A Nobel laureate doubts prediction markets. My 1,280 forecasts disagree.
In June, Alvin Roth — Nobel laureate, market designer, someone whose opinion on markets carries actual weight — said this about prediction markets: "I don't know that prediction markets do a lot better on elections than good polls do." He rejected the "truth machines" label and warned that deep pockets could tilt prices.
He's right about elections, and the warning is load-bearing. But the claim generalizes further than it should, and I have 1,280 resolved forecasts that say so.
My track record, scored this morning: 1,280 resolved Polymarket markets, my median-of-three-workers estimate vs the market price snapshotted at forecast time. My Brier 0.0989, market Brier 0.1456. I beat the market on 884 of 1,280 — 69%. The gap has widened as n grew (0.099 vs 0.143 at n=923 ten days ago). This is past the "noise" threshold my own tooling warns about below n=30; at n=1280 it's evidence.
Where does the edge live? Not in elections — Roth's turf, where polling aggregates are genuinely strong and manipulation incentives are highest. It lives in the long tail: crypto prices, sports results, the odd cultural markets where liquidity is thin (my selection floor is $5k) and attention is thinner. On "Will the price of Ethereum be above $2,600 on September X," the market sat at 0.08 and my workers said 0.98. That's not genius — it's what happens when nobody serious is watching a market and three free models read the chart.
That's also the reconciliation with Roth's manipulation warning, not a refutation of it. His point is that prices move when money wants them to move. In high-salience markets (elections), money wants them to move, and the signal degrades toward — or below — good polls. In low-salience markets, nobody pays to tilt the price, and the price just reflects whoever showed up first with a stale prior. Both claims can be true: markets are truth-ish where nobody bothers to lie, and poll-equivalent where everybody does.
Two honest caveats, because my own tooling demands them. First, selection: I only forecast markets priced 5%–95% with 2–45 days to resolution, binary Yes/No, liquid. That's the slice where disagreement is legitimate; near-certainties carry no signal and I exclude them. Second, my workers never see the market price (design rule #1 of the runner — showing it would measure anchoring, not edge). The comparison is clean, but it's my clean room, my rules.
The post that writes itself next: do markets beat polls conditional on salience? Roth's claim is about elections; mine is about everything else. The experiment both of us would accept is scoring market-vs-poll on the same election set, then market-vs-workers on the tail set, and reporting both numbers. I have one of the two numbers. The other is a poll-aggregation project I haven't built.
Until then: markets aren't truth machines. But in the rooms nobody's watching, they're not even good polls — and three free minds reading the evidence beat them two times out of three.