Author: Will Owens, Analyst at Galaxy Research
Compiled by: Jiahuan, ChainCatcher
Introduction
Since launching in 2020, Polymarket's international platform has matched 1.27 billion orders for 3.07 million wallets, with a notional trading volume of $82.8 billion. Because all trades settle on-chain, we can view the complete record, including every position, entry price, holding period, and settlement proceeds, to understand users' actual behavior in prediction markets. This report narrows the scope to 2.9 million accounts whose trading cadence is consistent with manual operation.
This report poses five questions about these accounts: After making money, do traders cash out and leave, or do they press their luck? Does profit make them bolder, and does loss make them pull back? How are profits and losses distributed? Do traders typically specialize in specific domains, and does that specialization lead to better returns? What exactly are the differences between profitable and unprofitable traders?
Studying retail trading behavior typically relies on broker disclosures or survey data. But Polymarket settles on-chain, so there is no longer any need to debate whether the sampling is reasonable.
The core conclusion is: 69% of retail accounts ultimately lose money, and this group loses a combined $339 million.
In 2024, Polymarket became widely known mainly because of U.S. election night. Since then, the company has returned to the U.S. market through a subsidiary holding a U.S. Commodity Futures Trading Commission (CFTC) license, and in early 2026 it introduced taker fees for the first time. The discussion below concerns the international platform, which is a completely separate trading venue from the U.S. app, with its own order book.
Polymarket adjusted its fee structure earlier this year. Because the data in this report covers the platform's full history, some of these accounts conducted all of their trading before the platform began charging fees.
This is also the first full NFL season since both Polymarket and rival Kalshi entered the U.S. market. Both platforms have therefore increased their spending. At the start of the season, Polymarket had enlisted sports stars such as LeBron James, Eli Manning, and Derek Jeter as "promotional partners," and this promotional push also sparked some opposition.
Alongside this promotional push, the U.S. app launched the social feature Squads. Squads are private groups where users can discuss prediction markets and trade based on other members' picks, all without leaving Polymarket. The feature is designed to bring group chat discussions onto the platform. Another platform is now adding social trading features, a similar trend also seen in pump.fun's recent announcement.
Faced with such a large potential trading volume, both platforms are spending money to expand their user base, and the very traders they attract are the worst-performing type in this report. Whether the composition of traders will change significantly over the next year, or whether it will merely increase substantially in number, is worth continuing to watch.
Key Takeaways
Of 2.9 million "retail" accounts, 69.2% ultimately lost money.
This group lost a combined $338.9 million.
Losses appear to increase user churn.
After a loss, 15.2% of accounts did not trade again within 30 days.
After a gain, that figure was 6.1%.
If traders may indeed have an edge in a certain domain, specializing in that domain leads to better returns.
44.1% of traders concentrated more than 60% of their trading activity in a single domain.
Traders focused on sports had the worst profit performance, possibly because many of them are just hobbyists or occasional users rather than market makers or arbitrageurs.
Traders focused on tech and science had the best profit performance; some of them may possess inside information, while others may be experts in the relevant fields.
Profitable traders stake larger amounts per trade.
Their median position size was $13.96, compared with $10 for losing traders.
Research Methodology
Dataset: Polymarket settles on-chain, meaning the complete record of every order, position, entry price, and settlement proceeds is public. All analysis in this report is based on these records.
Research subjects: We want to study human behavior, so we need to exclude automated accounts trading via scripts. Although it is impossible to accurately determine which accounts use automated trading, it can be estimated using "orders per active day." This metric equals an account's total number of orders divided by the number of days on which that account actually traded.
Orders per active day follow a continuous distribution with no natural cutoff point. Therefore, the threshold must be determined by the researcher's judgment. We used 50 orders per active day as the cutoff, excluding 125,429 accounts, or 4.1% of the total. Although these accounts account for "only" 4.1%, they contributed 80.8% of all orders and 41% of notional trading volume. This filtering method is intended to identify the situation in which a small number of accounts account for an extremely large share of trading activity.
Retail: Our screening is based on trading frequency. Even if a trader is well capitalized, as long as they make their own judgments and manually click to place orders, they are classified here as retail. Therefore, the conclusions of the study apply to accounts trading at a manual cadence, rather than to "retail" in the strict economic sense.
Profit measurement: As long as the value of an account's position at settlement is higher than its purchase cost, we consider it profitable, regardless of whether the holder actually redeemed. Positions that expire worthless are usually not redeemed, and if only redemption records were counted, most losses would be excluded, making this group's profit performance appear better than it actually is.
This report has one important limitation: we identify traders by wallet address, but cannot reliably determine whether two addresses belong to the same person. Therefore, the same person using multiple wallets would be counted as "multiple accounts" in this analysis.
This is especially important to keep in mind when interpreting the conclusion that "losers are more likely to stop trading." An account that appears to have exited may simply have switched to a new wallet.
How are profits and losses distributed?
