Galaxy Research analyst Will Owens has published a study of Polymarket’s international platform, using its full onchain settlement history to track how users actually traded and how much they made or lost. Since launch in 2020, the platform has matched 1.27 billion orders for 3.07 million wallets and generated $82.8 billion in notional trading volume. Because every trade settles onchain, the dataset includes each position, entry price, holding period, and settlement outcome.
Galaxy narrowed the analysis to 2.9 million accounts whose trading pace looked consistent with manual execution. The headline finding: 69.2% of those accounts ended in losses, and the losing cohort posted aggregate losses of $338.9 million.
The report says the international platform should be viewed separately from Polymarket’s US app, which runs on a different venue with its own order books. It notes that Polymarket became widely known in 2024 because of the US election night cycle, later re-entered the US market through a subsidiary with a Commodity Futures Trading Commission, or CFTC, license, and introduced taker fees for the first time in early 2026. Because the dataset spans the platform’s full history, some accounts in the study traded entirely before fees were introduced.
Galaxy also points to a commercial shift now underway. This is the first full NFL season after both Polymarket and Kalshi entered the US market. Polymarket has signed sports figures including LeBron James, Eli Manning, and Derek Jeter as promotional partners, while its US app also rolled out a social feature called Squads, a private group product that lets users discuss markets and trade based on other members’ picks without leaving the app. The report says the platforms are spending to grow their user bases while targeting the same type of traders that performed worst in this dataset.
How Galaxy built the sample
The report is based entirely on public settlement records. That lets Galaxy study trading behavior without relying on broker disclosures or survey data.
Because the goal was to analyze human trading patterns, the report tried to remove automated accounts. Galaxy used average orders per active day as its proxy, defined as total orders divided by the number of days on which an account actually traded. There was no natural cutoff in the distribution, so the study set the line at 50 orders per active day.
Using that threshold, Galaxy excluded 125,429 accounts, equal to 4.1% of the total. Even so, those accounts were responsible for 80.8% of all orders and 41% of notional volume. The report says that pattern is consistent with a small set of accounts driving a very large share of activity.
In the report, “retail” does not mean small by capital base. It means accounts trading at a pace that appears manual. A well-funded trader who makes their own decisions and clicks to place orders would still be counted in this group.
Galaxy defines a profitable position as one whose settlement value exceeds the purchase cost, whether or not the holder redeemed it. That choice matters because expired positions that go to zero often are not redeemed. Looking only at redemption records would leave out much of the losing side and make the results look better than they are.
The report also flags a limitation. Wallet addresses are treated as separate traders, but there is no reliable way to know whether multiple addresses belong to the same person. A user trading through several wallets would be counted as several accounts. Galaxy says that is especially relevant when reading the finding that losing accounts are more likely to stop trading, since some of those accounts may simply have moved to a new wallet.
Most accounts lost money, but most losses were small
Among the 2.9 million manually traded accounts, 69.2% ended in losses. Together they lost $338.9 million.
The median profit and loss across these accounts was roughly a $3 loss. Half of all accounts fell between a loss of $36.64 and a gain of $0.40. Galaxy says those amounts are too small to be life-changing for anyone. As expected, larger wins and losses appeared in the tails: the 1st percentile account lost $4,804, while the 99th percentile account made $3,381.
Measured against capital put at risk, the median account lost about 0.5% of the money it deployed. The 10th percentile account lost 90%. In other words, most accounts lost only a small amount, while a much smaller group lost several thousand dollars.
The 125,429 excluded accounts that Galaxy classified as automated ended with aggregate profits of $246.8 million. The report says their distribution also fits expectations. Many repeatedly traded to capture incentives, while a much smaller set appears to have engaged in market making or arbitrage.
The gains of automated accounts and the losses of manually traded accounts do not fully net out. Galaxy says about $92 million of the gap comes from factors outside the trader groups, mainly unsettled positions.
Profits kept traders around; losses pushed more of them away
One of the report’s main questions was whether traders cash out and leave after a win or keep trading. Most keep going. Still, the difference between winning and losing was clear.
