On Polymarket every trade is public and every contract ends worth exactly $1 or nothing. That makes it possible to answer questions about traders that stock-market data leave open. Does luck decide who keeps trading? Does money drift toward the skilled? Do traders sell winners too early and hold losers too long, and does it cost them? This page follows 2.1 million wallets across 74,000 resolved markets. It is the fourth in a series, after how much prices know, who wins and the Israeli election markets.
Whether a trader keeps going after a win or a loss mixes two things: winners may be better traders, and better traders stay. To separate them, use bets whose outcome was close to random. A market priced between 40¢ and 60¢ a day before it resolved was nearly a coin flip as far as the market could tell. The wallets that held one side won or lost largely by luck. For each wallet, take its first such market and follow how often it trades in each week, from 8 weeks before to 12 weeks after.
Before the event the two groups are indistinguishable week by week: the difference never exceeds one percentage point. There are more losers than winners, because more wallets held the side that lost, and the losers were somewhat smaller wallets (median buying over the previous 8 weeks of $401 against $594). But they traded at the same rate. After the event the paths split at once and stay apart for the whole 12 weeks. The winners trade about 3.7 points more often each week, roughly a fifth more than the losers, and fewer of them quit altogether. The gap is there too when the stake was under a quarter of the wallet's recent buying (4 points a week on a higher base), so it is not simply that winners had more money to trade with. Traders respond to outcomes they could not control, much like investors who keep buying IPOs after lucky first experiences (Kaustia and Knüpfer, 2008).
4,448 resolved markets whose last traded price a day before resolution was between 40¢ and 60¢. A wallet enters at its first such market if it held at least 100 net shares of one outcome there and bought anything in the 8 weeks before (52,510 winners and 67,958 losers). "Trading" = any purchase in the week, in any resolved market. Shaded band: 95% interval from 300 bootstrap draws of wallets.
| Winners / losers | Weeks before: winners minus losers | 12 weeks after | Losers' weekly trading rate after | Never traded again (12 weeks): winners / losers | |
|---|---|---|---|---|---|
| All | 52,510 / 67,958 | +0.1 pts | +3.7 pts | 18% | 32% / 37% |
| Stake under 25% of the previous 8 weeks' buying | 21,276 / 23,686 | -1.0 pts | +4.0 pts | 28% | 21% / 24% |
| Stake 25% or more | 31,234 / 44,272 | -0.7 pts | +2.1 pts | 13% | 39% / 44% |
The formal version. The regressions follow each wallet week by week, from 8 weeks before its coin-flip market resolved to 11 weeks after. They include a fixed effect for each wallet, which absorbs how active it is in general, and one for each cohort-by-week cell. A cohort is the calendar week the coin-flip market resolved, so winners are compared with losers whose market resolved the same week, at the same distance from the event. The coefficient on winning in each week traces the effect over time. The pooled version compares weeks +1 to +11 with weeks −8 to −2. Week −1 often contains the purchase of the coin-flip position itself, and in week 0 other markets resolving the same day carry a mechanical share of the result, so both are left out.
The weeks before are flat: relative to their own average, the coefficients for weeks −8 to −2 stay within about half a percentage point. Week −1 dips, because wallets that bought late, in the final week, ended up on the losing side more often. Then activity jumps: winners are 5.5 points more likely than losers to trade in the week after the market resolves, and about 2.5 points more likely in each of the following ten weeks. Pooled over weeks +1 to +11, winning raises the chance of trading in a week by 3.0 points against a loser rate of 18%. It raises dollars bought by about 0.2 log points ($105 a week against $261), dollars sold by 0.12 and the number of trades by 7%. The narrower 45–55¢ window gives larger effects (4.2 points, 0.34 log points) and the wider 30–70¢ window slightly smaller ones. Comparing only wallets that bet on the same side, Yes or No, changes nothing.
Two things do not change. Winners' profits on markets in other events are no higher in weeks +1 to +11: +$0.36 a week, not distinguishable from zero (in the 45–55¢ window, whose wallets are larger and profitable on average, +$5.70, about what trading more at the same return would give). All of the apparent profit gain in the 40–60¢ window sits in week 0, when related events resolve together. And their buying shifts only slightly toward favourites: the long-shot share of their purchases is about 1.4 points lower than the losers', and the average price they pay barely moves. Luck changes how much people trade, not how well.
