Favorite-longshot bias on Polymarket sports markets: 416,548 large bets, priced against the result

Make a large buy of a Polymarket sports contract and, on average, you pay what the result is worth. Across 416,548 such buys from April 2 to September 11, 2026, buyers paid 60.6 cents on average and won 60.9% of the time.

The exception is the cheap end. Under 10 cents, a dollar staked came back as 44 cents; from 10 cents up, as about a dollar (the chart below). That is the favorite-longshot bias every sportsbook dataset shows, and here it is a lean: the interval on its edge runs from minus 5.22 to plus 0.81 points and reaches zero. Treat a price under 10 cents as the one Polymarket sports price to distrust by default.

This study first ran on September 12. It scored every buy on a market that settled void, or that Polymarket could not resolve, as a loss: 1,625 buys in this window, most of them tennis and esports. A void market pays half a dollar a share on either side, so those buys were neither wins nor losses.

Re-run on October 10, 2026 with those markets left out, every figure moved up (the chart below). The whole sample moves from plus 0.02 points to plus 0.23. Under 10 cents the edge moves from minus 3.52 points, with an interval that cleared zero, to minus 2.97, with one that does not. MMA moves from minus 5.34% on the dollar to minus 1.99%. The rest of Polymarket moved most: buyers of non-sports favorites no longer miss by six to ten points.

The September query run on the same data gives plus 0.01 points on 418,173 buys, so the change comes from the void markets, not from markets that settled since. Everything after this section is the corrected study.

The sample is every buy 0xinsider recorded as a large trade on a Polymarket sports market from April 2 to September 11, 2026, on a market that settled yes or no after the trade cleared, at a price between 2 and 98 cents. The smallest recorded buy was about $3,000 until July 2, 2026 and $10,000 after. That is 416,548 buys by 16,190 wallets across 34,293 settled markets, carrying $8.25 billion in stakes. Soccer, esports, the NBA, tennis, baseball, hockey, cricket, MMA, the WNBA, and golf are all in it.

Most of it comes from April to June (the chart below): April alone is 44% of the buys, and the three months together 89%. August and September are thin because the alert floor rose to $10,000 on July 2, 2026.

Buys only, because a buy at a price on a named outcome is a clean bet on that outcome. A sell can be a hedge, an exit or a market-making fill, and its price says nothing about what the seller believed.

Each buy is scored two ways. The edge is the share of buys that won minus the average price paid, in percentage points: buy at 60 cents, win 61% of the time, and your edge is plus one point. The dollar return is what a dollar staked came back as, weighted by stake, held to settlement, before fees.

The two can disagree. Ten-cent contracts that pay out 8% of the time lose 20% of an equal stake on an edge of only minus two points, and unequal stakes pull a bucket's dollar return further from its win rate. That is why both numbers appear below.

Confidence intervals resample markets rather than trades. Fifty buys of the same side of the same game are one bet on that game, and treating them as fifty is how a study manufactures precision it does not have. 5,000 resamples, seed 20260912.

Across the whole sample the average price paid was 60.6 cents and the share of buys that won was 60.9%. The edge is +0.23 points. The dollar return, before fees, is +0.88%. On average, a large buyer on a Polymarket sports market paid what the outcome was worth.

Split into ten price buckets, every bucket from 20 cents to 98 cents sits within two points of zero, and every one of those intervals straddles zero. Buyers at 40 to 50 cents paid 45.4 cents and won 46.1%. Buyers at 50 to 60 cents paid 54.1 and won 53.8%. Buyers at 80 to 90 cents paid 84.7 and won 84.3%. Buyers at 90 to 98 cents paid 93.5 and won 93.0%.

The cheapest bucket is the one that leans. Buys under 10 cents paid 6.50 cents on average and won 3.53% of the time, an edge of minus 2.97 points with a 95% interval from minus 5.22 to plus 0.81. With void markets counted as losses, the September run put that interval below zero; it now reaches zero. The textbook longshot penalty shows on Polymarket only under 10 cents, and there it is a lean.

In five-cent steps the picture is sharper. Buys between 2 and 5 cents paid 3.19 cents on average and won 1.35% of the time; weighted by stake, a dollar came back as 71 cents. Buys between 5 and 10 cents paid 7.62 cents and won 4.27%; weighted by stake, a dollar came back as 37 cents. Those 2,040 buys put $28.3 million into contracts and got $12.5 million back, less than half.

