Over or Under? 21,863 Polymarket games and 47,121 large bets, checked against the result

The oldest rule in totals, take the under because the public loves the over, buys you nothing on Polymarket. Across 47,121 large buys, each side won about as often as its price said and lost less than a point to it.

What decides a total is its line. Each line settles at its own rate, and the price already charges for that rate. Judge a total by its line and its price, and ignore which side it is.

Not one total in the sample could push. Polymarket settled 83,465 total lines on 21,863 games with a kickoff from April 2 to September 13, 2026, and every one of the 82,182 lines with a recorded value carried a half point. That is the markets universe: every game total with outcomes Over and Under, counted whether or not anyone bought it large.

The buys universe is every buy 0xinsider recorded as a large trade on one of those markets, at 2 to 98 cents, on a market that settled Over or Under after the trade. The smallest recorded buy was about $3,000 until July 2, 2026 and $10,000 after. That is 47,121 buys from 3,041 wallets across 7,240 markets, $827.3 million in stakes, split almost evenly: Overs were 48.2% of the buys and 49.2% of the dollars.

Edge is the share of buys that won minus the average price paid, in percentage points; the dollar return is stake-weighted and held to settlement, before fees. Every interval resamples games, because every line on a game settles on the same final score, so the buys on one game are in effect one bet. A share for a single line, such as 2.5 goals, counts each game once and carries a plain binomial half-width.

The window starts on April 2 for the reason the calibration study documents.

On soccer's standard line, 2.5 goals, the Over hit: 52.8% of 9,731 matches went over, give or take a point. Move the line one goal and the answer flips. At 3.5 goals the Under hit 68.7% of the time, and at 4.5 goals 83.8%. Each line has its own rate, set by how many goals matches produce.

Average every line together and 52.6% of the 83,465 settled lines went Under, on an interval from 52.1% to 53.1% with games resampled. That average describes which lines Polymarket lists more than how games end. A soccer ladder from 0.5 to 5.5 goals leans Under because its high lines usually go Under, while the same matches went over the standard line more often than not.

Esports totals are mostly 2.5 games in a best-of-three series, and 56.4% of 6,962 series ended in two games. Baseball's lines went Under 49.7% over 2,149 games, on an interval from 48.2% to 51.2% that includes an even split, and the hockey and WNBA intervals include 50% as well. College football and the NBA leaned Over on fewer than 300 games each, with intervals that reach 50%. MMA round totals went Under 37.8% of their lines over 299 fights.

None of these rates is an edge by itself. Under 3.5 goals settled in about two matches out of three, and large Under 3.5 buyers paid 66.1 cents and won 65.3% of their buys. The share tells you how often a line lands. The price tells you whether anyone is paying too much for it.

Under buyers came in minus 0.21 points against their price and Over buyers minus 0.89, on game-resampled intervals from minus 3.46 to plus 3.15 and from minus 4.39 to plus 2.67. Losing less than a point is what crossing a one-cent spread and paying a fee looks like, and neither interval comes near excluding zero.

Unders cost about two cents more, 58.4 against 56.5, because of the lines buyers chose: an Under on a high line is an expensive bet that usually wins. Each side paid about what its lines were worth. There is no free Under on Polymarket, and there is no Over tax either.

The dollar returns split wider, 1.5% for the Under against -2.96% for the Over, because they weight each buy by its stake and a few large losing Overs move them. The point edge weights every buy the same, so read the edge as the finding.

Soccer, 26,533 of the buys, comes closest to the folklore. Soccer Over buyers paid 60.1 cents and won 57.5%, minus 2.64 points and minus 8.12% on the dollar, while soccer Under buyers came in plus 0.56. The Over's interval runs from minus 8.18 to plus 2.74: the lean points the way the folklore predicts, and the data cannot confirm it.

Baseball runs the other way: Overs plus 1.13 points and plus 7.22% on the dollar, Unders minus 2.25 and minus 5.09%, both intervals across zero. The NBA's Over came in at plus 2.73 points on 4,101 buys, interval minus 5.77 to plus 11.06. Hockey's 1,393 buys are too few to say anything: its intervals span about twenty points.

The intervals are wide because totals buys pile onto a few games. The 22,703 Over buys sit on 3,144 games, about seven large buys a game, and the big games carry dozens. Resampling games instead of buys gives the honest interval, and it says a two-point lean in soccer is inside the noise of one summer.

Over 3.5 goals is the number a trader remembers. Its buyers paid 51.2 cents for a bet that won 46.0% of the time: minus 5.14 points and minus 20.82% on the dollar across 2,223 buys. The interval runs from minus 15.98 to plus 5.48, so this window cannot say whether that is a mispricing or a bad summer for four-goal games. Under 3.5 buyers, at 66.1 cents, won 65.3%, on the price.

