« Interpersonal Morality | Main | Touching Vs. Understanding »

July 29, 2008

Comments

I agree. Great explanatory post. Reputation points should accrue to Peter.

Much appreciation to Peter.

These may represent the first major case study of real-money conditional futures markets.

Correction: The Iowa Electronic Markets had markets that could be interpreted as conditional forecasts of election given nomination, advising parties on who to nominate for President. But these new markets of Peter's arguably offer more socially valuable info, helping voters see for whom to vote.

I cannot see any way that Intrade lets you know the market is subsidized. It seems like this should be useful information for traders since, since it lets them know that their expected return is positive. I too am extremely happy to see experimentation with prediction market subsidization.

Are these bets really 'conditional', or rather conjunctive? That is, do you get your money back if the condition doesn't obtain in the end (say I bet on non-dem.DebtIncr, but then a Democrat is elected)? A cursory glance at the contract rules made it look like you instead lose -- "contract expires at 0" -- but maybe I've misunderstood something...

'Increase in US government debt (over $10 billion)'

Wait, what? Is this per year or increases over current rates or something I am misunderstanding? Did you mean '$10 trillion'? (That makes more sense to me.)

Just a pleasant reminder: Back on the original OB thread where we discussed these conditional futures, I proposed an alternate conditional futures market: instead of betting specifically on whether e.g. oil futures and the Democrat contract move the same direction on election day, what we should do instead, is bet on the overall correlation over all the days up through the election, between *changes* in oil futures each day, and the *changes* in the Democrat contract each day.

So e.g. you would bet on whether oil's price always goes up as the market deems Obama more likely to win, and vice versa.

This is obviously a superior metric in that it is robust against noise and/or manipulation that can happen on that one vital day. It inputs variation on many days, exploiting much more relevant information.

I explain this all and argue the point with Robin Hanson and Peter McCluskey in the thread linked above, but unfortunately, this is kind of like talking to a wall. They seem to spend very little effort reading my posts, resulting in responses that stem from a misunderstanding, rather than being serious criticisms. And, Peter McCluskey makes arguments against my metric that apply several times over to his own.

(Of course you're not obligated to read my posts, but if you're going to respond, don't do a half-job.)

Does anyone here agree with my point that my metric is superior and would be a far better prediction market contract? Would I get the same appreciation and reputation points if I gathered this data and funded such a contract? (Oh, and can someone turn off the italics in the linked thread?)

What I find most pleasing about this design is that it avoids the awkward construction that Richard alludes to, and at the same time spurs trade by virtue of the changing denominator. Excellent.

I think the actual subsidy is much smaller although it did seem that over $1000 was given away at inception by setting the initial prices at 50.

I can't agree that the claims I'm subsidizing have been a success yet. Note that NONDEM.PRES-GOVT.DEBT is priced as if it could be expired above 100, but the rules say the maximum expiry price is 100. That's a clear sign that there hasn't been enough trading to make the price of that claim reasonable.
A few of the "Impact of Next Pres." claims have traded. I bought and sold 2 contracts of PRES.McCAIN+LESS.CRIME in June. I'm not sure what time period the Intrade volume figures cover.
Gwern, "over $10 billion" and "over 2000" mean that you divide debt changes by $10 billion and divide troop levels by 2000 to get the corresponding contract price.
Silas, it's too late to change the existing contracts, and I'm not about to subsidise more contracts for this election. I will watch carefully for possible manipulation, and will consider better ways of minimizing manipulation risks if I subsidise similar contracts in the future. I think the rest of your arguments have been adequately answered. I certainly encourage you to subsidise additional contracts.
Jason, two of the contracts will expire at zero when the election is settled. I will lose nearly $8000 on those. For the others, it's unpredictable.

Peter: Silas, it's too late to change the existing contracts, and I'm not about to subsidise more contracts for this election.

Were you able to understand how I wasn't asking you to change the existing contracts or subsidize others, but rather, proposing to subsidize my own superior ones that have better safeguards than "I will watch carefully possible manipulation"?

Silas, if you improve my ability to understand the present and predict the future, the yes, I will definitely award you proportionate appreciation and respect.

When I told my father the current price for McCain on Intrade, he wanted to bet on him, but his bank blocked the credit card transfer.

(My father predicts that Obama will win the popular vote but lose the electoral vote, guessing that he'll sweep the blue states and lose enough red states, but barely.)

