Subject: Re: idea futures markets (was Re: Tom W. Bell paper)
From: "Ben Tilly" <btilly@gmail.com>
Date: Tue, 5 Sep 2006 21:16:47 -0700

I got most of the way through responding, then I realized the
fundamental disconnect.  Kragen, you are seeing prediction markets as
a way of eventually *paying* developers for doing development.
Stephen is looking at prediction markets as a way of *funding*
development that hasn't happened yet.  They are completely opposite
uses of the possible market, and behave in opposite ways.

To pay developers to do development, you want developers to bet they
will succeed.  They put in money to get more back later if they do
their work.

Standard financial theory says that to fund developers, you do the
opposite.  Developers have lots of exposure to the possibilty of
succeeding, so they want to hedge that by betting that they fail.
That is developers are given money now but will owe more later if they
succeed.  This is an identical financial position to, for instance,
venture capital but you're making the bet on a market, not in a
privately negotiated transaction.

Keep that in mind throughout.

On 9/5/06, Kragen Javier Sitaker <kragen@pobox.com> wrote:
> [Good lord, 3200 words, that's at least ten minutes, terribly sorry]

For the future, it would be nice if you sent one email in reply to
each email you're replying to.  That would make it easier to read and
respond in pieces...

> Tom Lord writes:
> > Don Marti wrote:
> > > Yes, and so you would have to (1) do R&D and (2) take
> > > a position in the market based on "predicting" the
> > > success of your own project in order to profit from
> > > the market....
> >
> > Your obligation here isn't simply to show how an
> > innovator could win a game in a Spex, given compliant
> > counter-parties.   You need to show why those counter-parties
> > would enter the market in the first place (and you've just
> > provided an argument that they would not).
>
> Here are some possible reasons:
> 1. You want a way to fund free-software development, and it's an
>    alternative to bug-bounty systems, dominant assurance contracts,
>    and the street performer protocol.

If it is an alternative that won't work, and there are reasons to
think it won't, then this doesn't hold.

> 2. You stand to make X amount of money if a particular piece of
>    technology is developed (e.g. save on your electric bills or the
>    cost of developing some new feature for your web site, or come to
>    market sooner) and you're looking for a way to reduce your risk
>    from that uncertainty by investing somewhere around X/2 in a
>    futures market.

This one might work, but I doubt it.  The problem here is that to be
useful for hedging, a market needs a certain amount of liquidity and
volume.  These markets will lack that liquidity.  If a large company
needs to hedge a bet, they'll do like AOL did with their IE-based
browser.  AOL had an IE-based browser, and they were afraid that
Microsoft would dominate the browser market, and would charge very
high rates when they went to re-negotiate that contract.  So AOL
bought a failing company (Netscape) and ran it at a loss to guarantee
that when they re-entered that negotiation, they had an alternative to
bring up while bargaining.

It worked.  They got a good rate.  Furthermore part of what AOL's
final price when they renegotiated the contract appears to be that
they would cut back on funding Mozilla.  Which was OK for them then
because it was clear that at the next negotiation cycle, IE would have
viable competitors.

My point is that if a company is large enough to need to hedge
something like this, it is easier for them to just make a direct
purchase or investment as a hedge than to try to find a big-enough
counterparty in a prediction market.  (Seriously, in my example what
other players in the browser market could have realistically been a
counterparty for AOL?)

> Even if a small fraction of the money made in the market (3%, 10%,
> 20%) goes to the people doing the R&D, that might still be more
> efficient than the proprietary-software market.  (I think Bruce wrote
> an essay with this number some years ago, but I don't have it handy at
> the moment.)

If I'm a company and I want a feature badly enough that I'm willing to
pay the whole cost, I can simply hire someone to do it and 100% goes
to the people doing the R&D.  That's far more efficient than the
prediction market.

