Quote from: edmundedgar on February 20, 2014, 03:00:05 AM

Quote from: gwern on February 20, 2014, 02:21:08 AM

Quote from: edmundedgar on February 20, 2014, 02:14:59 AM

I don't doubt that the market can sort out the problem of ambiguous or badly-worded contracts (to within reasonable tolerances) but that's not the problem I'm talking about. The key innovation in this proposal is to use a p2p system of consensus witnesses to settle the contracts. It looks to me like the p2p system will collapse into de-facto centralization, at which point the incentive system, which is supposed to be designed to incentivize getting to true outcomes, will actively prevent the market from correcting.

Alright, you didn't like my concrete example of why the system can be self-correcting; so could you offer a concrete scenario of how an existing prediction market contract from Intrade or somewhere could collapse into a perverse equlibrium? (Being partially centralized is not a bad thing: Bitcoin mining is pretty centralized these days. I don't mind if everyone is getting their figures from an authority like the BLS.)

...
1) Somebody publishes a convenient data feed saying what happened, pulling from the BLS and a bunch of other sources.
3) Everybody has an incentive to match the data feed.
5) The data feed produces bad data, because their systems are buggy or they're getting paid off by someone.
...
Anyway let's see what the OP says about the centralization issue - maybe I've misunderstood the proposal, or maybe they've got a solution to it.

tl;dr A little of both. Truthcoin designed such that voters know what to vote without consulting any external source at all. Authors should not create high-search-cost-Decisions as they risk loss of funds.

INCENTIVE COMMENTS - Why centralization might not matter
I am of the same opinion of gwern that bad data will be ignored.

Firstly, I state in the whitepaper that these markets are only appropriate in certain conditions, one being low search costs (and I don't consider relying on an automatic feed to be actually 'searching for the information' at all, as the automation literally avoids doing just that, which implies that searching for the info was not worth the voter's time [ ie high cost] ). Much more on this later.

To the main issue: in order to prevent people from lying about reality, people vote in a simple binary True/False way but are actually punished continuously, such that the most deviant person is punished more (much more) than the second-most-deviant, and so on. Therefore actually every voter has incentive to actually lie to each other, including tricking rival voters into using a broken feed, hoping to fool someone into voting incorrectly. For this reason voters may anticipate that >51% of other voters have noticed and caught the error in the data source, and so themselves correct the error.

Additionally, voting involves investment. To vote, one one needs to buy some coins (or forgo the opportunity cost of selling preowned coins), and yet one only realizes the full benefit of this investment in the future, in the form of collected trading fees. "Screwing up" crashes the trust of the system, meaning low (or zero) trading thus ruining the profitability of the investment. For this reason, voters might vote against a bad data feed even if they believed that it would likely win >51% of the voting power. If the bad data successfully alters the network, all coins might be worthless, so 'risking' them on an honest vote (and just hoping and praying to grab the critical 51%) isn't much of a risk at all. Quite the reverse, actually.

So really it is a nice mix: traders are paranoid enough not to trust each other (which might lead to herding as you indicate), and yet greedy enough to cooperate to promote the value proposition of the service. It is quite a bit like Bitcoin mining this way, where miner-competition secures the network, yet the mining pools are ready-and-able to hold these emergency cooperation-meetings to solve Bitcoin problems.

Finally, different Truthcoin branches can change the decision fee magnitudes. Branches with higher decision fees are more profitable, and voters would have more to lose by disappointing Traders, and so more to gain by 'double-checking'. I once worked in marketing dept for an upscale trading firm, every single email, pamphlet, website edit was literally triple-checked by the hawk-eyed lawyers / grammar PhDs of Compliance. I literally mean three times by experts. Internal Auditing would regularly force someone to drop everything and spend a week preparing for what the firm would do IF we got 'externally audited' for the accuracy of our published claims (which, as far as I could tell, never even happened). Creating a new branch is (theoretically) very easy, so if some voter-failure is inevitable, competition should at least produce exactly an efficient level of failure.

SCOPE - Avoiding Centralization
Now my discussion of the scope of this project. I regret not specifying this in the paper as clearly it really is quite important.

Unfortunately, (and I mean it), I could not find a magical solution to everyone's prediction market problems. However, I focused on a design that should be able to accomplish "the easiest" PMs, like "Will Hillary Clinton win the US Presidency in 2016?", impending wars, critical regulatory outcomes, disasters, or breakthrough technological/biomedical research. Items that, given the widespread nature of the info, and social benefits, we really """should""" have. I'm not a very emotional guy but I literally experienced sadness and disappointment, that we as a society (including Silicon Valley, scientists, philanthropists) simply did not care enough about progress to construct PMs ("institutions that actually work"). Bitcoin isn't perfect. For a while it was really hard to use, it still is to most people. 1 hour confirmation time? Soft forks? Having to spend days downloading and verifying the entire transaction history? legal ambiguity, crazy exchanges, network analysis of the blockchain, malleability, it goes on and on.