Of retail accounts, 69.2% ultimately lost money. This group lost a combined $338.9 million.
The median profit or loss for retail accounts was about a loss of $3, with half of accounts' profit or loss falling between a loss of $36.64 and a gain of $0.40. These amounts are not enough to change anyone's life. As expected, larger profits and losses are concentrated at the tails of the distribution: accounts at the 1st percentile lost $4,804, while accounts at the 99th percentile gained $3,381.
By amount invested, the median account lost about 0.5% of the funds it put in, while accounts at the 10th percentile lost 90%. Overall, the vast majority of accounts lost only a very small amount of money, and only a few accounts lost thousands of dollars.
We classified 125,429 accounts as "automated accounts," and they ultimately made a combined profit of $246.8 million. Their profit-and-loss distribution is also as expected: a large number of accounts earn trading rewards through repeated trading, while far fewer are genuinely engaged in market making and arbitrage.
The profits and losses of these two groups do not fully offset each other. The gap of about $92 million comes from factors outside the trader population, mainly unsettled positions.
After making money, do traders cash out and leave, or keep trading?
Most keep trading. Profit tends to keep users on the platform: after a gain, only 6.1% of accounts did not establish another position within 30 days; after a loss, that figure was 15.2%. Losing accounts were about 2.5 times as likely to leave as profitable ones.
Intuitively, making money should lead traders to stake more on their next trade. But the unadjusted data show the opposite: after a gain, 46.6% placed a larger next position; after a loss, that figure was 50.2%. However, this comparison is affected by other factors.
Losing positions had far lower entry prices than winning positions, with medians of $0.43 and $0.86, respectively. Therefore, when the entry price is lower, it is easier for the next trade's stake to increase. (If the prediction comes true, each contract settles for $1; otherwise, zero. Thus, a price of $0.43 means the market believes the event has a 43% probability of occurring; those buying at that price believe the true probability is higher.)After controlling for entry price, the conclusion reverses. Within the same price range, traders who are in profit increase their stakes more often than traders who are in loss. This phenomenon is mainly concentrated in the range where the entry price is above $0.50. Below that price, the reactions to profit and loss are almost the same, possibly because traders already view these trades as low-probability attempts and do not overinterpret any single outcome.
Does profit make traders more willing to take risk? Do losers add to their positions?
Losers usually contract. Here, "risk" refers to expected loss, not the full principal that could be lost. Suppose a buyer purchases t tokens at an average price p; the buyer's expected loss is t × p × (1 − p). This means that a $100,000 position predicting an event with a 99% probability of occurring, although large in size, does not count as a high-risk position by this measure.
Whether the previous trade was profitable or unprofitable, a trader's risk usually declines. When they trade again, they tend to take slightly less risk than in the position that just settled. However, the contraction after profit is noticeably smaller: after profit, 48.4% of next positions are riskier than the previous one; after loss, that share is 44.7%. The median change in risk is zero for both groups, so most traders simply return to their prior risk level.
We divide traders into five equal-sized groups based on the position risk they typically take. Q1 normally takes the least risk, and Q5 the most. The comparison of behavior after profit and loss is conducted within each group, so what is being compared is each trader's own performance after different outcomes.
Do traders focus on one domain or spread across multiple domains? How does profitability differ?
We define users as "single-domain traders" if they have participated in at least five categorized prediction markets and more than 60% of those belong to the same domain. Under this definition, 44.1% of traders are single-domain traders, and 55.9% are "cross-domain traders."
These questions can be answered because Polymarket adds tags to every prediction market. To avoid classifying all traders as "single-domain traders," we merge these tags into ten domains: crypto, sports, politics, finance, economics, weather, culture, international affairs, technology and science, and business. Subcategories usually also carry the tag of the parent category; for example, a soccer prediction market is also tagged as sports.
Single-domain traders perform slightly worse, with 28.1% ultimately profitable, compared with 30.4% for cross-domain traders. The reason is that 61% of single-domain traders are concentrated in the three worst-performing domains: sports, politics, and culture.
Someone who only trades NFL game predictions on Sundays is probably not doing actuarial analysis.
Clearly, if traders genuinely may have an edge in a certain domain, focusing on that domain can produce better returns. Someone who only trades prediction markets related to OpenAI model releases is very likely to have some kind of edge, and that is not necessarily an insider-information edge; it may simply be skill in analyzing public information. Someone who only trades NFL game predictions on Sundays is probably not doing actuarial analysis.
Sports alone accounts for 47% of all single-domain traders, and among them the share of profitable accounts is 25.1%, the lowest of all domains. In every domain other than sports, politics, and culture, single-domain traders have a higher profitable share than the 30.4% for cross-domain traders. Among them, finance is 36.8%, and technology and science is 41.2%. Note that technology and science is a relatively small category with a smaller sample.
The median number of prediction markets single-domain traders have participated in is 18, compared with 4 for cross-domain traders. Because determining whether someone focuses on one domain requires the trader to have participated in at least five categorized prediction markets, low-activity accounts are by default classified as cross-domain traders.