Coefficients bk from Yik = ai + dc(i),k + Σk bk·Wini·1[week = k], wallet i, event week k, cohort c, estimated with week −1 as the omitted week and shown relative to the average of weeks −8 to −2. Cohort = the calendar week the coin-flip market resolved, so winners are compared with losers whose market resolved the same week. Bars: 95% intervals, standard errors clustered by wallet. Week −1 often contains the purchase of the coin-flip position itself. Coin-flip markets priced 40–60¢ a day before resolution. Long-shot share and average price are measured in weeks with purchases only.
| Outcome (weekly) | 40–60¢ | 45–55¢ | 30–70¢ | 40–60¢, same side held | Losers' mean (40–60¢) |
|---|---|---|---|---|---|
| Trades at all (pts) | +2.97(0.16) | +4.19(0.33) | +2.74(0.14) | +3.1(0.17) | 18.1 |
| log(1 + $ bought) | +0.2(0.01) | +0.336(0.023) | +0.184(0.0089) | +0.196(0.011) | 0.919 |
| log(1 + trades) | +0.072(0.005) | +0.134(0.012) | +0.06(0.0043) | +0.0669(0.0052) | 0.427 |
| log(1 + $ sold) | +0.117(0.0083) | +0.215(0.019) | +0.104(0.0072) | +0.103(0.0085) | 0.608 |
| Profit on other markets ($) | +0.362(0.26) | +5.7(1.7) | +0.229(0.17) | +0.563(0.27) | -3.04 |
| Long-shot share (pts) | -1.35(0.23) | -1.26(0.29) | -1.39(0.22) | -1.19(0.25) | 16.5 |
| Average price paid (¢) | +0.431(0.23) | +1.4(0.29) | +0.421(0.22) | +0.468(0.24) | 51.9 |
| $ bought | +105(8.3) | +496(48) | +75.8(5.4) | +92.8(8.5) | 261 |
b from Yik = ai + dc(i),k + b·Wini·Postk, comparing weeks +1 to +11 with weeks −8 to −2 (the purchase week −1 and the resolution week 0 left out); standard errors clustered by wallet in brackets. Columns are the price a day before resolution. "Same side held": cohort × week effects also by the outcome the wallet held, so winners are compared with losers who bet on the same side (Yes or No). Shares in percentage points, prices in cents, money in dollars.
| Price a day before | Markets | Experienced wallets | …share that won | First-bet wallets | …share that won | First-bet wallets with no purchase in the 12 weeks: winners / losers |
|---|---|---|---|---|---|---|
| 40–60¢ | 4,448 | 120,468 | 44% | 122,538 | 46% | 41% / 43% |
| 45–55¢ | 2,173 | 41,665 | 41% | 20,758 | 37% | 22% / 41% |
| 30–70¢ | 8,803 | 158,607 | 40% | 167,195 | 43% | 37% / 43% |
First bets. A wallet whose very first market was a coin flip has no history, so there are no "weeks before" to compare, but also no record of skill that could differ between winners and losers. Of 122,538 such wallets in the 40–60¢ window, those that happened to win trade in about 5 more weeks out of every 100 over the next 11 weeks than those that lost, from a loser rate of 10%. That is about half as often again, three times the effect for experienced wallets. They also buy about $10 a week more against $20. Winners had, on average, bet on the side priced a little higher (54¢ against 48¢), as favourites win slightly more often. Comparing only wallets that bet on the same side, and holding that side's price fixed, makes the effect larger still: 7 points. In the narrow 45–55¢ window first-time winners trade about 80% more often than first-time losers.
First-time losers also turn to long shots: in the weeks they trade, 3 to 5 points more of their buying goes into contracts under 10¢ than winners'. They do not win back their losses: neither group's later profits differ by much, and first-time winners, who trade more, lose slightly more in total ($0.09 a week). A newcomer's first result, which is mostly luck, shapes how much they trade for months.