Yet Bitcoin restricted its focus to what it could accomplish, and so despite abuse it's still alive and kicking. I also placed restrictions on Truthcoin, but mine are easier to misunderstand. For example, early in the paper I note that the voters are instructed to vote .5 (the "information unavailable" vote [ recall that this vote punishes the Authors who created this Decision ]), even when the information IS available if THEY also DECIDE that the search costs of that info are prohibitively high (as in 'more than a google search')! I expect this attempt to generally fail, with voters doing a lot of extra research-work in the hope of not upsetting any Authors/Traders, thus improving "the service" but in the process opening up a can of systemic risk.

So Truthcoin was designed with a focus on after-the-fact outcomes of which everyone is already AWARE (as in, already thinking about RIGHT NOW, such as the identity of the US President). Because different humans can be aware of different things, the 'branching' capability allows us to cheat this limitation. "Truthcoin - Main" might be full of hobbyists who just google the latest DJIA, but "Truthcoin - Serious Finance - US CRUDE Wednesdays" might be a niche market with restrictions and incentives such that only professional oil speculators would decided to own those coins, and so they have absolutely no need to rely on any outside source because the information they provide to their branch every Wednesday evening was what they already spent their entire day doing. This is an extreme example to make the point, I don't actually expect to see a branch of this kind (but whatever the market will support is fine with me).

CONCLUSION
For these reasons, my expectation is that point 5) will actually not occur, because although lazy voters might (conveniently and efficiently) "auto-pre-fill" their Ballots, I expect them to be so paranoid as to at least glance over it manually to see if someone has tricked them into casting a vote indicating that John McCain was elected president.

For further thought: one idea I had to increase reliability at the expense of speed, would be to have multiple rounds of voting, where (for each Decision) the "newest" round is more influential than all previous rounds. Then, the truth would have to deliberately and slowly build up to some set parameter of "certainty", with attacks delaying this build up but not preventing it. As feeds would (we hope) eventually become correct, attackers would have to be constantly attacking the validity of the vote system. I concluded that that was overkill for the basic scheme but it might be appropriate for a branch.

------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Now a rambling comment for EE about my failed attempts to create a decentralized time-series feed:

I get the feeling that you may be comparing Truthcoin and RealityKeys a little too directly (not through any fault of yours). As I indicated on Twitter, I feel the services are so different that they can hardly be competitors, with RealityKeys being a low-cost high-volume service and Truthcoin being high-cost low-volume (although with branching who knows). For starters the information in a data feed is generally not in anyone's awareness, because we've passed the job completely to robots.

Specifically, I doubt that this service will be used for 'feeds' at all as the cost would be prohibitive. Think about it: a time series needs maybe 10 partitions per time unit to guarantee any useful information. If you wanted a daily feed, that's 300 per month for one variable. I expected Voters to do about 100-200 Votes per month per "branch" (as in, in TOTAL).

"Why not have the voters just vote on a level, such as 4, 70, 8, instead of just on TRUE FALSE?" Well I'm glad you asked.
1] I exploit .5 as a maximally-uninformed outcome in 1 vs 0 land, whereas this is impossible with continuous variables because the mean, median, and mode a) are just estimates of 'maximally-informed' and not logical requirements for 'maximally-uninformed, b) can all be gamed, and c) are so flexible that I lacked the ability to even determine the entire attack space.
2] Confusion over rounding, units, language. Is 3.0104 different from 3.01040013? Which is better? What if the contract says "rounded to the 6th decimal place", but the 2-6th decimal places are zero?
3] PCA actually assumes that all the variables are normalized, implying a measurable mean and sd. What happens to those numbers if some lunatic with 1% of the voting power types in 6 quadrillion?
4] True / False harmonizes the units, which are identical [0,1] range, preventing the incentive scheme from varying with a shift in the average-unit-variability of Decisions.
..and MORE.

I did derive this insane way of doing a continuous ts variable "feed", which ended up using regular expressions to constantly search the marketplace for a set of markets worded in an pre-standardized way and determine the relevant range in which a continuous variable was expected to exist. Then I tried to make it so that entrepreneurs would always have an incentive to create-and-trade-in markets on a log scale as close to the real-world level as possible such that the range would shrink to an acceptable level of clarity. I could not figure out a way to make this viable. In fact I am going to have nightmares tonight just re-living this explanation.