This article was compiled and organized by BlockWeeks
Since its launch in 2020, the Polymarket international platform has matched 1.27 billion orders, covering 3.07 million wallets, with a nominal trading volume of up to 82.8 billion USD. Because all transactions are settled on-chain, every position, entry price, holding period, and final payout is fully recorded, allowing outsiders for the first time to observe how users actually behave in prediction markets using full data rather than sampling or questionnaires—and the record delivers a rather unflattering report card.
An in-depth study of this public settlement record narrowed the account scope to 2.9 million accounts that "trade like humans," finding that 69.2% of retail accounts ultimately ended below the break-even line, with the entire group suffering a combined net loss of 338.9 million USD. The data indexing and delivery were completed by the prediction market and perpetual contract oracle network Stork.
The report attempts to answer five questions: Do winners cash out or press their advantage? Does profit make traders bolder, and loss make them more conservative? How are profits and losses distributed? Do traders specialize in certain themes, and does specialization pay? And, what exactly separates those who make money from those who lose it?
First, remove the bots: 4.1% of accounts consumed 80.8% of orders
To ask how "humans" trade, one must first exclude script accounts. The study used "number of orders placed per active trading day" as the criterion, but this distribution is a continuous curve with no natural breakpoint, so the cutoff is a subjective judgment. The study ultimately adopted 50 orders per active day as the threshold, excluding 125,429 accounts, or 4.1% of the total.
The key point is that this "mere" 4.1% of accounts contributed 80.8% of all order volume and 41% of nominal trading volume. This filter is precisely meant to capture the fact that "a very small number of accounts complete a very large proportion of trades."
It should be noted that "retail" here is defined by trading frequency rather than capital size: a well-funded full-time trader who manually clicks to place orders is also counted as retail. The report's conclusions depend on the group that trades "at a human pace," not on any economic definition of "retail."
Profit and loss: most people lose money, and losers are more likely to disappear for good
Profitability is determined by whether the value of a position at settlement exceeds its cost, regardless of whether the holder has already redeemed it.
The results are not optimistic: 69.2% of accounts ended below the break-even line, with the group suffering a net loss of 338.9 million USD.
The exit effect caused by losses is particularly pronounced. After a loss, 15.2% of accounts did not open a new position within 30 days; after a profit, this proportion was only 6.1%. In other words, a losing account is about 2.5 times more likely to "wash its hands of it" than a profitable account.
It is worth noting that this platform does not consist only of retail traders. If Polymarket were full of experts, no one would trade. This also points to an inherent structure of prediction markets: it is a "truth machine," and at the same time it can be a money-losing machine for most participants.
Do winners increase their bets? The answer depends on price
"Win and bet bigger" seems intuitive, but the raw data says the opposite: after a profit, 46.6% of positions were larger than the previous one, while after a loss this proportion was 50.2%. However, this comparison has an obvious confounding variable.
The entry price of losing positions is usually much lower than that of winning positions—a median of 0.43 USD versus 0.86 USD. Since each share pays 1 USD if the prediction comes true and zero if it does not, a price of 0.43 USD means the market believes the event has a 43% probability of occurring, while those willing to buy at this price clearly believe the true probability is higher. At low prices, "adding to a position in dollar terms" is naturally more likely to occur.
Once the entry price is fixed, the conclusion reverses: within the same price range, winners increase positions more frequently than losers, and this effect is most concentrated when the price is above 0.50. Below 0.50, the reactions triggered by profits and losses are almost identical—possibly because traders already regard such low-priced contracts as a gamble and do not care much about a single win or loss.
Risk appetite: both winners and losers pull back, but winners pull back less
Here, "risk" is defined as expected downside loss, not capital invested. When buying t tokens at an average price of p, the buyer's expected loss is t × p × (1 − p). This means that a 100,000 USD position with a 99% win rate is not considered high-risk, despite its large size.
Whether winning or losing, risk declines—traders usually return with a slightly smaller size than the position just settled. But the contraction after a profit is noticeably smaller: after a profit, 48.4% of positions had higher risk than the previous one, while after a loss this was 44.7%. The median change for both groups is zero, indicating that most people simply return to their habitual risk level.
The study also divided traders into quintiles based on their own typical position risk, with Q1 being the lowest and Q5 the highest, and compared wins and losses separately within each group to ensure each trader was compared only with themselves.
Does specialization pay? Sports is the worst, technology and science the best
44.1% of traders concentrated more than 60% of their activity in a single theme. The returns to specialization are highly differentiated.
Sports specialists had the worst profitability, most likely because this group is full of amateurs and dabblers rather than market makers or arbitrageurs. In contrast, technology and science specialists were the most profitable—some of them may be insiders, while others may be genuine domain experts. In other words, specialization pays only when traders truly possess an information or cognitive advantage.
Another detail is position size: the median position of profitable traders was 13.96 USD, while that of losing traders was 10 USD.
Why the platform does not care whether users make money
The report specifically notes that it covers Polymarket's international platform, which is completely separate from the independent U.S. venue (with its own order book). Polymarket recently re-entered the U.S. market through a CFTC-licensed subsidiary and introduced taker fees for the first time in early 2026—since the report's data covers the platform's entire history, some accounts conducted all their trades before fees existed.
This is also the first full NFL season after Polymarket and its rival Kalshi both established a foothold in the United States, and both are spending heavily. Polymarket opened the season with LeBron James, Eli Manning, and Derek Jeter as "promotional partners," a marketing move that sparked controversy. According to Front Office Sports, Polymarket pays James 15 million USD per year, about four times his annual NBA salary; to circumvent NBA rules, his deliverables were limited to football market promotion. The marketing launched alongside the Squads feature—Squads are private communities within the U.S. app where users can discuss markets and follow others' choices to place orders directly, with the aim of bringing group chat conversations onto the platform. Social trading features are becoming standard for prediction platforms.
Interestingly, these marketing investments target precisely the group with the worst performance in this dataset. From a business perspective, this makes perfect sense: Polymarket most likely does not care whether users are profitable. At least in the short term, its incentive is to bring in more indiscriminate taker order flow rather than smarter market makers. As potential volume grows larger, both platforms are spending money to expand this "least profitable" group, and how the composition of traders changes over the next year will be worth watching.
Kalshi controversy: on-chain data makes external auditing possible
The report also mentions a recent controversy involving Kalshi. On September 20, a quantitative trader using the pseudonym beniduboss accused the exchange on X of inflating crypto perpetual contract volume, claiming its 24-hour ETH-PERP volume was about 538.6 million USD while open interest was only about 3.1 million USD—extremely unusual for a perpetual market. For reference, Hyperliquid's ETH-PERP typically shows about 1.3 billion USD in daily volume corresponding to about 3.1 billion USD in open interest. Kalshi officially denied the allegation, saying prediction market contract counting and perpetual contract notional value had been conflated; as of the report's writing, no regulator had intervened.
The truly interesting part is this: no one outside the exchange can really tell the truth. Kalshi's public market data does not identify who the counterparties are, so the question of "whether the same entity controlled both legs" cannot be answered from the data.
Polymarket is different—every trade counterparty is a public address, and anyone who disagrees with the analysis's conclusions can verify it themselves. As prediction markets continue to expand into regulated venues with their own order books, externally auditing a volume figure will no longer be taken for granted.
Returning to the most fundamental point: this report does not diminish the value of prediction markets as "truth machines." Polymarket can simultaneously be a place where most participants lose money and a forecasting tool in the hands of non-participants. Noise trading is precisely the part that pays for prediction.
