This is a sample article — and the showcase of the design space. Its tags are cs + economics, so this page renders between the terminal and the library: mid-tone background, grid and paper grain both faintly present.
A prediction market looks like a gambling product: shares in "X happens by March" trade between 0 and 100 cents, and pay out a dollar if X happens. But squint at the mechanism rather than the product and you find something stranger — a computer, of an unusual architecture, that pays its components to work.
The architecture
Consider what the system does, in computational terms:
- Input is private information, held by scattered people who may not even know it is relevant until they see a price they disagree with.
- The incentive to write is arbitrage: if you know something the price does not reflect, the price is an offer to pay you for it.
- Aggregation is the order book. Each trade moves the price by an amount that reflects both the trader's confidence (position size) and the market's resistance (liquidity).
- Output is the price itself: a single number readable by anyone, which under the right conditions approximates the probability of the event.
Hayek made the general version of this observation in 1945 — prices are a "system of telecommunications," compressing dispersed knowledge nobody holds in full. A prediction market is the purified lab version: it strips away the goods and leaves only the information-aggregation machinery running bare.
The update rule is a program
The cleanest formulation is Hanson's logarithmic market scoring rule, where a market maker quotes prices from the current net position :
price(q) = exp(q/b) / (1 + exp(q/b)) # a sigmoid over net bets
cost(q→q') = b · [log(1 + e^{q'/b}) − log(1 + e^{q/b})]Trading against an LMSR maker is performing a Bayesian-flavored update on
the public estimate, with the liquidity parameter b playing learning rate:
small b and every trade slams the price around; large b and the machine
is stubborn, demanding a large position to move the output. Anyone who has
tuned gradient descent will recognize the failure modes on both ends.
Where the computer breaks
Every architecture has its bugs, and this one's are economic rather than electronic:
- Thin liquidity — with few traders, the "output register" is mostly noise from individual position-taking, not aggregated belief.
- Correlated inputs — if all traders read the same news, the market multiplies one signal rather than aggregating many, and confidence is an illusion of consensus.
- Oracle risk — the payout rule needs someone to decide whether X happened. The computer's ALU is a committee.
- Reflexivity — when the measured world reads the output and reacts (markets on elections, on Fed decisions), the computer is inside its own input loop, and the fixed point may not be the truth.
None of these are fatal; all of them are the reason "the market says 70%" deserves the same skepticism as any other program's output. A computer is only as good as its inputs, its incentives — and its operating conditions. The interesting move is refusing to treat either economics or computation as the metaphor, because at this level of abstraction they are the same subject: mechanisms that turn dispersed local decisions into a global result.