The Machines Passed: Grading the 2026 Midterm Markets

The first prediction-market election is over

Polls across the United States closed on the evening of 3 November 2026, and by the time most people had gone to bed the result was already priced. On Kalshi and Polymarket, the two venues that between them handle almost all of the country’s event-contract volume, the implied odds of Democratic control had crossed ninety-nine percent for both the House and the Senate hours before the Associated Press was willing to call either chamber. The next morning confirmed it: Democrats had taken both. Bloomberg’s post-mortem gave the markets a passing grade, and the platforms spent the week publishing accuracy scorecards that read like end-of-term report cards.

What makes this election worth a closer look is not that the markets were right. Favorites usually are. It is who, and what, was doing the pricing. The order books that set those odds were not a crowd of politically obsessed humans refreshing a browser. They were, to a remarkable degree, software: autonomous trading agents and automated strategies that quote, hedge, and re-price around the clock. For the first time, a national election was forecast at scale by machines trading against one another, and we can now hold that forecast up against reality.

This piece grades the machines. Not the horse race, which the markets won comfortably, but the harder questions underneath it. How calibrated were the odds, and not merely how often the favorite was correct? Where did the markets miss, and why? Does the fact that bots now dominate the trading make the consensus smarter, or only faster and more alike? And what does any of it tell us about whether a prediction market actually forecasts, or just reacts to the same incoming numbers everyone else is watching? The 2026 midterms are the largest real-world test these questions have ever had.

What the markets said, and what happened

Going into election day the markets were not uncertain about the House. Both venues had Democrats above ninety percent to flip the chamber, a number that barely moved in the final week. The Senate was the live question. For much of the autumn the order books had actually leaned Republican on Senate control, in line with a map that favored the incumbent party, before traders flipped the bet and Democrats overtook Republicans in the closing stretch. By the Monday before the vote the Washington Post put combined wagering at nearly 1.9 billion dollars, with Polymarket pricing Democratic Senate control around seventy-eight percent and Kalshi close behind (washingtonpost.com).

Then the results came in, and both chambers went blue. The markets had the direction right on the two questions that mattered most. The table below is the scorecard the week produced, drawing on the post-election accuracy review published by RotoWire, which compared the two platforms on both their pre-close implied probabilities and their Brier scores, the standard measure of how well a probabilistic forecast is calibrated.

The headline is flattering. Across every contest Kalshi tracked, its favorite won 479 of 487 races, a 98 percent hit rate. Polymarket reported much the same, and edged its rival on the tightest measure, the House Brier score. But the scorecard also contains the whole story of where the machines fell short, and it is sitting in plain sight in the Senate row.

The accuracy numbers, decoded

A Brier score runs from zero to one, and lower is better. A forecast that says ninety-five percent and turns out right scores close to zero; a confident forecast that turns out wrong is punished hard. On the House, both platforms posted Brier scores in the thousandths, about as good as a binary forecast gets, and Polymarket’s 0.0030 was marginally better than Kalshi’s 0.0045. That gap is real but tiny, and it mostly reflects Polymarket’s willingness to push its House number a little higher. On the Senate, both scores jumped by more than an order of magnitude, to 0.0600 and 0.0625. That is the figure to watch. It is not a failure, the favorite still won, but it tells you the markets were far less sure of the Senate than of the House, and that their uncertainty was, if anything, not large enough.

Below the chamber level the granular record is strong. According to RotoWire’s review, favorites won 356 of 360 House races (95 percent), 32 of 33 Senate races (94 percent), 33 of 34 governorships (94 percent), and 26 of 27 attorney-general races (96 percent). Kalshi’s own recap claimed 95 percent accuracy on the House and 94 percent on the Senate; Polymarket claimed 98.4 percent across all resolved races, including 98.6 percent on the House. The platforms are measuring slightly different things, implied-probability level versus share of races called correctly, so the numbers do not line up exactly, and readers should treat any single self-reported figure as marketing until an independent reviewer reproduces it.