Wallets whose first market was a coin flip (40–60¢ a day before resolution, at least 10 net shares). Each week's coefficient is from Yi = dt + bk·Wini on that week's observations, with cohort × week fixed effects; 95% intervals. Week 0, the week the market resolved, is shown but left out of the pooled estimates.
| Outcome (weekly) | 40–60¢ | 45–55¢ | 30–70¢ | 40–60¢, same side held | …and its price | Losers' mean (40–60¢) |
|---|---|---|---|---|---|---|
| Trades at all (pts) | +4.76(0.1) | +9.43(0.3) | +5.18(0.091) | +5.43(0.11) | +6.93(0.18) | 9.85 |
| log(1 + $ bought) | +0.2(0.0049) | +0.379(0.015) | +0.221(0.0046) | +0.218(0.0054) | +0.279(0.0093) | 0.397 |
| log(1 + trades) | +0.0781(0.0024) | +0.168(0.0073) | +0.0883(0.0022) | +0.0887(0.0026) | +0.114(0.0045) | 0.154 |
| log(1 + $ sold) | +0.101(0.0038) | +0.228(0.011) | +0.108(0.0034) | +0.122(0.0042) | +0.15(0.0066) | 0.172 |
| Profit on other markets ($) | -0.0947(0.019) | -0.127(0.11) | -0.103(0.021) | -0.0749(0.02) | -0.141(0.038) | -0.74 |
| Long-shot share (pts) | -2.9(0.23) | -5.24(0.5) | -3.75(0.2) | -2.86(0.26) | -3.35(0.34) | 16.8 |
| Average price paid (¢) | +2.75(0.31) | +8.43(0.59) | +5.32(0.26) | +4.22(0.34) | +5.7(0.4) | 51.7 |
| $ bought | +10.2(0.46) | +21.9(2.2) | +12.2(0.48) | +10.1(0.5) | +12.3(0.92) | 20.3 |
b from Yik = dc(i),k + b·Wini (+ the price of the side held, last column), weeks +1 to +11; standard errors clustered by wallet in brackets. No wallet fixed effects: these wallets have no history before their first bet.
If markets select for skill, money should drift from unskilled to skilled traders over time and prices should get sharper as it does. The first half of that happens on Polymarket, but on the wrong signal. Wallets that made money on the markets resolving in a quarter are more likely to trade in the next one. Among wallets that spent at least $10,000, 57% of winners trade again against 42% of losers. The winners' share of the money rises from 60% to 69%. Returns, however, barely persist from one period to the next (traders page), and the coin flips show that luck alone moves people this much.
Measured by record rather than by last quarter's luck, there is no drift toward skill. Each quarter, wallets with a statistically proven record supply 3–4% of the money, proven losers 4–10%, and wallets with little or no record between a quarter and two-fifths. Prices did get sharper over 2025–26 (Brier skill a day out rose from 0.44 to 0.63 between early 2025 and mid-2026), but not because skilled capital took over. And individual traders do not get better with experience: following the same wallets over their first 100 markets, their returns are flat.
| Return on markets resolved this quarter | Wallet-quarters | Trade next quarter | Median next/this quarter spend, if they trade | Share of this quarter's dollars | Share of next quarter's |
|---|---|---|---|---|---|
| Lost over half | 131,781 | 26% | 0.80× | 3% | 3% |
| Lost 10–50% | 120,212 | 41% | 0.34× | 6% | 7% |
| Lost under 10% | 974,691 | 27% | 0.25× | 31% | 21% |
| Gained under 10% | 361,134 | 39% | 0.41× | 50% | 57% |
| Gained 10–50% | 99,977 | 45% | 0.50× | 7% | 10% |
| Gained over half | 47,639 | 37% | 0.73× | 3% | 2% |
Records use only markets resolved before the quarter began. Proven forecaster: profit t-statistic above 2 over at least 10 markets; proven loser: below −2; market maker: at least 20 markets, under 20% of fills taking liquidity; new or short record: fewer than 3 resolved markets. Quarters from mid-2024 to the third quarter of 2026.