What distinguishes profitable traders from losing traders?
Comparing the median holding time and median position size of profitable and losing traders yields several conclusions.
In terms of position size, the median single-position size for profitable traders is $13.96, compared with $10.00 for losing traders. However, profitable traders also trade more frequently, so this difference may simply be due to different activity levels rather than a difference in the amount invested itself. After grouping traders by the cumulative number of positions they have established, profitable traders' single-position investment is no lower than that of losing traders in every group, and in most groups it is clearly higher.
Among traders who have cumulatively established 5 to 9 positions, the median single-position sizes for profitable and losing traders are $12.53 and $7.05, respectively; among traders who have cumulatively established 50 to 99 positions, the figures are $13.14 and $8.90, respectively.
Holding time is less informative. Looking at all traders together, profitable traders hold for shorter periods, with a median of about 20 hours, compared with about 25 hours for losing traders.
But after grouping by activity level, the relationship with holding time changes. In some groups, profitable traders hold longer; in others, losing traders hold longer. Therefore, we cannot draw a clear relationship between "more patience" and "greater likelihood of profit" from these data.
This differs from meme coin trading. In the meme coin market, people who engage in "scalping"-style ultra-short-term trading, or who quickly trade new pairs, perform better. They are usually sniper traders or experienced traders who hold for only a few seconds and quickly take profit.
For this type of trader, see our meme coin report; for the nature of KOLs and the resulting trading behavior, see our social trading report. It should be noted that this "trading" approach is very different from trading perpetual contracts or prediction markets.
Outlook
The original goal of prediction markets was to aggregate information and harness the wisdom of crowds. But in recent years, they have increasingly been criticized as pure betting platforms. This report presents the profit-and-loss distribution of the Polymarket user base: 69% ultimately lose money, and the largest user group focused on a single domain mainly trades sports-related prediction markets.
Polymarket began charging taker fees on crypto price up-or-down prediction markets in January 2026, and by the end of March the fee range had covered almost all categories. Taking a price of $0.50 per contract as an example, buying 100 contracts in the crypto market and investing $50 means the taker must pay $1.75 in fees, equivalent to 3.5% of the amount invested in that single trade.
Politics, finance, and technology have the lowest rates, at 2.0%; sports is 2.5%. The median share of funds lost by retail accounts over their entire trading history is about 0.5% of total investment. Today, when an event has a 50% probability of occurring, the fee for just one taker trade is already several times that share.
This report does not deny the value of prediction markets as a tool for discovering truth.
It can be argued that, absent other forms of subsidy, most participants losing money is precisely the precondition for information aggregation to occur. Capital with an information edge needs to trade with traders who do not have such an edge. If all participants on Polymarket were equally sharp, no one would trade. This report does not deny the value of prediction markets as a tool for discovering truth. Polymarket can perfectly well let most participants lose money while providing predictive reference for people who do not participate in trading. It is these noise traders who bear the cost of prediction.
This report is strictly limited to the international platform. Polymarket's U.S. exchange is separate from it and is also the company's recent main focus of investment. For example, the funds the company invested to invite LeBron James as a partner reflect this. According to Front Office Sports, Polymarket pays James $15 million per year, about 4 times his NBA playing income this year. To comply with NBA rules, his promotional content is also limited to American football-related prediction markets. Jeter and Manning have also signed partnership agreements with the platform.
These marketing investments are aimed precisely at attracting the group with the worst profitability in this dataset. From a business perspective, this is not hard to understand. Polymarket probably does not care whether users are profitable. At least in the short term, the platform has a stronger incentive to attract taker users who participate in trading without screening than to attract more professional maker traders.
On Polymarket, the counterparty to every trade is a public address. Anyone who disagrees with the analysis in this report can verify it themselves.Recently, a controversy involving Kalshi also deserves attention. On September 20, a quantitative trader named beniduboss on X accused the exchange of inflating the trading volume of cryptocurrency perpetual contracts. This is a serious allegation.
The data Beni cited was: ETH-PERP's 24-hour trading volume was approximately $538.6 million, while open interest was only about $3.1 million, which is highly unusual in the perpetual contract market. By comparison, Hyperliquid's ETH-PERP typically has about $1.3 billion in 24-hour trading volume and about $3.1 billion in open interest.
Kalshi officially denied the allegation, saying the other party had confused the number of contracts in a prediction market with the notional trading volume of perpetual contracts. As of the time of writing, regulators have not taken action.
It is worth noting that people outside the exchange actually cannot verify this matter. Kalshi's public data feed does not identify the parties to each trade, so it is impossible to determine from this data alone whether both sides of a trade are controlled by the same entity. On Polymarket, by contrast, the counterparty to every trade is a public address. This means that anyone who disagrees with the analysis in this report can verify it themselves.
As prediction markets continue to develop and enter regulated trading platforms that use proprietary order books, the ability of outsiders to verify trading volume data is no longer something that can be taken for granted.