Still, the broad picture is not in dispute. As a machine for turning scattered information into a single, continuously updated probability, the markets worked. The interesting arguments start when you ask what that probability was actually made of, and whether being 98 percent right is the achievement it sounds like.

The misses: where the odds were wrong

The Senate Brier score is the tell. The markets gave Democrats a roughly three-in-four chance of taking the chamber, and in the aggregate that was correct, but the odds were built on a set of individual races the order books priced on the wrong side. Several seats the markets had leaning Republican broke for the Democrats, and the cluster of them is why the Senate number looks so much worse than the House.

The post-election tracker at Covers catalogued the most visible misses. In South Carolina, Democrat Annie Andrews beat a Republican the markets had priced around seventy percent to win. In Texas, Democrat Gina Hinojosa unseated the Republican incumbent in the governor’s race, a candidate favored at roughly sixty-five percent on the morning of the vote. The tracker listed a run of other Senate upsets in the same direction, in Iowa, Maine, Michigan and Ohio, all breaking toward the Democrats against the market’s lean.

Read individually these look like errors, and in a sense they are. But the deeper point is calibration, not box score. A market that prices a candidate at seventy percent is not promising that candidate will win; it is promising that candidates priced at seventy percent win about seventy percent of the time. Lose a handful of those and you have not necessarily miscalibrated, you have paid out the three-in-ten that was always baked in. The problem in 2026 is that the misses were not scattered randomly. They clustered on one side, which is the signature of a systematic bias rather than honest noise: the machines did not merely lose a few coin flips, they underpriced the size of the Democratic wave. Post-vote reviews converged on the same conclusion, that the markets identified the eventual winners about as often as the polls did and, like the polls, underestimated the Democratic margin (The Hill).

Election night in fast-forward

The clearest way to see what the machines do, and do not do, is to watch the Senate number move on the night itself. Going in, both venues had Democratic control of the Senate at around seventy-five percent, the softest of their headline calls. As the first eastern returns landed, the order books did what order books do: they re-priced, fast. By seven o’clock on the East Coast the Covers tracker had Democratic Senate control up to 89 or 90 percent, and by mid-evening both platforms had pushed past ninety-nine percent for both chambers (Covers).

That trajectory is the whole argument in miniature. On one reading it is impressive: a liquid market absorbing thousands of county-level data points and converging on the right answer within a couple of hours, no anchor desk required. On another it is almost tautological, because by the time the number hit ninety-nine percent the result was already most of the way decided by the votes themselves. The market was not forecasting the Senate so much as narrating the count, in real time and with a dollar sign attached.

What the fast-forward does not show is foresight about the surprises. The same tracker that caught the Senate surging to near-certainty also recorded the individual seats the market had on the wrong side until the returns forced the correction. Speed cuts both ways: a market that re-prices in seconds looks clairvoyant when it is confirming the expected and merely slow-footed when the expected is wrong, and 2026 served up both inside the same six hours.

Called it first is not the same as called it right

The platforms’ proudest claim was not accuracy, it was speed. By ten o’clock on election night both venues had pushed past ninety-nine percent confidence in Democratic control of both chambers, well ahead of the television networks, and Polymarket said it had called ten key races ahead of the major outlets by a median of thirty-one minutes (CNBC). RotoWire noted that traders on the older PredictIt venue called the House more than three hours before the Associated Press. If you treat being first as the goal, the machines crushed it.

But being first is a different thing from forecasting, and it is worth being precise about the difference. A race call is not a prediction once the votes start arriving; it is an inference from partial returns. A market that reaches ninety-nine percent an hour before the AP is not seeing further into the future than the AP, it is applying a looser standard of proof to the same incoming county-level numbers. That can be genuinely useful, because a fast and liquid consensus is a real public good, but it is not the thing the markets are usually sold as, which is an oracle that knows the outcome before the evidence does.