This is the classic shape. A 7-cent contract is a bet on something the market says happens one time in fourteen. The people making large buys of it are paying for a fourteen-to-one payout, and in this sample they got a payout that arrived one time in 23. Polymarket has no bookmaker to shade that price, so the overpaying comes from the buyers.

Two things about the size of the effect. It is concentrated: those buys are 0.49% of the sample by count and 0.34% by dollars. And it is large: minus 56% on the dollar is the worst return of the ten price buckets, by a wide margin. On Polymarket sports markets, a large buy under 10 cents has been a bad bet.

Buys between 10 and 20 cents show a dollar return of plus 28.8% and an edge of plus 4.94 points. Read alone, that says the second-cheapest bucket was the best bet on the platform. It was not. The confidence interval runs from minus 4.98 to plus 19.73 points, the widest of any bucket, and the reason is a single market.

One market, 'Will Spain win the 2026 FIFA World Cup?', drew 490 large buys at an average of 15.2 cents, $7.48 million in all. Spain won. Those buys paid out $41.31 million in profit. The whole 10-to-20-cent bucket, all 5,962 buys across 1,577 markets, shows a combined profit of $27.17 million. The other 1,576 markets in the bucket lost $14.14 million between them.

This is what market-clustered intervals are for. A study that treated those buys as independent observations would report a significant longshot edge that is one tournament. Read correctly, the 10-to-20-cent bucket is indistinguishable from zero, and the lean below 10 cents does not obviously extend above it.

The fee comparison re-runs the fixed trade window on October 21, 2026, including eligible fills and settlements recorded since the original study. It sums a taker fee at each fill's price, at the rate that buy's market carries on Polymarket's schedule read October 21, 2026, then divides by total stake. This uses a stake-weighted price, so a large buy contributes more to the fee basis than a small buy. The data does not identify which buys took liquidity; the after-fee chart assumes every buy paid the taker fee. The fills record no fee, and fills before July 10, 2026 paid the 0.03 sports rate of the time, so this is today's schedule applied to the window, not what each buyer paid.

The textbook says favorites are underbet. At a sportsbook a -400 favorite still loses money over time, because the house margin is built into the line, but it loses less than the longshot on the other side. The favorite is the better of two bad bets.

On Polymarket it is not a bad bet at all, before fees. Buyers at 80 to 98 cents came out within half a point of their price, and buyers at 70 to 80 cents came out plus 0.98 points and plus 3.64% on the dollar ahead, across 49,786 buys.

Fees cost a known amount, and only if you take liquidity. On most sports Polymarket's taker fee is 0.05 times the share count times the price times one minus the price: 0.50% of the stake at 90 cents, 0.25% at 95. NFL and college football games carry 0.03 and NBA games no fee, as of October 21, 2026. A resting limit order pays nothing. Charged as a taker at each market's own rate, the 70-to-80-cent favorites stay ahead, the 80-to-90-cent ones keep +0.09% and the 90-to-98-cent ones slip to -0.83% (the chart below). Either way it is a long way from a sportsbook, where the same favorite carries the full hold.

Until October 21, 2026 this comparison charged every buy 0.05, including the NBA games that carry no fee today and hold 16.7% of the window's stake. It showed the 80-to-90-cent favorites at -0.27% after the fee. At each market's own rate, with the fills recorded since, they come out +0.09%.

Run the same ten buckets over every large buy on politics, crypto, geopolitics, culture, finance, world and business markets in the same window and the picture is less even. Those 71,646 buys miss the price by three to eight points from 20 to 50 cents, against the buyer, and beat it by four to five points from 60 to 80 cents.

Buyers at 20 to 30 cents on those markets paid 25.2 cents and won 17.6% of the time, minus 7.5 points. Buyers at 40 to 50 cents paid 45.5 and won 42.4%, minus 3.0. Buyers at 60 to 70 cents paid 65.1 and won 68.9%, plus 3.8, and at 70 to 80 cents plus 4.8. Buyers at 90 to 98 cents paid 94.1 and won 92.2%, minus 1.9, across 24,131 buys.