If I'm a company and I want a feature but not badly enough that I'm
willing to pay much of the cost, then I'm going to hope that someone
else will do it and I won't pay anything for it.  (In between there is
room for a small group of companies to form a viable consortium - but
the politics get complex as each company tries to pay as little as
they can while getting the project as a whole funded.)  This is the
classic problem with funding public goods.

> I'm not sure it would actually work, for the following reasons:
> 1. It gives the innovator a strong incentive to keep their
>    work-in-progress secret as long as they can afford; they only have
>    an incentive to reveal their results in order to engage in
>    profit-taking.

Which isn't helpful in an open-source environment.

> 2. I should probably try do the math before I shoot off my mouth, but
>    I suspect that adept speculators will be able to see price changes
>    induced by innovators' buy and sell decisions and infer the
>    information possessed by the innovator --- and take most of the
>    profits available due to the resulting information arbitrage.
>    (Maybe Warren Buffett's success in the stock market is a sufficient
>    counterexample.)

Adept speculators are only going to put the energy into analyzing the
market if it is big enough for them to make much money at it.  With
something like this, they will likely stay away in droves.  Unless it
is really simple to manipulate, which it might be if people try to use
it like you suggested.

> But it's probably worth a try.  It also has some advantages over
> bug-bounty systems, which share disadvantage #1 above.
> 1. While (as in a bug-bounty system) multiple innovators working in
>    parallel will reduce the expected payoff, the innovators know about
>    the reduced payoff when they buy their options and start their work
>    --- they just don't know why --- and they don't lose their payoff
>    if somebody else fixes the bug first.

This is only true if innovators use it as a way of getting paid.  If
they use it as a way of getting funded, then one innovator is fine and
the rest are royally screwed.

> 2. Knowledgeable speculators will adjust the payoffs of innovations so
>    that less-likely innovations have higher payoffs.

There really won't be that many knowledgable speculators.  The ratio
of profits to be made from the speculation to the knowledge needed to
do so is not very good.

[...]
> Stephen Turnbull writes:
> [reordered to improve the clarity of my reply]
> > Suppose you want money *now*.  You have web 3.0 in development, so
> > you set up a claim on web 3.0.  To get money, you must *sell*
> > claims, right?  You are now liable to pay somebody a lot of money if
> > web 3.0 actually works.  That's OK, you're going to make much more
> > profit than that, and in the end your VC is backing this play, it's
> > no more your money in this scenario than in the patents world.
> > Problem is, *I* invent web 3.0 *first* and get the lion's share of
> > the market.  Now how do you plan to pay off those claims?
>
> I don't think it's supposed to work like that, although I haven't read
> the paper, the excerpt Don quoted makes it sound like a standard
> prediction market, which doesn't have these problems.  You have web
> 3.0 in development, but just like the patent system, the system
> doesn't award any money for things that are in development.  You use
> some of your VC money to bet heavily on things that you hope your
> development is going to cause to happen: maybe the advertising
> networks go out of business, teenagers in Bolivia write software
> that's used in millions of US businesses, and people start writing
> car-crash simulation codes in JavaScript.

As I said above, it depends on how you expect the market to be used.
You and Stephen are looking at it in opposite ways, for opposite
purposes.

> The traditional prediction-market way of doing this is that you buy a
> pair of tickets, one that pays off $1 if DoubleClick goes bankrupt by
> 2009 and one that pays off $1 if it doesn't, the pair for $1; and then
> you sell the second one of these to someone else, maybe someone who
> stands to make a lot of money in 2009 if DoubleClick does go bankrupt,
> and wants to reduce their exposure to DoubleClick staying in business.
> Maybe they pay you $0.75 because the market thinks there's only a 25%
> chance of DoubleClick going bankrupt, so your total exposure is 25
> cents, which you have already given to the prediction market operator.

The net result of this scenario is that the developer pays 25 cents
now in return for (hopefully) a dollar later.  This is using the
prediction market as a way of paying the developer.