| Markets traded so far | Markets | Return per market vs. the wallet's own average | Return per dollar |
|---|---|---|---|
| 1–5 | 65,270 | +0.51 pts(+0.11 to +0.90) | +0.8%(-4.3 to +6.7) |
| 6–10 | 65,270 | +0.06 pts(-0.29 to +0.44) | +0.6%(-3.9 to +4.6) |
| 11–25 | 195,810 | +0.28 pts(+0.04 to +0.52) | -1.7%(-4.9 to +1.5) |
| 26–50 | 326,350 | +0.32 pts(+0.17 to +0.48) | -1.1%(-5.7 to +1.7) |
| 51–100 | 652,700 | -0.30 pts(-0.41 to -0.21) | +2.9%(+0.9 to +5.5) |
13,054 wallets that traded at least 100 resolved markets (with at least $10 spent), followed over their first 100. The middle column subtracts each wallet's own average return, so it compares the same wallets early and late; returns per market capped at −100% and +300%.
The disposition effect, investors' readiness to sell winners and reluctance to sell losers, is one of the best-documented biases in finance (Shefrin and Statman, 1985; Odean, 1998). Measuring it needs each investor's purchase price and, to say whether it costs anything, what the shares were eventually worth. Polymarket gives both exactly. Every order by a wallet that already holds the outcome it trades is a moment at which it could sell. The position is at a gain or a loss by comparing the price with the wallet's average purchase price.
The effect is clear: wallets sell at 50% of the moments when they are ahead and 40% when they are behind. It is weak among the smallest wallets and about 15 points among those that traded $10k–1m, and 58% of wallets with enough events show it. But it costs little. Across wallets, shares sold at a gain ended up worth just 0.9¢ more, relative to the price at the time, than shares held at a loss; for the largest wallets the difference is zero or reversed. The wallets most prone to the effect earned more per dollar and were more often profitable than the least prone. On a market whose prices partly reverse after moves, selling after a rise is close to the right thing to do.
| Wallets | Sold at a gain | Sold at a loss | Difference | Wallets selling winners more readily | Cost per share of selling winners, not losers | |
|---|---|---|---|---|---|---|
| All wallets | 916,217 | 50% | 40% | +9.9 pts(+8.7 to +10.9) | 58% | +0.9¢(+0.7 to +1.1) |
| Traded < $1k | 565,401 | 81% | 79% | +1.4 pts(+1.2 to +1.7) | 51% | +2.1¢(+1.6 to +2.8) |
| $1k–10k | 246,296 | 71% | 62% | +8.4 pts(+7.8 to +9.1) | 56% | +1.5¢(+1.1 to +1.9) |
| $10k–100k | 91,488 | 61% | 46% | +14.7 pts(+13.7 to +15.7) | 61% | +0.0¢(-0.4 to +0.4) |
| $100k–1m | 11,012 | 52% | 36% | +15.4 pts(+14.0 to +16.8) | 68% | -0.6¢(-1.2 to -0.1) |
| ≥ $1m | 2,020 | 39% | 28% | +10.3 pts(+8.6 to +12.5) | 67% | +0.5¢(-0.6 to +1.6) |
An event is an order by a wallet that already holds the outcome it trades, when the order's price is at least 1¢ above (gain) or below (loss) the wallet's average purchase price. "Worth at resolution minus price": the final payoff (0 or 100¢) minus the price at the event, per share; the last column compares shares sold at a gain with shares held at a loss, averaged across wallets with at least 10 events of each kind (positive = selling the winners cost money).
| Disposition quintile | Wallets | Median difference | Return per dollar | Profitable |
|---|---|---|---|---|
| Q1 (least) | 10,217 | -22 pts | -1.9%(-3.2 to -0.8) | 33% |
| Q2 | 10,090 | -1 pts | +4.1%(+0.9 to +7.1) | 32% |
| Q3 | 10,177 | +7 pts | +0.4%(-0.3 to +1.1) | 43% |
| Q4 | 10,117 | +23 pts | +1.3%(+0.8 to +1.9) | 47% |
| Q5 (most) | 10,151 | +49 pts | +2.3%(+0.5 to +4.0) | 51% |
Wallets with at least 10 gain and 10 loss events and at least $1,000 spent, in quintiles of (share sold at a gain − share sold at a loss).