After a profitable trade, 6.1% of accounts did not open another position within 30 days. After a losing trade, that number rose to 15.2%. Galaxy says an account was about 2.5 times more likely to stop trading after a loss than after a gain.
At first glance, the next-position data looked counterintuitive. Unadjusted results showed that 46.6% of accounts increased the size of their next position after a gain, compared with 50.2% after a loss. Galaxy says that comparison is distorted by entry prices.
Median entry prices were much lower for losing positions than for profitable ones, at $0.43 and $0.86 respectively. Since each contract settles at $1 if the event occurs and $0 otherwise, a $0.43 contract implies a market-assigned probability of 43%. At lower entry prices, it becomes mechanically easier for the next trade to look larger in dollar terms.
Once entry price is controlled for, the result flips. Within the same price bucket, traders were more likely to increase their stake after a gain than after a loss. Galaxy says that effect is concentrated in positions entered above $0.50. Below that level, reactions after wins and losses looked almost the same, possibly because traders already viewed those positions as low-probability attempts and did not read too much into a single outcome.
Risk generally moved lower, especially after losses
Galaxy measures risk as expected loss rather than headline notional. For a position of t tokens bought at average price p, expected loss is t × p × (1 − p). By that measure, a $100,000 position priced at a 99% probability is large, but not especially risky.
Across the sample, traders usually reduced risk after a position settled, whether that position was a win or a loss. Their next trade tended to carry slightly less risk than the one that had just resolved. The difference lies in how much they pulled back.
After a gain, 48.4% of next positions carried more risk than the previous one. After a loss, that figure was 44.7%. The median change in risk was zero for both groups, which means most traders simply returned to their usual risk level.
To avoid comparing fundamentally different trader types, Galaxy split the sample into five equally sized groups based on the level of risk each trader typically takes. Q1 contained the lowest-risk traders and Q5 the highest-risk traders. Galaxy then compared how each group behaved after wins and losses, rather than comparing one group’s averages against another’s.
Category focus helped only in some areas
The report also looked at specialization. Galaxy defines a “single-domain trader” as someone who participated in at least five categorized markets and directed more than 60% of their activity to one category. On that basis, 44.1% of traders were single-domain traders and 55.9% were cross-domain traders.
Polymarket labels its markets, and Galaxy merged those labels into 10 broad categories: crypto, sports, politics, finance, economics, weather, culture, international affairs, technology and science, and business. Subcategories usually also carry the parent label. A football market, for example, is tagged under sports.
Single-domain traders performed slightly worse overall. Just 28.1% of them ended profitable, compared with 30.4% of cross-domain traders. Galaxy traces that to concentration: 61% of single-domain traders were clustered in the three weaker categories of sports, politics, and culture.
Sports alone accounted for 47% of all single-domain traders, and only 25.1% of those accounts were profitable, the lowest share of any category. Outside sports, politics, and culture, every other single-domain category had a profitable-account share above the 30.4% posted by cross-domain traders. Finance came in at 36.8%, while technology and science reached 41.2%. Galaxy notes that technology and science is a smaller category with a smaller sample size.
The report argues that specialization can help when traders genuinely have an edge in a field. Someone who only trades markets tied to OpenAI model releases may have an information or analysis advantage, not necessarily inside information but perhaps a stronger read on public data. Someone who only trades NFL markets on Sundays, by contrast, is probably not doing actuarial analysis.
Single-domain traders were also more active. The median number of markets they participated in was 18, versus 4 for cross-domain traders. Galaxy notes that traders must have entered at least five categorized markets to qualify as specialized, which means low-activity accounts naturally fall into the cross-domain bucket.
Winning traders tended to place larger bets
Galaxy compared profitable and unprofitable traders by median holding period and median position size. On size, the difference was clear. Winning traders had a median position size of $13.96, while losing traders came in at $10.00.
The report says that alone does not settle the question, because profitable traders also traded more often. So Galaxy grouped traders by the cumulative number of positions they had opened. Inside every activity bucket, profitable traders still matched or exceeded losing traders on median position size, and in most buckets the gap stayed noticeable.