What Stephen is envisioning is something like a traditional futures
market, where someone sells a contract saying, "I will pay you X at
time Y if Z happens."  The seller of that contract gets money now, in
return for a potential liability later.  The contract is not covered,
but if the author of the contract doesn't pay up, they can face
penalties (up to and including jail).

> If you (or somebody else) ships web 3.0 and it causes these things to
> happen, then you cash in your DoubleClick-goes-bankrupt ticket and get
> back your $1, quadrupling the amount you initially put on the table.
> If unforeseen technical problems prevent you from shipping it, or
> unforeseen social problems prevent it from getting adopted, you lose
> those 25 cents.

If you ship web 3.0 and cause those things to happen, you'll stand to
make enough money that your piddling prediction market investment is
meaningless.  If you don't, you lose development costs (which you'll
lose anyways).

> It's not quite that all-or-nothing, though.  If you ship a web 3.0
> protototype in Q3 2007, maybe Wired will have a big feature about how
> it's going to put DoubleClick out of business by 2009, and all the
> speculators making the same bet that you did back in 2006 drives the
> price up, so now you can sell your DoubleClick-goes-bankrupt ticket
> for 40 cents instead of the 25 you bought it for.
>
> If you need money *now*, you need to find a claim that your research
> *already* gives you inside information about.  You bet in the
> direction your research tells you to bet, wait until the market knows
> what you already knew, then take your profits.  If you can persuade
> the players in the market quickly by announcing your results in a
> press release, publishing a paper, demoing at PC Forum, shipping a
> prototype, what have you, then you can take your profits a lot sooner,
> perhaps immediately.

And fear of exactly this will keep investors away in droves.  Rational
investors don't particularly like risk, and this is very risky.

> ("DoubleClick goes bankrupt" would seem to have some moral hazard
> associated with it.  Is that avoidable?)
>
> > Second, even with the legal changes, these markets may suffer from a
> > severe "market for lemons" problem, which will result in
> > underfunding of desirable projects.
>
> In the scenario Don's proposing, the "market for lemons" problem
> doesn't seem particularly severe.
>
> If I understand correctly, the classic "market for lemons" problem is
> that the seller of a car has more information about it than the buyer;
> the buyer is willing to pay only their expected utility from the car
> (minus epsilon), which is something like the average of their
> subjective values for all the actual cars on the market, while the
> seller is willing to accept only their utility (plus epsilon) for the
> car.  So if buyers and sellers have exactly the same utility function,
> only sellers whose car is worse than the average car in the market
> will continue to participate, which of course reduces the value of the
> average car, and the vicious cycle begins.

Yes.

> Of course, you need at least some variation in utility functions for
> people to bother with a market at all, so it's just a question of how
> large the information asymmetry is, relative to the ratio of
> utilities.
>
> In this case, the problem you're talking about is that the innovator
> (who might be either a buyer or a seller) knows, say, that cold fusion
> doesn't actually work, or that their new algorithm for solving systems
> of linear equations is about to speed up scientific computing
> substantially; while the insurance buyer thinks cold fusion might
> work, or scientific computing is going to stay slow.  In this case,
> the variation in utility is supposed to be provided by insurance
> buyers being exposed to risks outside the market, while the innovators
> want to be able to buy exposure to those risks, since they know which
> side of the bet to take.

The problem is that investors don't like investing in things they are
likely to lose money at.  If a big participant in these markets are
innovators who are hoping to take money from the markets, who prove to
be successful at it, then random investors are likely to lose money.
So they'll stay away and there is nobody putting up money to be taken.

> > we want to be able to extract funding for R&D, but there's no net
> > surplus to be extracted,
>
> There is a net surplus to be extracted from risk arbitrage.  See
> below.

Not really.

