A lot of people believe it is difficult to walk out of a casino with more money than you walked in with. It is actually incredibly easy.
Simply go to the roulette table and wager the table minimum—likely $10—on red or black. If you win, leave. If you lose, double your bet. Lose again, double again. Stop when you win, or when you’ve lost $310.
If 100 people walk into a casino with $310 and use this strategy, roughly 96 of them will leave with a $10 profit—even though all 100 were making the same bad bets. The other four will lose all $310.1
Those winners will have made a profit playing a game in which they were expected to lose. Nobody has an edge in roulette, so everyone who came out ahead was lucky.2
In other gambling games—poker, sports betting, prediction markets—it’s not that simple. Because there is skill involved, some players do have an edge, meaning they are likely to profit in the long run. But a winning run doesn’t tell you whether someone is one of those skilled few.
Poker players are acutely aware of how treacherous small samples can be. Entire sessions often boil down to a single hand. A recreational player’s pocket aces getting cracked by pocket kings may well be the difference between a winning month and a losing one.

For online players, a common rule of thumb is to play 100,000 hands before putting much confidence in an estimated win rate, and even then there is substantial room for luck to impact results.3
Most gamblers never produce anything close to that kind of sample. Yet the percentage who finish ahead is routinely presented as evidence of how many can beat the game.
Take Kalshi.
The company and its founders repeatedly cite the proportion of users who have made money as evidence that customers fare better there than elsewhere.
In June, COO Luana Lopes Lara said about 70% of prediction market users lose money, compared with 95% in other forms of gambling and financial speculation.
In July, CEO Tarek Mansour told Axios that “⅔ of people on Kalshi lose money.” In a rebuttal to a Roosevelt Institute report on user losses, Kalshi reported a “one to three ratio of net profitable users,” compared with casinos where “~0% of users are profitable.”4
The ratio comes from a figure Kalshi gave the Wall Street Journal in May: 2.9 unprofitable users for every profitable one, based on data from the past month.5
The past month.
Obviously, people sometimes leave casinos with more money than they brought in; casinos wouldn’t have many customers otherwise. But suggesting those lucky winners can expect to keep making money gambling is misleading. Kalshi understands this, and (correctly) excludes lucky casino winners from its comparison. But it counts its own lucky traders to suggest that more of its customers can beat the game.6
To see how little Kalshi’s statistic tells us about who is likely to profit in the long run, imagine someone who, instead of placing trades on Kalshi, plays online roulette. Each day they log in, dutifully wager $100 on red, watch the wheel spin, and log out after the ball lands. (The typical prediction market user places about one trade per day.)7
After 31 days, they will have wagered $3,100 on a game with a 2.7% house edge. Their expected loss is about $84. Yet they have a 44% chance of finishing the month profitable, compared with roughly 26% of Kalshi users in the month the company reported. A game nobody can beat would produce a larger share of monthly winners than Kalshi did.8
I asked more than a dozen experienced prediction market traders, employees, and analysts what percentage of Kalshi users they would expect to profit over the long run.9 Estimates ranged from 1% to 10%, mostly between 2% and 5%.10 For “regular” users—those without a strong quantitative or gambling background, who wouldn’t understand what an API connection is—nobody put the figure above 1%.11
Those estimates do not contradict Kalshi’s monthly figure. A quarter of users can finish a particular month in the green, but this month’s winners can be next month’s losers. The longer people keep making bets that lose money on average, the more likely they are to end up down overall.
Treating that monthly figure as a measure of how many customers can beat the game is statistical malpractice.
It’s one thing to cherry-pick data to skirt the uncomfortable reality that, as in any game where the pot doesn’t grow and the house takes a cut, users collectively lose money.
It’s another to use that data as evidence you’ve built “one of the most equitable, egalitarian marketplaces in existence.” Especially when your fees are so high that the average retail user would expect to lose less betting the same amount on online roulette.12
This is known as the Martingale strategy. The five bets are $10, $20, $40, $80 and $160, summing to $310. On a standard American double-green wheel, betting red or black wins 18/38 times, meaning the odds of losing 5 bets in a row are (20/38)^5 = 4.04%. The expected loss of employing this strategy is $2.93 (10*.9596) - (310*.0404).
