The Behavior of Polymarket Traders

Introduction

Since launch in 2020, Polymarket’s international platform has matched 1.27 billion orders across 3.07 million wallets, worth $82.8b in notional. Because all of it is settled onchain, it’s possible to read the entire record (every position, entry price, holding period, and payout) and see how users actually behave on the prediction market. For the purposes of this report, we will narrow accounts to the 2.9 million accounts that actually trade like people.

This report asks five questions about those accounts. Do winners cash out or press their advantage? Does winning make traders bolder, and does losing chasten them? How are profit and loss distributed? Do traders typically specialize in certain topics, and does specialization pay? And what separates the profitable from the unprofitable?

Retail trading behavior is typically studied through broker disclosures or survey data. Because Polymarket settles onchain, there is no sampling frame to argue about.

The data behind this report was curated by Stork, an oracle network for prediction markets and perpetual futures. Every figure here is derived from Polymarket’s public settlement record, with Stork handling the indexing and delivery.

The TLDR is that 69% of retail accounts finish below break-even (at a loss) and the population is down $339m in aggregate.

Back in 2024, Polymarket was mainly known because of election night. The company has since re-entered the U.S. through a CFTC-licensed subsidiary and introduced taker fees for the first time in early 2026. Everything below describes the international platform (which is a completely separate venue from the U.S. app, with its own order book).

Polymarket revamped its fee structure earlier this year. Because the data in this report covers the platform’s full history, some accounts here traded entirely before fees existed.

This is also the first full NFL season in which both Polymarket and rival Kalshi have established U.S. presences. Both have been spending accordingly. Polymarket already opened the season with stars LeBron James, Eli Manning, and Derek Jeter as “promotional partners,” a campaign which has faced some backlash.

That campaign launched alongside Squads, a social feature inside the U.S. app. Squads are private groups where users chat about markets and trade from each other’s picks without leaving Polymarket. The feature is designed to move group chat conversations onto the platform. Once again, a platform has started to add social trading features (see pump.fun’s recent announcement).

With so much potential trading volume on the line, both platforms are spending to enlarge the exact cohort this report finds least profitable. It will be interesting to see whether the composition of who trades (or just how many of them there are) changes drastically over the next year.

Executive Summary

• 69.2% of the 2.9 million “retail” accounts (defined below) finished below break-even

• The population is down $338.9m in aggregate

• Losing appears to raise churn

  • After a loss, 15.2% of accounts had not traded again within 30 days

  • This is against 6.1% after a win

• Specialization pays where traders plausibly have an edge

  • 44.1% of traders concentrate more than 60% of their activity in one topic

  • Sports specialists are the least profitable, probably because many are amateurs or dabblers, as opposed to market makers or arbitrageurs

  • Tech and science specialists are the most profitable (some may be insiders, while others could be subject matter experts)

• Profitable traders bet bigger

• A median position of $13.96 against $10 for unprofitable traders

Methodology

Dataset: Polymarket settles onchain, which means the complete record of every order, position, entry price, and payout is public. Everything in this report is derived from that record.

Population: We are asking how people behave, so automated accounts that trade by script have to come out. It’s impossible to know exactly which accounts are automated, but one way to get an idea is by looking at orders per active day. These are orders placed divided by days on which the account traded at all.

The distribution of orders per active day is a single continuum with no natural break. Therefore, it’s a judgement call to make the cutoff. We use 50 orders per active day as the cutoff, which removes 125,429 accounts (4.1% of the total). Even though these accounts are “only” 4.1% of the total, they placed 80.8% of all orders and 41% of all notional volume. The filter is meant to catch the fact that a small number of accounts do an enormous share of the trading.

Retail: Our filter separates accounts by trading frequency. A well-funded discretionary trader clicking manually counts as retail here. The findings depend on the population trading at a human pace instead of being ‘retail’ in any economic sense.

Measuring Profit: An account is profitable if its positions were worth more at settlement than they cost, whether or not the holder redeemed them. Because positions that expire worthless are typically left unclaimed, counting only redemptions would drop most losses out of the record and make the population look more successful than it is.

A notable limitation of this report is that traders are identified by wallet addresses, and there is no reliable way to establish that two addresses belong to one person. Anyone trading from multiple wallets appears as “multiple accounts” in this analysis.