That is not a statement about which markets are mispriced in the abstract. Non-sports markets settle over months, the sample is thinner, and a run of surprising political outcomes in one summer can move every bucket at once. It is a statement about where a large buyer's price has matched the result. In sports it has, within two points in every bucket from 20 cents up. Elsewhere on the platform, in this window, it has missed by several points in both directions.

Soccer is the largest sample, 158,865 buys across 11,813 settled markets and $3.06 billion, and it is calibrated: buyers paid 60.9 cents and won 61.6%, plus 0.68 points and plus 1.43% on the dollar. Esports, 70,494 buys, came in at plus 0.39 points. The NBA, 63,037 buys across 1,841 games, minus 0.26. Tennis plus 0.21, baseball minus 0.47, hockey minus 0.57, all on tens of thousands of buys.

Cricket is small and positive, plus 2.21 points over 6,022 buys, most of them in-play. Basketball outside the NBA and WNBA is 1,968 buys at plus 0.78 points. The WNBA is 1,484 buys at minus 0.46. Golf, 1,092 buys, came in at plus 4.79 points but minus 1.65% on the dollar, because its winners were cheap favorites and its losers expensive ones.

MMA is the outlier. Across 5,197 buys, buyers paid 58.1 cents and won 53.8%, minus 4.33 points and minus 1.99% on the dollar. The six largest sports all sit within seven-tenths of a point of the price.

Split every sport into buys under 30 cents, between 30 and 70, and at 70 and above, and the favorite band is where the sports differ. Soccer favorites at 70 cents and up paid 83.8 cents and won 83.9%. Tennis favorites paid 84.1 and won 85.6%. Esports favorites paid 81.7 and won 82.5%. NBA favorites paid 84.3 and won 82.7%. Baseball favorites paid 83.3 and won 81.2%, the second-worst band.

MMA favorites paid 80.4 cents and won 61.6% of the time. That is minus 18.7 points and minus 16.8% on the dollar across 1,478 buys. Two fights carry most of it. Khamzat Chimaev went off at 81 cents against Sean Strickland at UFC 328 and lost, with 222 large buys and $4.29 million on him. Ilia Topuria went off at 80 cents against Justin Gaethje at UFC Freedom 250 and lost, with 148 large buys and $1.57 million.

MMA has 333 settled markets in this sample against 11,813 for soccer, so one card moves the whole sport. Read the MMA number as a warning about small-sample sports rather than a finding that fight favorites are systematically overpriced. What it does establish is that calibration is not automatic on Polymarket: it holds in deep, frequently traded markets like soccer, tennis, and basketball, and a fight card is neither.

Polymarket tags each sports market with a type, and the price held on every one of them. Moneylines carry most of the money, 271,155 buys and $5.60 billion, at 61.2 cents paid and 61.4% won. Spreads did best, plus 0.64 points, and totals worst, minus 0.54; the table below has every type. Child moneylines, which cover a single map, set or period, came in at plus 0.71.

The soccer specials behave the same way. Team-to-advance markets, 4,828 buys, plus 0.05 points. Both-teams-to-score, 3,682 buys, plus 0.87. Exact-score markets are bought almost entirely as favorites, at 86.7 cents on average, and they won 86.7% of the time.

Split each type at 30 and 70 cents and the tails look like the whole: every type's longshots and favorites landed within 3 points of their price (the chart below). Total longshots fell furthest behind, minus 2.63 points on just 690 buys. The one corner that leans your way is spread favorites, plus 2.42 points on 5,079 buys. This run put no interval on it, so treat it as a lead to test, not an edge.

On a Polymarket sports market the price is a fair estimate of the outcome. No bookmaker folds a margin into it, and outside the cheapest end the crowd has not pushed it off its value. With no margin to beat and no systematic mispricing to lean on, your edge has to come from knowing something the price does not yet reflect.

One place to look is who holds each side. 0xinsider grades every Polymarket wallet from S to F on its settled P&L, and the sharp money study found that the grade told sports buyers apart only in play: S, A and B wallets beat D and F by 2.45 points after kickoff and by nothing measurable before it (the chart below).

Each sports board shows which side the profitable wallets hold on each game, live on Pro and delayed or locked without it, and Pick of the Day publishes up to 25 of those sides a day on a public record.

And the one rule this study supports on its own: do not buy a 7-cent contract because it pays fourteen to one. In this sample, contracts at 5 to 10 cents paid out one time in 23.

