Three weeks out from the US election, the polls report a coin flip. Polymarket reports 60-40 for Trump.
The polls have said 50-50 for months, always inside the margin of error. Polymarket carries hundreds of millions in cumulative volume behind the drift. So the discourse picked its fighter: "a whale is manipulating the odds". Maybe. That is the wrong question. The right question is what each number measures.
Polls herd, measurably
A poll is an instrument wrapped in incentives. Response rates sit under 1%, so every published number is mostly weighting model rather than raw sample. Pollster incentives point inward. Publish an outlier and be wrong, and you are the industry joke. Publish the consensus and be wrong, and everyone was wrong together.
Nate Silver proved the effect statistically a decade ago. The method is the beautiful part, because it is a variance floor argument. Random sampling forces a minimum spread. Two honest polls of 800 people each should disagree by about 3,5 points on average from noise alone. Real methodological differences only add more.
The late-October polls of this race cluster tighter than that mathematical floor. Someone nudged the numbers toward the pack and suppressed the spread. Herding is an inequality violated, and anyone can check the arithmetic. A 50-50 that every pollster reports is insurance.
Markets invert the incentive. Being right against consensus is the only way to profit. Nobody gets paid for agreeing with everyone.
This is the older technology, by a century
Rhode and Strumpf documented the enormous election betting markets that ran on Wall Street from the 1880s to the 1930s. In peak weeks the betting turnover exceeded trading on the stock exchange itself. Newspapers printed the odds as news. The prices called every clear winner from 1884 to 1940, except the photo finish of 1916.
Inaccuracy did not kill them. Scientific polling arrived in 1936 as a cheaper information substitute, and legal gambling alternatives took the rest. Gallup undercut the business model. Betting priced elections in production for fifty years. The open 2024 question is whether these markets, with this liquidity, still do.
The manipulation experiment already ran
The whale question has a body count. In 2012 one Intrade trader spent the final two weeks buying Romney contracts. He lost an estimated 4 to 7 million dollars, at times a third of all Romney volume, apparently to keep the race looking close. Rothschild and Sethi dissected all ~84.000 transactions afterward. The whale built a price firewall. Obama capped near 70%, Romney floored near 30%, and the wall held right until Ohio was called. Then reality marked his book to zero.
That episode proves two things at once. A determined actor can distort a thin market for a while, so the number you read carries that risk. Today's talk about a large Polymarket account is a legitimate version of it. And every cent of distortion is someone else's subsidized arbitrage. Push a 50% probability to 60% and you hand free expected value to everyone who disagrees. Hanson, Oprea and Porter ran the experiment in a lab. Manipulation attempts got absorbed with little lasting price impact, because informed traders eat the manipulator.
Robin Hanson pushed the logic one step further with his subsidized market makers. If noise and manipulation pay informed traders to show up, a manipulator funds price discovery like a philanthropist.
Elections resolve. The result marks the painted tape to truth in November. Herding is free, invisible, and career-positive. Pick your poison, but only one of them has an immune system.
Markets leak
Wisdom-of-crowds arithmetic is the weak reason to read a market price. The strong one is the property Wolfers and Zitzewitz formalized. A market is the one public number private information cannot resist touching. A campaign's internal polling, a fund's commissioned surveys, an insider's nervous certainty: none of that reaches a pollster's weighting model. All of it reaches a market, because sitting on tradeable information costs money every hour. The price is a leak aggregator.
Three honest caveats before you treat the price as gospel.
- Price is not the crowd's mean belief. Manski showed that heterogeneous beliefs and budget constraints leave a 60% price bounding the mean belief inside a wide interval. Wolfers and Zitzewitz later gave conditions where price ≈ mean belief, but the equality is an assumption rather than a theorem. Read 60% as a region.
- The favorite-longshot bias is real. Betting markets overprice longshots and underprice favorites at the extremes. A 5% contract is usually worth less than 5%. At 60-40 the bias is small, and it says cheap-looking tails are cheap for a reason.
- Thin books leak noise along with information. Polymarket's election book is deep by crypto standards and shallow next to anything a bond desk calls liquid. Depth is the exchange rate between dollars and distortion.
One caveat runs the other way. Markets are partly parasitic on polls. Watch the tape after a big poll drops, because polls are an input to traders. The two instruments are a food chain rather than independent witnesses. Polls still do what markets cannot: crosstabs, state-level structure, the why behind the number. The market compresses everything into one probability. The poll tells you which voters moved. You want both, and you want to know which one copied the other's homework this week.
So read the 60-40 as a leak. Someone, somewhere, is betting real money that the polls are herding again. A probability is not a prediction. The losing side of 60-40 should win four times in ten, and the "markets were wrong" takes will write themselves. In three weeks we get one draw from the distribution. The polls will explain themselves afterward. The market already committed its statement.