This is an important caveat to the finding that losers tend to stop trading. An account that appears to have quit may simply have just moved to a fresh wallet.

How is Profit and Loss Distributed?

Out of retail accounts, 69.2% finished below break-even (unprofitable). In aggregate, the population is down $338.9m.

The median retail account is down ~$3.00 and half the population falls between -$36.64 and +$0.40. These are amounts that wouldn’t change anybody’s life. As expected, the money is in the tails. The first percentile is -$4,804 and the 99th is +$3,381.

Per dollar committed, the median account gives up about half a percent of what it puts in. The tenth percentile loses 90% of it. Generally, the vast majority of accounts lose trivial amounts and a small number lose thousands of dollars.

There are 125,429 accounts we classify as “automated.” They finished +$246.8m. The shape of their PnL is also as expected. A large number of accounts churn for trading rewards, and a much smaller number do genuine market making and arbitrage.

Do Winners Cash Out, or Keep Going?

Mostly, they keep going. Winning tends to keep people on the platform, with just 6.1% of accounts having not opened another position after a win within 30 days. After a loss, 15.2% of accounts hadn’t opened one. A losing account is ~2.5x as likely to walk away.

It may seem intuitive that winning makes the next bet bigger. Raw rates actually suggest otherwise (46.6% of post-win positions were larger, against 50.2% after a loss), but that comparison is confounded.

Losing positions are entered at much lower prices than winning ones (a median of $0.43 against $0.86), so sizing up in dollar terms is far more common at low prices. (Each share pays $1 if the prediction comes true, and zero if it doesn’t, so a $0.43 price implies the market gives the event a 43% chance; a buyer at that price thinks the true odds are higher.)

Holding entry price fixed reverses the takeaway. Within a price band, winners size up more often than losers. The effect concentrates above 0.50. Below it, winning and losing produce nearly the same response, plausibly because traders treat longshots as such and don’t read much into either outcome.

Does Winning Breed Risk-Taking? Do Losers Chase?

Losers generally retreat. “Risk” here is defined as expected downside (as opposed to capital at risk). For t tokens bought at average price p, the buyer expects to lose t * p (1 – p). This means that a $100,000 position at 99% odds is not a risky position, even though it’s a large position.

Risk falls after both outcomes. Traders typically come back a little smaller than the position that just settled, whether it won or lost. But they cut back considerably less after a win: 48.4% of post-win positions were riskier than their predecessor, against 44.7% after a loss. The median change is zero in both groups, so most traders simply return to the risk level they left at.

Traders are split into five equal groups by their own typical position risk. Q1 takes the smallest usual risk; Q5 is the largest. The win/loss comparison runs inside each group, so traders are only compared against themselves.

Are Traders Specialists or Generalists? How Does Performance Vary?

We define users as specialists if more than 60% of the markets they traded fall under a single topic, measured across at least five categorized markets. By this definition, 44.1% of traders are “specialists” and 55.9% are “generalists.”

These questions are answerable because Polymarket tags every market. In order to avoid classifying every trader as a “specialist”, we collapse the tags into 10 topics: crypto, sports, politics, finance, economy, weather, culture, world, tech and science, and business. Sub-topics reliably carry their parents (for example, soccer markets are tagged sports).

Specialists do slightly worse: 28.1% of them finished profitable against 30.4% of generalists. This is because 61% of them concentrate on three topics that underperform (sports, politics, culture).

Someone who trades nothing but NFL games on Sunday is unlikely to be doing an actuarial analysis.

Obviously, concentration pays where a trader plausibly has an edge. An individual who trades nothing but OpenAI model release markets likely has some kind of edge (and not necessarily an insider’s edge; they could be astute readers of public information). Someone who trades nothing but NFL games on Sunday is unlikely to be doing an actuarial analysis.

Sports alone accounts for 47% of all specialists at a 25.1% profitable rate, the worst of any topic. Specialists in every topic outside sports, politics, and culture beat the 30.4% generalist rate, rising to 36.8% in finance and 41.2% in tech and science. It’s important to note that tech and science is a smaller category (smaller sample).

Specialists have also traded a median of 18 markets against four for generalists. The concentration test requires at least five categorized markets, so low-activity accounts land in the generalist group by default.