The standard line was closer to fair. Over 2.5 buyers paid 54.4 cents and won 51.8%, minus 2.66 points; Under 2.5 buyers paid 54.4 and won 56.2%, plus 1.81. At 1.5 goals, Over buyers paid 73.7 and won 71.7%, minus 1.98.

The thin lines came in as close as a few hundred buys allow: Over 0.5 at 88.1 cents won 85.0%, Under 4.5 at 78.1 won 76.3%, Under 5.5 at 86.3 won 86.3%. The market knows what a 4.5-goal game costs.

The middle of the board is where totals are priced best. 66% of the buys sat between 40 and 60 cents, and there both sides came within half a point of their price. Pay more than 60 cents for either side and the price was ahead of you, by the most between 60 and 80 cents.

Under 40 cents the edge and the dollar return come apart. On 2,313 cheap buys the Over came in at plus 1.30 points and the Under at minus 1.87, yet both lost money, minus 5.19% and minus 16.26% on the dollar, because the cheap side that lost was bought bigger than the cheap side that won. The favorite-longshot study found the same shape on every sports market.

None of these bands is resampled on its own, so read them as the shape of the board, not as four edges.

Before kickoff both sides sat near a coin flip, Overs at 53.9 cents and Unders at 54.3, and both lost under a point to the price, with intervals from about minus 5 to plus 3.

In-play the prices move away from the middle, because the clock runs and goals either come or do not. Live Unders were 58% of the in-play buys; they paid 64.4 cents and won 64.8%, plus 0.45. A live Under is a bet that the game stays where it is, and the market priced it to within half a point. Live Overs paid 62.1 and won 61.4%, minus 0.70.

The timing study found the in-play price fairer to the buyer than the pre-game price across every market type, and totals are no exception. What the phase split adds is where the Under lean lives: before kickoff the buys split 52 to 48 for the Over, and in-play 58 to 42 for the Under.

The best-graded wallets lean Under and the worst do not. From June 1, with each buy graded as its wallet stood when it traded, wallets graded S, A or B put 60.0% of their totals buys on the Under. Wallets graded D or F put 49.8% there, an even split. That preference is a count, not an estimate, and it is ten points wide.

On the Under, S, A and B wallets paid 63.2 cents and won 65.1%: +1.85 points and +5.31% on the dollar across 4,960 buys, interval -3.92 to +7.61. On the Over they came in at -2.95 points and -16.20% on the dollar, the dollar figure driven by a handful of large losing buys. D and F wallets lost on both sides, -3.43 points on Overs and -2.48 on Unders, with intervals across zero.

That fits two other studies: the sharp money study found totals the market type where the poorly graded cohort did worst, and the fade study found a one-sided crowd on totals trailing its price. The graded wallets' Under edge still has an interval that includes zero.

Corrected October 7, 2026. The first version graded each buy with a ranking row 0xinsider can rewrite after the trade, and reported S, A and B wallets at +2.12 points on the Under and D and F wallets at -4.10 on the Over. Graded at the trade, over the markets settled by October 7, 2026, the lean holds and the edges are smaller.

A real bias shows the same sign every month. In the four months with real size, April to July, both sides changed sign at least twice. August and September hold 1,525 buys between them, after the alert floor rose to $10,000.

The soccer Over's minus 2.64 for the window is mostly June, June is mostly the World Cup, and a tournament is one draw from the distribution of summers. Nothing in the months says the Over is systematically overpriced.

Total corners produced the biggest numbers in the study and the weakest evidence. Corner Unders came in plus 9.03 points against their price and corner Overs minus 7.53, on 368 and 355 buys across a few dozen markets, too few to resample. That is a number to re-check on a larger sample, not evidence yet.

First-half totals and soccer team totals came in within four points of their prices either way. Team totals returned plus 27% on the Over and minus 19% on the Under, swings a single game can produce. The study records all three families so a rerun has something to compare.

The first version of this page was surer than its data. Published September 13, 2026, it counted each of 82,080 settled total lines as an independent game, said the Under settled 52.7% of the time give or take 0.34 points, and read that as the folklore's premise confirmed. Most games list several lines, and every line on a game settles on the same final score, so that interval was too narrow and the share depended on which lines Polymarket listed.

The recount on September 14, 2026 keeps each line's game. The average over every line is 52.6%, on an interval from 52.1% to 53.1% with games resampled, and soccer's standard 2.5-goal line went over in 52.8% of 9,731 matches. The buy-level points did not move. Their intervals, which resampled markets, now resample games, and none of them excludes zero.

