis Markets

It looks like some version of the Policy Analysis Markets will be launched after all, but without any government involvement. As noted in my previous post, such markets could be effective aggregators of information if structured properly. As far as I can tell, the markets will be run by a private company (San Diego-based Net Exchange). If profit condiderations drive decisions about market structure, then it seems unlikely that the market rules and procedures will be optimal from the perspective of aggregating information.

Policy Analysis Markets

In an extraordinary sequence of events last month, a Defense Department project aimed at constructing a futures market in catastrophic events appeared briefly in the media spotlight, received blistering attacks from all corners of the political spectrum, and was hastily dismantled, all in the space of a few days.

Notwithstanding the sudden demise of this particular initiative, the basic idea underlying it is important enough to deserve a calm, sensible and non-partisan public debate. Similar markets already exist. Dublin-based Tradesports.com offers futures contracts linked to the capture of Saddam Hussein and Osama Bin Laden, as well as the search for unconventional weapons in Iraq. The Iowa Electronic Markets offers contracts tied to the outcome of the Democratic presidential nomination process, as well as vote shares in the 2004 election. The Iowa Markets have managed to forecast vote shares consistently better than most opinion polls. Perhaps such markets can forecast attacks, assassinations and social upheavals and in doing so, reduce their incidence and mitigate their most harmful effects. At the very least, it's worth giving the idea a thoughtful public airing.

The main argument in favor of such markets is that they efficiently aggregate dispersed information to produce a more reliable forecast than could be obtained by more conventional methods. If this argument is to be persuasive, then the same asset traded in different markets should fetch roughly the same price. A quick comparison of the Democratic nominee contracts in IEM and Tradesports is enough to give pause even to the most ardent market enthusiast. It is easy to find significant and persistent pricing disparities for identical contracts across the two markets.

At the time of this writing, contracts that pay $10 if Gephardt is chosen to be the nominee can be purchased for 70 cents on Tradesports but cost $1.19 on IEM. Over the first two weeks of August, such contracts have been available for purchase in large quantities on Tradesports at prices never exceeding 80 cents. Over the same period, equivalent contracts have traded on IEM for prices as high as $1.59, and for an average price of $1.18. Similar disparities have persisted in the pricing of Lieberman contracts, but in the opposite direction: such contracts have been significantly cheaper on IEM relative to Tradesports over this period. Why the difference? What prevents the equalization of prices through cross-market arbitrage? And which, if any, of these market forecasts should be taken seriously?

While there is no substantive difference in the terms of the individual contracts, there are several differences in the rules and procedures governing the two markets. The most important of these concern the information available to traders about orders placed by others, and the types of orders that market participants are permitted to place. Such differences in rules can affect the ability of markets to serve as efficient aggregators of information.

Traders in the Iowa Markets can see only the best available prices at which the asset can be currently sold or purchased (the bid and ask prices). They cannot see the amounts that could be traded at these prices, nor can they see the range of other prices and quantities at which orders have been placed. In contrast, Tradesports allows participants to see the entire order book, consisting of the prices and quantities associated with every active and unfilled order. Having access to the order book allows traders to make inferences about the information available to others, which can potentially lead to better information aggregation. In this respect, the Tradesports design is superior.

A second major difference in market structure pertains to the kinds of orders that can be placed. The Iowa Markets allow for the purchase and sale of contract “bundles” either directly from the exchange at a fixed price, or from other traders at currently posted prices. A bundle consists of the entire set of contracts, exactly one of which will eventually pay off while the others expire worthless. The exchange buys and sells bundles at a dollar apiece, since this is exactly what the eventual payoff to the bundle will be. Having bought a bundle, traders can sell off one or more of its individual components on the market: this is how they can bet “against” an event.

The ability to buy and sell bundles can have dramatic effects on the speed and efficiency with which markets respond to new information. It ensures that when the price of one asset rises, the others almost simultaneously decline. Furthermore, it allows traders with good information but small budgets to periodically recycle their funds and have a disproportionate impact on the market. For instance, if they are able to systematically identify temporarily underpriced contracts, they can accumulate bundles over time and sell them back to the exchange, thus freeing up funds for subsequent purchases. In Tradesports they would have to sell these contracts to other traders on an individual basis, which is more time-consuming and less reliable. Along this dimension, therefore, the Iowa design is superior.

There are other differences in market structure. Trading is free in the Iowa Markets, while Tradesports participants pay a small percentage of the contract face value on each trade. Prices on IEM can be posted in multiples of one-tenth of a cent, while the tick size in Tradesports markets can be as high as 10 cents. Tradesports shuts down for an hour a day, while trading is continuous on IEM. In general, lower fees, smaller tick sizes, and continuously open markets can be expected to result in broader participation and better information aggregation.

If the original purpose of a market in catastrophic events is to be met, the structure of the new market should combine the best features of existing ones. Even the most carefully constructed market, however, may fail to yield reliable forecasts. Policy Analysis Markets are likely to be thin, subject to large price swings and relatively easy to manipulate. Low probability events would tend to be overpriced since it is expensive and risky to bet against them. For similar reasons, high probability events could be systematically underpriced. Despite these potential pitfalls, it is worth paying some attention to the question of how such markets should be structured, because their emergence, with or without governmental involvement, is surely inevitable.

A Field Experiment on Labor Market Discrimination

A recent paper by Marianne Bertrand and Sendhil Mullainathan has been getting a lot of press in the shadow of the debate over the University of Michigan’s affirmative action policies. The paper, entitled Are Emily and Brendan More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination, is accessible to the non-specialist and well worth reading.

The authors started with a collection of resumes that were downloaded from web sites catering to jobseekers, and modified them to remove identifying information. They then sent out resumes in response to vacancies advertised in Boston and Chicago, after first randomly assigning to each resume either a "white-sounding" or a "black-sounding" name. The names were obtained from historical birth records, and were based on the relative frequencies of names in black and white households. Hence names that were either much more common in white households than in black ones (such as Meredith and Todd) or much more common in black households than in white ones (such as Tanisha and Hakim) were selected for the study. Names such as Michael and Vanessa, which are common in both communities and hence carry little or no racial association, were not used. The authors found that resumes which had been randomly assigned white-sounding names elicited significantly higher callback rates than those assigned black-sounding names.

Many economists will be tempted to interpret these findings through the lens of statistical discrimination theory: if names carry information about the populations from which subjects are drawn, and these populations differ with respect to their productive characteristics, then names will carry statistical information about worker productivity. The authors try to test this interpretation by selectively improving some resumes prior to the random assignment of names. This is done, for instance, by filling in gaps in employment history and adding volunteer work. They find that such improvements significantly increase callback rates for resumes assigned white-sounding names but have negligible effects for those assigned black-sounding names. This flies in the face of the standard statistical discrimination model, in which greater information about candidates should narrow rather than widen discriminatory treatment.

So what’s going on? The authors are careful to leave open the possibility that more sophisticated models of statistical discrimination can account for their findings. But sometimes the simplest explanation really is the right one, and in my own view this is the case here. It seems to me that there is a subset of employers who have a strong negative gut-reaction to a black-sounding name and don’t bother to scan resumes for additional information once this reaction is triggered. The number of such individuals may be small relative to the population of employers, but they must be sufficiently numerous for their behavior to result in statistically discernible aggregate effects.

To see just how much publicity this article has generated, try a google search for black-sounding names, where you'll find links to the Chicago Sun Times, Boston Globe, New York Newsday, and the New York Post among dozens of media outlets.