> Who would buy a "server CPUs stay hot in two years" ticket at $0.90,
> knowing the counterparty on most of the tickets was Transmeta in 1998?
> Presumably anyone who thought there was less than a 10% chance of
> Transmeta conquering the market (or maybe 2% if your internal ROI is
> already 4%, since it's going to come out of your cash flow for two
> years).  But a risk-averse agent who stood to save a lot of money if
> server CPUs got cooler might be willing to do it even if they thought
> Transmeta had a 20% or a 90% chance --- their overall expectation on
> the bet is negative, but its upside diminishes their exposure to the
> risk that they'll have to keep installing bigger air conditioners in
> their data centers.  In this case, they can hedge that risk completely
> by buying $1 of tickets for every $2 of air conditioners and other
> power expenses they expect in two years.  If they win the bet, they
> get $1.11 in winnings to offset each $2 of air conditioners, and
> they're out 89 cents, and if they lose, they're out the 90 cents per
> ticket but don't have to install the air conditioners, so they no
> longer worry about whether Transmeta wins or loses (assuming 90 = 89).
> If they have *some* risk tolerance, they can hedge against a smaller
> fraction of their risk, so it starts to matter whether they think
> Transmeta's chances are 5%, 20%, or 90%.
>
> So, in this example, the variation in utilities comes from
> risk-averseness, and as risk-averseness approaches infinity, even
> large information asymmetries approach insignificance.
>
> Also note that in this example, if Transmeta folks know they have a
> 90% chance of success but the market thinks they have a 10% chance
> (ignoring the ROI parenthetical above), then they have an expectation
> of getting out 9 times the money they put into the market.  And if
> Transmeta had been funded this way, they wouldn't have had to patent
> their inventions, so maybe Intel or AMD would have adopted them, and
> they *would* have been in every server by 2000.

But if Transmeta funds themselves like you suggested, they get no
money up front but lots of money after they succeed.  Which cash
profile does not lend itself to succeeding.

If Transmeta funds themselves like Stephen expects them to (which
actually gives them cash in hand, which is what they really needed),
then they are well and truly sunk if the prediction comes true but
because Intel succeeded in re-inventing the technology.  And in fact
with the liability of that prediction market hanging over Transmeta's
head, they have just made themselves a bad investment for anyone else.

> (Of course there are zillions of other risks: power costs may rise,
> Freon might get cheaper, your data centers might grow faster or slower
> than you expect, whatever.  That bet is only a hedge against one
> particular risk.)
>
> > the loss is a multiple of the amount of research they were able to
> > fund through it, while if they succeed, the profit is a fraction of
> > what they could have made.  This just doesn't make sense unless VCs
> > (or inventors) are *extremely* cash constrained now but expect to be
> > very rich in the future regardless of the success of the project.
>
> Those statements don't hold for the kind of market I described above,
> although since I haven't read the SPEx paper, I don't know if they
> hold for it.  The loss is limited to your up-front investment (in the
> market, not in the research), and the profit is a multiple of the
> up-front investment, possibly a large multiple if you make something
> happen that people thought was unlikely.  Furthermore, there's no
> necessary connection between the amount you bet in the market and the
> amount you spend on the research, but:

And the problem with your strategy is that you're asking someone who
has no money but wants to get funded to spend money that they don't
have to get money when they won't need it.  There is a flaw in this
approach...

[...]
> > What you're looking at here is a large collection of individual
> > securities, each subject to "events" that could imply swings of
> > hundreds of billions of dollars---and the inevitable bankruptcy risk
> > on the side facing the adverse change.  Would I buy a share on a
> > market?  Not a chance, [unless...]
> > ...
> > *What* was I thinking?  ...  the swing per market will be millions
> > of dollars at most, and there's plenty of room for diversification
> > by "large" participants.
>
> In the traditional prediction-markets scheme described above, the
> millions or hundreds of billions of dollars at risk are already
> deposited in the account of the prediction-market operator well before
> the claim comes due, and the only question is who they're going to get
> paid out to, not whether or when.  Perhaps the interest can pay for
> the referees.
>
> Some non-play-money prediction-market operators allow bettors to trade
> on margin, complete with margin calls and everything.

To be useful for funding companies, you need non-play-money prediction markets.

[...]

Cheers,
Ben