What Separates Profitable Traders From Unprofitable Ones?

When comparing profitable and unprofitable traders by their median holding time and median position size, we can draw a few conclusions.

On position size, a profitable trader’s typical position is $13.96 against $10.00 for the median unprofitable position. Profitable traders also trade more often, so the difference could just be activity rather than sizing. Splitting traders into bands by how many positions they have ever opened, profitable traders bet at least as much in every band and clearly more in most.

Among traders with five to nine positions, the gap is $12.53 (profitable) against $7.05 (unprofitable). Among those with 50-99, it is $13.14 (profitable) against $8.90 (unprofitable).

Holding period is less of a signal. Pooled together, profitable traders hold for less time (~20-hour median versus ~25 for unprofitable traders).

But when we break the traders into activity bands, holding periods change. In some bands, profitable traders hold for longer, and in others unprofitable traders do. Because of this, we can’t draw a clean relationship between a trader being “patient” and a trader being “more profitable” here.

This is in contrast to memecoin trading, where traders who “scalp” or trade new pairs quickly are more profitable. These are typically snipers or experienced traders who hold for just a few seconds and realize quick gains. See our memecoin report for more on these types of traders, as well as our social trading report for more on the nature of “key opinion leaders” (KOLs) and the trading behavior enabled here. Disclaimer: it’s a very different style of “trading” compared to perps or prediction markets.

Outlook

While the original telos behind prediction markets was to aggregate information (the wisdom of the crowd), they have been increasingly criticized as purely gambling platforms. The distribution in this report describes the population on Polymarket. In this population, 69% finish down and the largest concentrated cohort trades primarily sports.

Taker fees arrived on Polymarket for crypto price direction markets in January 2026 and covered nearly every category by the end of March. At 50 cents, a taker in crypto pays $1.75 per 100 shares on a $50 position (3.5% of capital committed on a single trade). Politics, finance, and tech run cheapest at 2.0%, sports 2.5%. The median retail account gave up about half a percent of everything it committed across its entire history. One taker trade at even odds now costs several times that.

Nothing in this report undermines the case for prediction markets as truth machines.

Arguably, a losing majority is a precondition for information aggregation (absent some other form of subsidy). Informed capital needs uninformed flow to trade against. If all participants on Polymarket were sharp, nobody would trade. Nothing in this report undermines the case for prediction markets as truth machines. Polymarket can be an unprofitable endeavor for most participants and a forecasting tool for nonparticipants at the same time. The noise trading is what pays for the forecast.

This report strictly covers the international platform; Polymarket’s U.S. exchange is separate and is what the company has been focusing on primarily as of late. This is evidenced by its spends to acquire LeBron James, for example. According to Front Office Sports, Polymarket is paying James $15m annually, which is ~4x more than what he will be making this year for playing in the NBA. His deliverables are also restricted to football market promotions to stay inside NBA rules. Jeter and Manning also signed deals with the platform.

The cohort that spending is designed to attract is the one with the worst record in this dataset, which makes sense from a business standpoint. Polymarket presumably doesn’t care whether its users are profitable. At least in the short term, it is incentivized to onboard indiscriminate traders who are takers, as opposed to more sophisticated makers.

With Polymarket, the counterparty to each trade is a public address. Anyone who disagrees with the analysis can check it.

A recent dispute at Kalshi is also worth noting. On Sept. 20, a quant on X (who goes by the handle beniduboss) alleged the exchange was inflating crypto perpetual volume, a serious allegation. Beni cited ~$538.6m in 24-hour ETH-PERP volume against ~$3.1m in open interest, which is highly unusual for perp markets. For context, Hyperliquid typically has ~$1.3b in 24-hour ETH-PERP volume against ~$3.1b in open interest.

Kalshi officially denied the allegation, arguing that prediction-market contract counts had been conflated with perpetual futures notional. No regulator has acted at the time of writing.

What’s interesting here is that nobody outside the exchange can really tell. Kalshi’s public feed doesn’t identify the participants on either side of a trade, so whether one entity controlled both legs is unanswerable from the data. With Polymarket, the counterparty to each trade is a public address. That means anyone who disagrees with the analysis can check it.

As prediction markets continue to scale into regulated venues with proprietary order books, auditing a volume figure from the outside stops being a given.