Rajiv Sethi, an economics professor at Barnard College who has studied these markets for years, offered the deflating version of the accuracy story. Calling both chambers correctly, he pointed out, is not statistically remarkable when both were already the favorite: “If you toss two coins, there’s a 25% chance that you’ll throw two heads in a row, and nobody would be terribly surprised if you did” (CNBC). The sharper critique is that a market which merely tracks the vote count and the polling average is not adding a distinct signal, it is repackaging the existing one faster. That is the same question crypto keeps asking of every new piece of market plumbing: does it discover a price, or just relay one? The honest answer for election markets sits in between. The odds are a real-time aggregator that beats the pundits on speed and matches them on direction, which is both less and more than the marketing. We have grown used to seeing a number attached to everything now, from a contested Senate seat to the eighty percent the market put on Bitcoin’s next soft fork, and the midterms are a reminder that the number is only ever as good as the information flowing into it.

Who did the trading: the bots won the turnout

Step back from the election for a moment and look at the order book. The reason the midterms are a crypto story, and not only a politics one, is that the odds that called the races were set in large part by autonomous software settling in stablecoins on public blockchains. By the time the campaign began, a review of Polymarket’s public leaderboard found that fourteen of the twenty most profitable wallets were bots, that AI agents accounted for more than thirty percent of wallet activity, and that over thirty-seven percent of those agents were running a positive profit and loss, against less than half that rate for human traders (CoinDesk).

The flagship of that fleet is Polystrat, an autonomous agent built on the Olas protocol by the firm Valory and launched on Polymarket in early 2026. Unlike a script-based bot that fires on fixed rules, Polystrat wraps a large language model, so a user can set a goal in plain English and let the agent interpret markets, size positions and trade around the clock. Its builder, Valory chief executive David Minarsch, has described it as “an autonomous AI agent that trades on Polymarket 24/7 on behalf of its human user,” and has been candid about the limits: simply prompting an off-the-shelf model with a market, he told CoinDesk, “usually results in outcomes no better than a coin-flip.” The edge, where there is one, comes from the workflow around the model, not the model alone. But in aggregate, Minarsch said, “agents tend to do better than humans.”

That is the backdrop to the 2026 scorecard. When commentators say the prediction markets called the midterms, what they are really saying is that a population of trading agents, competing for profit and re-pricing continuously, converged on the right answer faster than the human analysts did. The accuracy is a property of the machines, and so, as we are about to see, are the blind spots.

Retail went pro, with antennas and Claude

The agents did not arrive in a vacuum. Around them a whole cottage industry of semi-professional traders has grown up, and the texture of that scene says a lot about where election markets are heading. CNBC spent time this year with the people who trade Kalshi and Polymarket full time, and the picture is less Wall Street than a very online arms race.

There is Steve Farmer, an Atlanta data engineer who runs an AI bot on economic contracts from his wife’s home office and reports a roughly seventy percent win rate, crediting the machine’s lack of feelings: a bot, he said, does not carry the impulse and greed that wreck human traders. There is Logan Sudeith, a twenty-six-year-old who bought a broadcast antenna so he could watch live television with less lag than the streaming feeds and trade ad-related contracts a fraction of a second faster, a setup he says earned him 250,000 dollars in a single month. And there is Kenneth Deneau, a thirty-two-year-old former institutional investor who uses Claude to scan newsletters and Discord servers for tradeable signals, calling the AI assistant one of the most transformative parts of his operation (CNBC).

The platforms have leaned into it. Kalshi’s Pro terminal, launched for its most active speculators, offers the kind of latency and multi-market tooling that used to live only inside trading firms. The point is not that any one of these traders moved the midterm odds; most down-ballot races were far too thin for that. The point is cultural. The marginal trader on an election contract in 2026 is no longer a partisan with a hunch. It is someone running a model, often with an antenna on the roof and a language model in the loop, and the wallet that lets them hold and move funds has quietly become part of the toolchain, the same shift we tracked in the move toward embedded, agent-ready wallets. The election odds are the output of that toolchain.