The edge and the dollar return are measured at each buy's own price against the settled outcome. They ignore fees and slippage, and they treat every position as held to settlement. A wallet that bought at 60 cents and sold at 75 before the game ended is scored on the final result, which is not what it made.

The sample is large buys 0xinsider recorded as alerts, not every fill on the platform. On sports markets the smallest recorded buy was about $3,000 until July 2, 2026 and $10,000 after, so the late months skew to bigger bets. Across every category, nine in ten recorded buys were at least $3,521 in April and at least $10,758 in August (the chart below). Large buyers are not a random sample of traders, and a market can be calibrated for them while retail buyers of $20 contracts pay a different price.

Only markets that settled yes or no are in the sample. A void market, or one Polymarket could not resolve, is left out: the stake came back or has not been paid, so it is neither a win nor a loss. Pre-kickoff and in-play buys are mixed together here; the timing study splits them, and the split matters for who is on each side.

And a calibrated market is not a market nobody beats. Prices that match results on average mean the winners and losers among large buyers net to zero, not that there are no winners. The sharp money study takes the same sports buys, from June onward, and splits them by the grade each wallet held when it traded. The grade separates the winners from the losers only on buys made in play.

The favorite-longshot bias is the oldest result in betting research. Richard Griffith found it at American racetracks in 1949: horses at long odds were bet more heavily than their win rate justified, and favorites less. Every large sportsbook dataset published since has shown the same shape. A bet at +900 returns less per dollar than a bet at -300, and both lose to the house, but the longshot loses far more.

The usual explanation has two parts. Bettors overestimate small probabilities and enjoy the size of a possible payout, so they overpay for it. And a bookmaker who knows that shades the longshot price further, because the demand is there. The bias is a fact about the customers and about the margin built on top of them.

Polymarket sports markets have no bookmaker. A contract on a team pays $1 if the team wins and nothing if it does not, and the price is whatever the last buyer and seller agreed on. If the bias comes from the buyers, it should show up here too. If it depends on a bookmaker shading the line, it should not.

Polymarket does take a cut from takers, and its shape favors the favorite. On most sports the taker fee is 4.7% of the stake at the 6.5-cent average paid under 10 cents and 0.3% at 93.5 cents (the chart below). Every number in this study is before that fee, so any bias it finds belongs to the buyers.

The page draws on dated, read-only runs against a copy of the production database, all public in the 0xinsider research repository under polymarket-sports-markets. The sample, the buckets and their intervals come from the run at 12:24 UTC on October 10, 2026: calibration-2026-10-10.sql, its output and its market-level export, from which bootstrap.py recomputes the intervals without a database. The stored-outcome defect by month and the fee schedule come from the run at 12:36 UTC on October 11, 2026, in calibration-charts-2026-10-11.sql and its output. The smallest recorded buy comes from the read at 14:12 UTC on October 11, 2026, in large-trade-floor-2026-10-11.sql and its output. The split by grade is the sharp money study's October 7, 2026 run, in the files whose names start grade-at-trade-2026-10-07. The September 12 run, calibration.sql and its output, stays beside them as the record of the first version.

The universe is every row of the large-trade alert table for a Polymarket buy on a sports-category market, joined to the settled outcome of its market where that outcome is a yes or a no, with the resolution timestamp after the trade timestamp and the price between 2 and 98 cents. The edge is the mean of a won-or-lost indicator minus the mean price. The dollar return is the stake-weighted sum of won divided by price minus one.

Do not start your window before April 2, 2026. Before that day the alert table stored a defaulted outcome zero for a large share of buys, and the favorites it stored that way "lost" at a rate no market produces: at 80 to 98 cents, they won 57.8% in February and 51.8% in March, against 89.1% and 88.7% for the same prices stored as outcome one. In April both win within a point of their price (the chart below).

Read February and March as stored and those outcome-zero favorites miss their price by 30 to 37 points: a reverse favorite-longshot bias that is a data defect, not a market. An early pass of this analysis fell into exactly that.

Expect your counts to come out higher than these. Every buy in the sample sits on a market that has settled, so the universe grows as open markets resolve. The bucket-by-bucket picture should not move; if a rerun moves a bucket by more than its interval, that is a finding, not a reproduction.