the futarchy proposal is old enough now that its rough edges should be widely known, and they are not. the pitch — vote on values, bet on beliefs, let conditional markets decide policy — is easy to state and hard to poke holes in with undergraduate economics. the subtle problem, which the proposal's defenders tend to wave at and its critics tend to overstate, is that a decision market does not price what the pitch claims it prices. it prices a correlation. the policy question wants a counterfactual.
set up the canonical example. a dao asks "should we hire alice as ceo?" and runs two conditional markets: "price of token if alice is hired" and "price of token if alice is not hired." the rule says hire alice iff the first price exceeds the second by some threshold. now ask: what belief does a trader with positive edge have to hold to push the alice-hired market up? the answer is not "alice would be a better ceo." the answer is "conditional on alice being hired, the token will be worth more than the market thinks." those are not the same statement.
the gap is the usual reason conditioning goes wrong: selection. if alice only gets hired when the dao is in some other state (say, confident, well-funded, aligned with alice's plan), then the conditional market on "if alice is hired" is pricing the token outcome under that joint state. the not-alice market prices the token outcome under the complementary state, which is a different world. the decision rule compares them as if they were interventional contrasts — "hire alice and see what happens" vs. "don't hire alice and see what happens" — when what the market actually trades is observational.
hanson's original formulation acknowledges this and handles it by randomizing the decision: with small probability, the opposite choice is made, which turns the observational market into an interventional one at the margin. if you actually randomize, the critique is toothless. the problem is that nobody who writes about futarchy in practice — including in dao governance contexts where the proposal has had its most realistic chances — includes the randomization step. they take the market prices and use them deterministically. which means they are running the mechanism in the mode where it is known not to work.
this would be a footnote if the failure were small, but it is not. in any setting where the decision is endogenous to the same variables that drive the outcome, the causal-correlational gap is the mechanism's dominant source of error. imagine you are deciding whether to ship a product. the "ship" market prices the outcome conditional on shipping, but shipping decisions are correlated with readiness, team morale, and fifty other unobservable signals. the "not ship" market conditions on the absence of those signals. the spread is almost entirely driven by the selection process and has almost no causal content.
people sometimes propose fixes that do not fix it. the first is "use a longer time horizon so selection effects wash out." they don't, because the decision does happen, and once it happens you are only ever in one of the conditional worlds. the second is "let the market size grow to absorb informed traders." this helps the efficiency of the conditional price as an observational statistic, which was never the problem. it does not bridge the gap to counterfactual. a third, more serious proposal is to combine the markets with a causal identification strategy (instrumental variables, synthetic controls) so the price movements can be interpreted causally. this works but requires real econometric machinery, and it is not a thing any working prediction-market platform has built.
the cleanest solution is still hanson's: make the decision genuinely probabilistic. flip a coin weighted by the market's posterior; execute whichever outcome wins; pay out against the realized world. with enough randomization, the observational and interventional distributions coincide. the cost is that you sometimes have to take a decision you think is worse in expectation, which is a tough political sell in a world that wants its decision markets to feel like optimization rather than lottery.
where i have landed is that futarchy-style conditional markets are genuinely useful for information aggregation — i want to know what the market thinks the token price would be under each scenario, independently of whether i use it to pick — and much more dangerous as automated decision rules than the literature admits. the belief-aggregation interpretation survives almost any amount of selection bias. the decision-rule interpretation does not.
if you are designing a conditional-markets platform from scratch, this suggests two things. first, expose conditional prices as information, not as a vote. let people trade and read them; do not auto-execute on the spread. second, build the randomization machinery up front, so that when a dao does want to hand over a decision to the market, the mechanism is running in its theoretically-clean mode. that is not hard to implement. it is just boring infrastructure that does not survive a pitch deck.
i think conditional markets become the most useful financial primitive of the next decade — there is almost no end to the set of questions you want to ask conditional on hypothetical policies. but building them as "decision markets" in hanson's original sense requires swallowing the randomization pill, and nobody in the prediction-market space has really wanted to. worth watching whether the dao-governance crowd forces the issue.