Smarter, or just faster and more alike?

So does the machine takeover make the markets better? The optimistic case is efficient aggregation: thousands of agents, each pulling in a slightly different slice of data, their errors cancelling, their prices converging on the truth faster than any committee could. The Senate shift is exhibit A for that story. When the autumn map moved, the order books did not wait for a pundit to change his mind; they re-priced, carrying Senate control from a Republican lean to a roughly three-in-four Democratic call, and that call proved correct.

The pessimistic case is monoculture. If most of the volume is coming from agents trained on the same public data, wired to the same polling feeds and running similar models, their errors do not cancel, they compound. A market can be fast and confident and wrong all at once, because every participant is wrong in the same direction. That is exactly what the Senate misses looked like in 2026. The upsets did not scatter; they clustered, every one of them in the direction of a Democratic wave the machines collectively underweighted. Correlated models produce correlated blind spots, and the more the order book is automated, the more the whole market can share a single mistake.

Liquidity made it worse where it mattered least to the bots and most to the public. The marquee questions, House and Senate control, were deep and efficient. But plenty of the individual battleground races that actually decide control traded on less than a hundred thousand dollars, thin enough that a few orders could move the line and too thin for the arbitrage that keeps deeper markets honest. That is the structural tension crypto has been circling for years, the gap between a headline price that looks precise and the shallow book underneath it, and it is the same question we have asked of automated market-maker design when the curve has to quote a real asset. On the big questions the machines were a genuine upgrade. On the long tail, a confident-looking number sat on top of a book too shallow to trust.

The resolution test: objective outcome, subjective timing

Every prediction market lives or dies on resolution, the step that decides who gets paid, and election markets are unusually lucky here. The outcome of a vote is about as objective as an event gets, which is why the 2026 races resolved cleanly where messier markets, on everything from corporate filings to geopolitics, routinely descend into disputes. Kalshi settles centrally against a named source, the official result or a wire service’s call, with its own review committee as the backstop. Polymarket settles through UMA’s optimistic oracle, where a proposer posts the answer with a bond and anyone can challenge it into a token-holder vote. On clean yes-or-no questions with an obvious source, both work well.

The subtlety is not what happened but when. The entire speed story rests on a choice about when a race is officially over, and that choice is not as objective as the outcome. Is a seat decided when the Associated Press calls it, when the losing candidate concedes, when the secretary of state certifies, or when the market’s chosen source says so? Different venues can resolve the same race minutes or hours apart depending on which clock they read, and the thirty-one-minutes-ahead bragging rights depend entirely on that definition. It is a small thing in a landslide and a lawsuit waiting to happen in a squeaker, which is why who controls the settlement rule, and who can be held to account for it, is the question that keeps resurfacing. It is the same accountability gap we examined in who actually answers for a DAO: when an autonomous system decides who gets paid, the law still wants a person at the other end.

Resolution is also, quietly, the next thing to be automated. UMA has been testing an AI proposer that reads the evidence and posts an answer for humans to challenge, with promising but still imperfect results in trials, and researchers have floated locking a specific model and a specific prompt into a market at the moment it is created, so the resolution rule is fixed and auditable before anyone places a bet. For a clean election night that is overkill. For the contested, open-ended markets that make up the long tail, it is where the real design fight is heading, and it means the machines are moving from one side of the trade to both: betting on the outcome and, increasingly, deciding it.

Integrity in the most-surveilled market

A market where the trading is radically transparent, every wallet and every position visible on-chain, ought to be the hardest place in finance to cheat. The 2026 cycle showed that transparency catches cheaters without stopping them. The clearest cases were the most obvious kind of conflict: candidates trading their own races. Kalshi suspended at least six 2026 candidates for betting on their own contests in breach of platform rules, banning an earlier batch of three, named as Ezekiel Enriquez, Matt Klein and Mark Moran, for five years in what the company called political insider trading (CNN). The US Senate, for its part, barred senators and their staff from trading on these markets at all.