Do traders take more risk after winning, playing with "house money" (Thaler and Johnson, 1990), or after losing, trying to get even? Here risk-taking is measured by the share of buying in long-shot contracts under 10¢ and by the average price paid. After any large result, wallets shift slightly toward long shots, more after losses (4–5 points) than after wins (2–3 points). But large wins and losses are not random: they happen to particular kinds of wallets at particular times. The coin-flip markets give the clean comparison. There the gap between unlucky losers and lucky winners shrinks to 0.6 points, at the edge of statistical significance. A weak break-even effect, and no sign of house money.
| The wallet's result in a resolved market | Events | Long-shot share of buying, before | Change after | Average price paid, change | Spend after / before (median) |
|---|---|---|---|---|---|
| Lost ≥ $5k | 15,736 | 7.1% | +5.2 pts(+4.4 to +5.9) | -1.3¢(-1.9 to -0.7) | 0.61× |
| Lost $500–5k | 95,874 | 9.7% | +3.8 pts(+3.4 to +4.2) | -1.0¢(-1.3 to -0.7) | 0.70× |
| Won $500–5k | 118,704 | 7.1% | +2.5 pts(+2.2 to +2.8) | +0.1¢(-0.2 to +0.4) | 0.88× |
| Won ≥ $5k | 18,113 | 6.3% | +2.7 pts(+2.0 to +3.4) | +1.6¢(+1.0 to +2.3) | 0.80× |
| Change from before to after | Lucky winners | Unlucky losers | Difference | Wallets |
|---|---|---|---|---|
| Long-shot share of buying (weeks +1 to +4 vs. −8 to −2) | +1.3 pts | +1.8 pts | -0.6 pts(-1.1 to -0.1) | 14,544 / 14,567 |
| Average price paid | +4.5¢ | +3.9¢ | +0.6¢(-0.0 to +1.2) | 14,544 / 14,567 |
A small group of wallets buys mostly long shots. Those putting half or more of their money into contracts under 10¢ (average price 10¢) lost 25% of what they spent, and only 27% of them came out ahead. Wallets that rarely buy long shots roughly broke even. The calibration page shows that contracts under 10¢ win about as often as their last traded price says. So the losses are probably not mispricing of long shots as such but how they are bought: a 1¢ spread is a tenth of the price of a 10¢ contract, and taking it again and again in thin markets adds up.
| Share of buying under 10¢ | Wallets | Average price paid | Return per dollar | Profitable |
|---|---|---|---|---|
| < 1% | 399,682 | 85¢ | +0.5%(-0.3 to +1.2) | 27% |
| 1–5% | 66,533 | 74¢ | +0.2%(-0.5 to +0.9) | 38% |
| 5–20% | 33,383 | 55¢ | -0.8%(-2.8 to +1.3) | 34% |
| 20–50% | 12,109 | 33¢ | -5.2%(-9.0 to -1.2) | 28% |
| ≥ 50% | 18,866 | 10¢ | -25.2%(-29.9 to -20.0) | 27% |
Data. Every trade in 74,099 resolved yes/no Polymarket markets (October 2023 to October 2026) from the public Data API, both sides of every fill: 173.9 million records. Activity is measured in markets in this sample, so trading in markets that had not resolved by October 2026 is not counted.
Coin flips. Markets whose last traded price a day before resolution was 40–60¢. A wallet enters at its first such market if it held at least 100 net shares of one outcome at resolution and bought something in the 8 weeks before; events in the last 12 weeks of the data are excluded so that every wallet has a full 12 weeks after.
Records. Quarterly snapshots built only from markets resolved before each quarter began: number of markets, total profit, and a t-statistic of profit per dollar across markets.
Disposition. Average purchase price is the dollar-weighted price of all the wallet's earlier purchases of that outcome in that market. Positions created by splitting a dollar into Yes and No shares are not visible as purchases, so orders by wallets without a recorded purchase are left out.
Limits. Wallets are not people: one trader can run many wallets, and a wallet that stops trading may have moved to another. Returns exclude fees and liquidity rewards.