Silas: Sorry to hear that.

Robin: That claim definitely demands some serious acknowledgment. It sure does sound like Silas proposed a better system for extracting info from markets.

Silas, as I explained before, we expect that on election day most of the causal direction of the correlation is from the vote to the consequences, while for most days before there are large causal forces in both directions. Also, we want a market estimate as useful as possible right up until the election, but just before the election the overall correlation price would be dominated by the previous correlation history. (And do you really imagine I had not considered your obvious variation when I chose my design?)

Without knowing anything about the trading algorithm Peter is using, I suspect that subsidize is probably the wrong word for what he's doing. Market-making is a very common trading strategy that can be highly profitable if done correctly, even without knowing much about the actual value of what you're trading. I'd assume that the algorithm adjusts the bid and ask prices whenever a trade is made, meaning that if the initial market is wrong, the algorithm will correct it. If there's enough trading volume, Peter will make enough from the bid/ask spread of 2.5 to cover any losses from the initial market being incorrect.

Props to Peter for taking on the risk in order to get the market going, and more props if he ends up making a profit. Anyone interested in setting up a prediction market based on whether Peter McCluskey comes out ahead?

I doubt Peter makes a profit on this batch. That would require a lot of trading with that small a spread.

But in principle you are right. There's got to be some sweet spot of volume and good guesstimation of the initial line where this is profitable even with the fairly small bid/ask.

If a somewhat larger spread would still invite interest in the market, it should be fairly easy to make it profitable as long as transaction costs to the market maker are low (much smaller than the spread).

Robin: Silas, as I explained before,

Where? Would you mind pointing me to the first comment you made in that discussion, in which you understood what I was proposing, and correctly stated what you meant? Because that was kind of an issue back then.

we expect that on election day most of the causal direction of the correlation is from the vote to the consequences, while for most days before there are large causal forces in both directions.

Well, you certainly *hope* so, but then, the market has usually largely incorporated expectations about the winner before election, making the signal really really small, and allowing the random influences to dominate.

Let me take a gander at Peter McCluskey's shock futures market. Is it consistent with my theory that bidders will assume that other random forces will determine the election day correlation?

"49.9-50.1"

Yep, that's betting on a coin flip alright!

Also, we want a market estimate as useful as possible right up until the election, but just before the election the overall correlation price would be dominated by the previous correlation history.

Yes, so in the unlikely event that the correlation of "probability of democratic win" and "oil futures" is very far from the unknown noise-filtered, manipulation-filtered election day correlation, this metric will have inaccurate results.

If you really think the election day's correlation is that much more important, then you can say:

"Silas, that's a really good idea, but to account for election day's more relevant information, the metric should treat election day as 30 days of price history for purposes of determining the correlation that resolves the bet."

Then, you give stronger weight to election day, but still incorporate all the available information and dilute irrelevant influences!

(And do you really imagine I had not considered your obvious variation when I chose my design?)

I invite anyone to read the thread and decide if Robin's responses indicate that he had considered the idea before.

P(Robin responds to misunderstandings of my idea | Robin has considered my idea before) = low
P(Robin responds to misunderstandings of my idea | Robin has not considered my idea before) = high

Being argumentative for the sake of argumentation isn't productive.

Biases?

(1) Column headings:
"Dem price non-Dem price Dem norm non-Dem norm"

Why not Democrat and Republican? Those are the two major parties. What about third
parties?

(2) "non-Democratic administration"

"Democratic" is the system of Democracy. Democrat is a political party in America. Although I'm sure the DNC loves how people conflate the two.

Why not Democrat and Republican? Those are the two major parties. What about third parties?

I suspect the intention was to make the probabilities ("Dem presidency" and "non-Dem presidency") add up to be exactly the same as the probability of any kind of presidency.

"Democratic" is the system of Democracy. Democrat is a political party in America. Although I'm sure the DNC loves how people conflate the two.

I don't think it's a conflation; it's simply a word with more than one meaning. The word Republican is the same: it also has a meaning distinct from the name of a party.

Now, if only Intrade worked in the United States without requiring a bank transfer...

Blogged and flogged.

I'll miss my posting privileges here. I really will :-(

The comments to this entry are closed.

Less Wrong (sister site)

May 2009

Sun Mon Tue Wed Thu Fri Sat
          1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
31