More corrosive are the cases where the trader has an information edge that is not obviously illegal. Analysts flagged a Kalshi account linked to a pollster that traded three races hours before that pollster’s former firm released surprising numbers on those same contests. Because the chain is public, surveillance tools and journalists can reconstruct these patterns after the fact, but reconstruction is not deterrence, and the line between a sharp read and misappropriated information is exactly where the law is murkiest. Michelle Kendler-Kretsch, a researcher at the Anti-Corruption Data Collective, put the warning plainly after the vote: the “risks and vulnerabilities didn’t change overnight just because the markets were accurate this time around” (CNBC). A clean scorecard is not a clean bill of health.

The agent angle sharpens the point. When a profitable wallet is an autonomous program, the question of who knew what, and who is responsible, gets harder, not easier. An agent that trades on a leaked internal poll does not have intent the way a human does, but someone deployed it, funded it and pointed it at the market. The transparency that makes prediction markets auditable is the same property that let the bots dominate them, and it does nothing on its own to answer the accountability question.

Washington’s verdict is still pending

The one verdict the markets cannot price is their own legality, and here 2026 ended in limbo. The dispute is jurisdictional and, for crypto readers, usefully clarifying: event contracts are the domain of the Commodity Futures Trading Commission, not the Securities and Exchange Commission. The SEC enters the story only at the edges, over the tokens that some of this infrastructure runs on, such as UMA, and over any stock or security an agent might buy on the back of a market signal. The contracts themselves, election markets included, are the CFTC’s to police.

The CFTC spent the autumn trying to draw that perimeter. On 9 October it proposed to define the word swap to include event contracts based on sports, politics, culture and weather, a move that would plant election markets firmly on federal ground, and it has sued nine states to defend what it regards as its exclusive authority (CoinCentral). The states are not going quietly. Three federal appeals courts have now split on whether these contracts are swaps at all, with the Third Circuit siding with the platforms and the Ninth and Sixth Circuits siding with state regulators, a three-way conflict that all but guarantees the Supreme Court will have to settle it (CNBC). Kalshi’s response to the pending petition fell due in the week after the election.

For the machines, the stakes are concrete. If event contracts are swaps, they inherit the Commodity Exchange Act’s anti-fraud and anti-manipulation rules, the same misappropriation framework that governs insider trading in other derivatives, which would land squarely on the bots and the desks running them. A nationwide federal rulebook would also be the precondition for agents to scale, because software cannot geofence itself around a fifty-state patchwork the way a human can simply be told which states to avoid. The clarity that institutions want is the same clarity the autonomous traders need, which is why the people building America’s finished stablecoin rulebook and the people building prediction-market agents are, improbably, rooting for the same outcome.

The money: billions, and a 2024 shadow

By any historical standard the numbers are enormous. By the Monday before the vote the Washington Post counted nearly 1.9 billion dollars wagered on midterm contracts across the two venues, roughly 1.3 billion on Kalshi and 600 million on Polymarket (Washington Post). Kalshi’s own recap put its full-cycle midterm volume near 2.5 billion dollars, roughly double what it saw in the 2024 cycle (RotoWire).

And yet the midterms still sat in the shadow of a presidential year. The single Polymarket market on whether Donald Trump or Kamala Harris would win in 2024 drew more than 3.6 billion dollars on its own. The headline growth is real, but the money is concentrated at the top of the ballot, and the long tail of House races stays thin, which is precisely where the odds were least trustworthy.