On October 10, 2026 the buy-level figures were re-run with void and unresolvable markets left out, where the first run had scored them as losses. That rule moved no side by more than a tenth of a point; the markets that settled since moved the rest.

There is no free Under on Polymarket. A trader who takes every Under at 58 cents is paying 58 cents for a 58% chance, and fading the Over does no better. Every blanket rule in the chart below has an interval that includes zero.

The edge on a total, where there is one, comes from the same places as on a moneyline. The counterparty: the fade study found a poorly graded crowd that agrees on a side of a total before kickoff trailing its price by 2.87 points, inside its margin of error. The timing: in-play totals came within a point of their prices on both sides. And the number itself: a trader who can say why a game produces four goals has a bet on Over 3.5 that this window priced at 51 cents for a 46% outcome.

The graded wallets' Under lean is the one habit in this study with some support. They put six of every ten totals buys on the Under and came in ahead on it, and behind on the Over. That is a lean, not a rule, and its interval says so.

It has no sportsbook in it. The claim that books shade totals toward the Over is about books. What this data can say is that on a market with no book, each line settles at its own rate and the price reflects it.

The sample thins out at the end. The alert floor rose from about $3,000 to $10,000 on July 2, 2026, so August and September record only the largest bets, and the monthly picture is four months deep.

The buy-level intervals are wide because totals buys cluster on popular games. A two-point lean a trader might notice over a season is inside them. A study of every totals trade, not just the large buys, would narrow them and keep the clustering.

The settled share is a fact about lines and results, not about traders; how Polymarket sets its lines is outside this data. And every position is scored as held to settlement at the buyer's price, before fees: a wallet that sold a losing Over at 30 cents counts as if it held to zero.

On Polymarket there is no line to shade, only a price. A game total is a market with its line fixed inside it, always a half point, and two outcomes, Over and Under, priced in cents that sum to about a dollar. Buying Over at 56 cents pays a dollar a share if the game clears the number.

Most games list several of these markets at different lines, 3.82 a game in this study, each priced on its own. So the question splits in two: how often does the Under settle at each line, and did the price know?

At a sportsbook the number moves and the price stays near -110 on both sides, and the folklore is old: the public bets the over because rooting for points is more fun, the book shades the total up to meet them, and the patient money takes the under. There are studies of it and counter-studies, and all of them use sportsbook data.

0xinsider holds both halves of the Polymarket version: every settled totals market since April 2, 2026 with its result, and every large buy on one with its price and its buyer.

With Pro, a game's event page lists every market on the game by family, totals included, with the Over and Under prices in cents; open one from any sports board, such as the soccer or MLB board. Each total line carries a Profitable wallets row: how the graded money splits between Over and Under, and the top graded wallets on each side. Without Pro the Markets section shows sample data. The over/under guide covers the bet itself.

The large-trade feed carried 1,059 trades on game totals from September 11 to October 10, 2026, each marked BUY or SELL with the wallet's grade beside it, live on Pro and 24 hours behind without it. A run of D and F buys on one side of a total before kickoff is the pattern the fade study measured.

We re-ran the soccer lines on the 1,986 matches that kicked off from September 14 to October 9, 2026, after this study's window closed. The 2.5-goal line went over 55.5% of the time, so the finding that the standard line goes over held and got stronger. The study's interval for that line did not hold: 44.5% Under sits below its 47.2% give or take 0.99.

The new matches are one month of autumn league football; the study's window ran through a summer of internationals and the World Cup. One month cannot say which is the normal rate, so treat any soccer line's share as a season's number, not a constant.

The buy-level figures and their game-resampled intervals come from one run of read-only queries on a copy of the production database at 12:48 UTC on October 10, 2026, with void and unresolvable markets left out; the settled shares come from the recount at 01:11 UTC on September 14, 2026. Their query text, raw output, exports and bootstrap output are public at github.com/0xinsider/research, under polymarket-sports-markets, in the files whose names carry the words over-under, and every interval reproduces from the exports with no database.

The buys universe is the calibration study's universe restricted to markets of type totals with Over and Under outcomes. The markets universe is every such market with a kickoff in the window and a settled outcome, and a market's game is its Polymarket event. Outcome 0 is Over and outcome 1 is Under on these markets; the settled share is the share with outcome 1.

The grade split is a later run: on October 7, 2026 it was re-measured with each buy graded as the wallet stood at the trade (the latest daily ranking dated on or before the trade day and last written at or before it), in the files whose names start grade-at-trade-2026-10-07 in the same folder, with intervals that group buys by market, by wallet and by game and show the widest.