The Now Economy

HedgeStreet issued a press release last week touting a new interface and improved platform. The UI is somewhat simpler, but I can’t tell that it is a big improvement. They mention two changes:

• Rather than paired yes and no contracts on each question, they now have a single instrument for any claim that you can take either position on. They describe it as buying or “selling short”, but I think their terminology is confusing.
• Their interface offers better charting, streamlined order entry, and centralised information about the holdings and account history.

Their description of betting against a position as “short selling” seems confusing to me. After reading all their help screens, it appears that when you sell short, you are investing money and acquiring an asset that may pay out a positive amount, just like when you buy. HedgeStreet wants to present it as simple by saying there is only one asset, and you can buy it or sell it short. The problem with their description is that it isn’t the same as short selling. Their model is actually more similar to the “buying complementary assets” model I described in my introduction to basic Prediction Market formats. Brokers in the stock market can’t sequester the entire amount that might be required to repay a short position, since the underlying asset can grow without bound (see Google for an example.) HedgeStreet’s contracts pay out at either $0 or $10, so HedgeStreet can (and does) collect the entire price up front. Short sellers pay $10 less than the stated price for an asset with a positive payout. Buyers pay the stated price, and also get an asset with a positive payout. The short selling terminology will be confusing to people who understand short selling, and will be a handicap to anyone else who tries to understand short selling later.

They could patch up this description by saying that rather than having a margin requirement, they simply collect the maximum loss up front, and repay it when the question is settled, but at that point, they’d be better off describing it as buying the opposite position rather than explaining how they stretched the analogy to short selling.

HedgeStreet’s new interface is certainly simpler than their previous version, which had paired markets in each outcome. When they divided possible crude oil inventories into 4 bands (x < A, A < x < B, B < x < C, and C < x), they had 8 distinct markets for one outcome. With their new software, they will have only four markets, which is an improvement, but as I said in my talk at the Prediction Market Summit in San Francisco, and will repeat in New York, you can do better by linking the markets together. I haven’t yet written up the explanation for that point; it will appear on this blog when I do.

It also appears that HedgeStreet is now saying that all trades will be made directly between their customers. I haven’t found their old documentation (the Wayback machine is slow-to-nonresponsive this afternoon) so I can’t verify what they used to say, but they seemed to have automated market makers in some cases with the old platform. I don’t see how it’s an advantage to either HedgeStreet or their customers to drop that possibility. In illiquid markets, an automated market maker can, at limited cost, increasing trading possiblities substantially.

We released a new version of Zocalo in December. (There have already been about 50 downloads, even without an announcement.) It is available on SourceForge. It includes some new configurable features for prediction market experiments: the main one is that experiments can be run in which traders keep their earnings across rounds within a session. (This allows experiments in which there is progressive revelation of information over the session about the underlying value.) In order to better support longer-lived sessions, we fixed a bug that allowed data to overrun the right edge of the trade history chart.

In parallel, we have started working on support for long-lived markets. This is the main step needed to support internal markets within a business or organization. In order to handle that, we’ll need persistent data for trader’s accounts and book orders, as well as an ability to display the price history to traders without requiring that they be continuously connected to the market (this is currently necessary for the experiments that Zocalo supports). I’m pleased to say that I have integrated both Hibernate (open source Object-Relational Mapping software) and JFreeChart (open source Chart drawing) into Zocalo. Neither is called to the full extent necessary for a usable long-term market, but both are being used thoroughly enough to show us that the integration works and they are being invoked appropriately. The next step is to ensure that all aspects of users, accounts, and orders are persistent. I’m hopeful that I’ll be able to demonstrate this functionality by the end of the month. (This functionality is not yet in the released versions. We’ll publish it as soon as its usable and stable.)

Smarket is an interesting new market that lets you bet on whether the rank of books and other products on Amazon will go up or down. This is another addition to a type of market I don’t think there’s a good theory for yet; markets with open-ended pricing. There are several other examples including the Hollywood Stock Exchange (HSX), ProTrade, and Yahoo!’s Tech Buzz Game. These markets all try to follow the real-world rise and fall of some abstract value (box office returns, on-field athletic performance, and search rank, respectively), and allow traders to bet on them.

So far, all the markets of this type use only play money, and that’s part of what I’m talking about when I say there’s no theory. One of the innovations that makes standard Prediction Markets (of the kind I described last week) workable is that a risk-neutral market operator can sponsor a market without taking a position in the trading; that is, without taking on risk. The traders buy and sell from one another, and the operator merely facilitates the trades. Even if you add an automated market maker, you can constrain the potential costs, while still facilitating unlimited trading.

Traditional bookies learned to handicap the events they were selling bets on or they went out of business quickly. Part of the development of insurance was inventing the actuarial techniques that allowed insurers to reliably estimate how their exposure would change as they sold various policies. Modern stock markets only allow short selling when the investors back their trading with assets that the broker can rely on if prices go the wrong way. Real money prediction market operators will need the same ability to constrain losses.

With the open-ended markets, the value to be paid out to stocks that gained isn’t limited, and isn’t related in a simple way to the price paid by the purchasers. I would expect a theory that showed how the market’s exposure is limited to start from the observation that one asset can only rise in value when another falls. This is true of Smarkets and TechBuzz because they’re based on relative rankings. It’s not obvious whether it’s true of HSX (Box Offices returns are only weakly limited) or ProTrade (the formulas for ranking athletes haven’t been published.)

It’s also not clear how to read the market results as predictions. In a standard Prediction Market, the traders’ incentive is to bid the price up or down until it matches the subjective probability, so prices should equal probabilities. David Pennock says the optimum target on TechBuzz is the square root of the search share. I think a separate analysis will have to be done of each institution, and if the rule isn’t easy to understand, it will be hard to claim that traders are producing prices that follow an optimum rule.

There also isn’t yet any theory regarding the feedback from underlying value to market value. Most of these markets seem to reflect changes in the underlying value by paying dividends to holders of the assets. Sometimes they (also? instead?) change the market maker’s current price for the asset. I expect this to make it difficult to come up with closed-form expressions of how the market operator’s exposure changes when assets are revalued. Remember that there’s no guarantee that similar amounts of money are invested in each market, or each side of any particular issue. And this kind of distinction is where the money pump was discovered in the TechBuzz Dynamic Parimutuel Mechanism. You can’t do this with real money if you don’t know what your exposure is.

These markets are definitely interesting. They allow Surowiecki’s Wisdom of Crowds to give us a hint about the rising and falling fortunes of things we care about. I welcome efforts to turn them into markets that can make predictions, or allow insurance, hedging, and the other side effects we expect from standard Prediction Markets.

Hat tip to Dave Pennock.