Ok maybe a couple people can beat roulette.
Poker players should play around with a variance calculator to see just how much luck is involved. For example, if a player has a true win rate of 3 big blinds per 100 hands, which is very respectable, after 100,000 hands there is a more than 17% chance they will have lost money overall.
Ironically, the Roosevelt report—which found that regular users had lost half a billion dollars as of May—drastically undercounted user losses because they excluded fees.
Kalshi has not (to my knowledge) made any other data public about user profitability, and I believe all these references, despite their slight variations (~70% for Lopes Lara, ⅔ for Mansour, 2.9:1 to the WSJ) are referencing the same data.
A 2013 Wall Street Journal analysis of two years of records for 4,222 online casino gamblers found that they made money on 30% of the days they gambled, and on those days made a median of 49 bets per day, which is more trades than the average Kalshi user places in a month.
Kalshi does not make individual trading data public, but analyses of Polymarket and other similar platforms have found average users place roughly one trade a day, and there’s no reason to believe Kalshi users would be placing significantly more trades.
Diligent readers may notice I’ve switched from double-zero roulette in the opening example to single-zero here. First, thank you. This blog is a labor of love, and I have an extra soft spot for the footnote readers. Second, double-zero roulette is the default in most U.S. casinos (although they’re trying to make triple-green a thing), while single-zero roulette is the default online. Third, this is my blog, and if switching wheels makes my argument look better, goddamnit, I’m going to do it.
For the math: red wins on 18 of 37 pockets, and doubles the amount wagered. Finishing ahead after 31 spins requires at least 16 wins, which happens 43.97% of the time (calculated by adding the probability of winning exactly 16, 17, 18 and so on times, with an 18/37 chance of winning each spin). With 18 winning pockets and 19 losing ones, the expected loss is 1/37 of the amount wagered: $3,100/37, or $83.78.
In practice, roulette and other casino games will often produce larger losses than trading on a platform like Kalshi because bets resolve quicker, resulting in users placing more of them. The comparison here is per dollar wagered.
This estimate—and all other analyses and estimates in this post—are focused on individual sports event contracts on Kalshi, which make up the vast majority of trading volume. It is not including parlays (sorry, combos), which have a far lower expected return for average users.
No one will ever have an exact answer to this question, because most users will never develop a big enough sample of trades. But anyone who understands anything about gambling or math should understand how absurdly low 70% of players losing would be. Here is one conversation with professional gamblers including myself on the question. Here (1:15:30) is another, where the hosts, who I respect and think are very knowledgeable, lay out the case for a higher % of winning users (~10%). Here is a response to their estimates, agreeing with me.
An API (Application Programming Interface) lets software interact directly with a platform. Traders can use APIs to monitor prices and place trades automatically, instead of doing everything manually in the app or on the site. If you have to ask…
As I noted in a previous post, buying fairly priced 50/50 contracts on Kalshi produces an expected return of about −3.4% after taker fees, compared to −2.7% per spin on a standard single-green online roulette wheel. Lower-priced contracts carry higher fees per dollar wagered. This assumes users break even before fees, which is generous. Various analyses have found takers perform worse than makers before fees, often due to biases such as overpaying for “Yes” contracts.
Kalshi objects to treating market takers as a proxy for ordinary users. This critique is dumb. Anyone who knows anything about these platforms or has ever met retail users knows the vast majority of them are takers, do not post resting orders, and have roughly zero edge before fees––meaning their expected loss is roughly what they pay in fees. While there are some profitable traders on the taker side, their inclusion pulls the average taker return up, making recreational takers’ losses look smaller. (For more on market takers and makers, see here or here for a video explanation).
High fees are not inevitable on exchanges. Sports prediction markets such as ProphetX and Sporttrade offer lower fees. Novig, true to its name, currently charges zero fees on pregame straight bets. On these platforms, the average customer can expect to lose less per dollar wagered than on Kalshi.
Lower fees don’t necessarily mean a larger share of long-term winners. If prices are more efficient, there will be fewer opportunities to find an edge.



Great column! Very illuminating.
Thirty percent of which users? If losers stop opening the app, next month’s win rate can improve without anyone getting better. I’d follow the same January signups for a year, including the people who vanish.