The money has also drawn the kind of backer that does not usually bet on elections. The parent company of the New York Stock Exchange has poured a sum reported in the billions into Polymarket, pitched not as a wager but as an investment in real-time data and settlement infrastructure. That framing matters, because it treats the live probability feed, the thing the agents produce, as the product. It is also the tension the whole sector now lives inside. The platforms are valued in the tens of billions; the governance tokens that secure some of their plumbing, such as UMA and the Olas token behind Polystrat, remain thinly traded micro-caps worth a tiny fraction of the venues that use them. The value is in the rails and the data, not the ticker, a gap this cluster keeps running into whenever framework quality and token price part ways.

What the machine midterms taught us

So, the grade. On the only question most people asked, who wins, the machines earned their passing mark: they called both chambers, got hundreds of down-ballot races right, and did it faster than anyone else. On the harder questions the report card is more mixed. They were well calibrated on the big contests and poorly calibrated on the Senate, where a correlated fleet of models underpriced the very wave the pollsters also missed. They were first but not clairvoyant, reacting to returns rather than seeing past them. And they were transparent without being clean, auditable after the fact but not immune to the oldest conflicts in betting.

What the 2026 midterms really demonstrated is what a prediction market is for in the machine age. It is not a crystal ball, and it never was. It is a fast, liquid, continuously updated consensus, now mostly produced by software, that other software can read and act on. The product is the probability feed, and election night was its loudest demo yet. The pessimist’s warning, that a market of correlated bots can be confidently wrong together, is the price of that speed, and it does not go away as the models converge; it gets quietly larger.

The open questions are the ones Washington will answer. If the Supreme Court and the CFTC hand the sector a single federal rulebook, the machines get the uniform, national surface they need to scale, and 2028 will run on markets that are deeper, more automated and more consequential still. If the states win, the patchwork that software cannot navigate will fence the bots back out of the biggest contests. Either way, the 2026 midterms were the moment the prediction markets stopped being a crypto curiosity and became part of how the country reads an election in real time. The machines passed. The harder test is whether we learn to read their grades correctly.

Frequently Asked Questions

Were prediction markets more accurate than polls in the 2026 midterms?

Not meaningfully. Kalshi and Polymarket both called Democratic control of the House and Senate, and favorites won about 98 percent of the races they tracked, but post-election reviews found the markets identified winners about as often as the polls and, like the polls, underestimated the size of the Democratic wave. Their real edge was speed, not foresight: the odds reached near-certainty hours before the networks called the chambers.

How much money was traded on the 2026 midterm prediction markets?

A lot. The Washington Post counted nearly 1.9 billion dollars wagered across Kalshi and Polymarket by the Monday before the vote, with Kalshi taking the larger share at roughly 1.3 billion dollars. Kalshi later put its full-cycle midterm volume near 2.5 billion dollars, about double its 2024 figure, though both remained below the 3.6 billion dollars that the single 2024 presidential market drew on Polymarket.

Do AI bots really dominate prediction-market trading?

Increasingly, yes. A review of Polymarket’s public leaderboard found that 14 of the 20 most profitable wallets were bots and that AI agents made up more than 30 percent of wallet activity, with a higher share of agents profitable than human traders. The flagship is Polystrat, an autonomous agent built on the Olas protocol that uses a language model to trade on plain-English goals around the clock.

Is it legal to trade election outcomes on Kalshi and Polymarket in the US?

It is contested. Event contracts, including election markets, fall under the Commodity Futures Trading Commission rather than the SEC, and the CFTC has moved to classify them as swaps under federal law. But three federal appeals courts have split on whether that classification holds, several states argue the contracts are gambling they can regulate, and the Supreme Court is likely to decide the question. Until it does, availability can vary by state.

Did anyone insider-trade the 2026 midterms on prediction markets?

There were clear cases and murkier ones. Kalshi suspended at least six 2026 candidates for betting on their own races, banning three of them for five years, and analysts flagged an account linked to a pollster that traded ahead of surprising poll releases. Because the markets settle on public blockchains, such patterns can be reconstructed after the fact, but detection is not the same as prevention.

By Marcus Okafor, senior editor at HOGE Wire, covering AI agents, crypto and the markets where they meet.