Among traders who had opened 5 to 9 positions, the median position size was $12.53 for winners and $7.05 for losers. Among traders who had opened 50 to 99 positions, the figures were $13.14 and $8.90 respectively.
Holding periods told a less consistent story. Looking at the full sample without adjustment, winning traders held positions for about 20 hours at the median, versus about 25 hours for losing traders. Once Galaxy grouped traders by activity level, that relationship stopped holding consistently. In some groups, winners held longer; in others, losers did.
Galaxy says the data do not support a clean link between patience and profitability in prediction markets. The report contrasts that with meme coin trading, where ultra-short-term scalping or rapid trading in new pairs often lines up with better performance. Those traders are often snipers or highly experienced participants who may hold for only seconds before exiting. Galaxy adds that this style of trading is very different from how traders operate in prediction markets or perpetual futures.
Fees and marketing may change the mix of users
In its closing section, Galaxy says prediction markets were originally built to aggregate information and draw on the wisdom of crowds, but in recent years they have increasingly been criticized as pure gambling venues. The current user base on Polymarket’s international platform fits that tension: roughly 69% of accounts lost money, and the largest single-domain cohort was concentrated in sports markets.
Fees are now a bigger part of the equation. Polymarket started charging taker fees on crypto price direction markets in January 2026, and by late March had expanded those fees to nearly every category. Galaxy gives an example: buying 100 contracts at $0.50 each in a crypto market requires $50 of capital, and the taker would pay $1.75 in fees, equal to 3.5% of that stake.
Rates vary by category. Politics, finance, and technology carry the lowest fee at 2.0%, while sports is 2.5%. Galaxy says the median manually traded account lost about 0.5% of the capital it deployed over its entire trading history. At a 50% event probability, a single taker trade now costs several times that amount.
Still, the report does not argue that prediction markets have no value as truth-discovery tools. Galaxy says the opposite may be true: in the absence of some other subsidy, having most participants lose money may be part of what allows information aggregation to work. Capital with an informational edge needs someone less informed to trade against. Polymarket can leave most participants in the red while still producing forecasts that are useful to non-traders. In that framework, noise traders absorb the cost of the predictions.
Galaxy stresses again that the study is limited to the international platform. Polymarket’s US exchange is separate and has become a major focus for the company. Citing Front Office Sports, the report says Polymarket pays LeBron James $15 million a year, about four times his NBA playing income this year. To comply with NBA rules, his promotional content is limited to American football markets. Jeter and Manning have also signed deals with the platform.
The report says those marketing efforts are aimed at the very cohort that performed worst in the dataset. From a business perspective, Galaxy says, that is not hard to understand. At least in the short run, the platform has more reason to attract indiscriminate taker flow than to prioritize more professional liquidity providers.
Why the report contrasts Polymarket with Kalshi
Galaxy closes by drawing a distinction on verifiability. It points to a dispute involving Kalshi on Sept. 20, when an X user known as beniduboss accused the exchange of overstating crypto perpetual futures volume. The cited figures were unusual: ETH-PERP showed about $538.6 million in 24-hour volume and only about $3.1 million in open interest. For comparison, Galaxy says Hyperliquid’s ETH-PERP typically shows about $1.3 billion in 24-hour volume and about $3.1 billion in open interest.
Kalshi denied the accusation and said the critic had confused prediction-market contract counts with perpetual-futures notional volume. As of the time the report was written, Galaxy says regulators had not taken action.
The larger point in the report is that outsiders cannot fully test the claim from public Kalshi data alone. Kalshi’s public feed does not identify both sides of each trade, so external observers cannot determine whether counterparties are controlled by the same entity. On Polymarket, by contrast, counterparties are visible as public addresses. Galaxy says anyone who disagrees with the report’s conclusions can examine the records directly.
As prediction markets expand and move into regulated venues that rely on proprietary order books, the report says outside verification of trading-volume data can no longer be taken for granted.