Is Retail Crypto Trading Better Than Gambling?
Memecoins, perpetual futures and prediction markets against casino odds
A casino game has a payout law, so its return per unit staked and the chance of walking away ahead can be computed exactly. A trade fits the same units once its cash in and cash out are known, and on a public blockchain they are known for every wallet. This paper rebuilds six weeks of retail trading on pump.fun, Hyperliquid and Polymarket from on-chain records and prices each venue against six casino games under tests registered before any outcome existed.
Memecoins paid back 75 for every 100
Retail memecoin traders got back 0.754 per unit staked (95% interval 0.726 to 0.777), below American roulette at 0.947 and below every other game tested. Only 14.1% of their wallets ended the six weeks ahead, against 19.7% to 49.1% of gamblers betting the same stakes.
No jackpot to pay for
A memecoin bet placed at a random time returns a hundred times its stake about once in 250,000 tries, against once in 1,928 at Plinko. The slot-machine story fails on its own terms, so the losses come from how retail trades and whom it trades against.
The other two venues pay casino rates
Per unit of peak exposure, Hyperliquid retail got back 0.9999 and Polymarket retail 0.984, level with craps, baccarat and Plinko. Their share of winners (33.5% and 39.0%) sits between the slots and the table games.
Overview
Memecoin traders are called gamblers all the time, mostly as a figure of speech. The paper takes the comparison literally: every position has a payout multiple, cash out divided by cash in, which is the same object as the payout of a casino bet, and so one set of statistics can price a launchpad and a roulette table side by side.
Four tests were registered before any profit or loss was computed. Two compare retail with the games (return per unit staked, and the share of wallets ending ahead against gamblers placing the same stakes), while the other two ask whether some retail traders are skilled and whether their edge persists, using a permutation null that keeps the trading day and the stake size fixed.
A trade is a bet with a payout law
For each pump.fun (wallet, token) pair, cash in is every SOL paid on buys including protocol, creator and network fees, while cash out is every SOL received on sells plus whatever the wallet still held at the end, valued by selling it into the token's final pool at the price that sale would produce. Hyperliquid episodes run from flat to flat with funding charged hour by hour, while Polymarket positions count purchases, mints, merges and conversions, with unresolved tokens at their last trade price.
Each position, episode or bet gets a payout multiple X, exactly as a casino bet does. A wallet ends ahead when its pooled multiple across all its positions is above one.
A venue's return to player (RTP) is total cash out over total cash in across all retail positions, with a 95% interval from a bootstrap over calendar days (10,000 resamples).
Retail is defined by rules fixed in advance that use no outcome information. On pump.fun the filters remove each token's creator, wallets that bought in the launch block or the two blocks after it, wallets that received tokens from the creator, the protocol's own trading wallet, atomic bots and any wallet with more than 200 trades in a clock hour.
Below every game in the set
Over the window, retail memecoin traders put 46.35 million SOL into positions and got back 0.754 per unit. Every 100 SOL came back as about 75, whereas 100 SOL on red at a European roulette table returns 97 on average, which left retail 11.42 million SOL poorer after six weeks (about US$1.07 billion at the window's mean SOL price).
Every gap between memecoins and a game is significant after a Holm correction across the six games, so H1 holds for all six. On Hyperliquid and Polymarket it holds for none, since neither venue is significantly worse than any game.
Fewer winners than at a slot machine
Whether a wallet ends ahead depends on how many bets it places and how big they are, so a raw comparison with a gambler's odds would mislead. The paper replays each retail wallet's actual sequence of stakes through each game's payout law 200 times and counts how many of those stake-matched gamblers end ahead.
Slot games produce fewer winners than table games because rare large wins carry most of their payout, and memecoin wallets end ahead less often even than slot players. On the other two venues, 33.5% of Hyperliquid wallets and 39.0% of Polymarket wallets end ahead, fewer than at every table game yet more than at Plinko or Sweet Bonanza.
Interactive · casino session replay
Play 4,000 sessions of equal bets at any of the six games, drawn from the paper's own payout laws, then sort the sessions by how much of their stake came back. The dashed outlines are the measured memecoin retail wallets in the same buckets.
sessions ahead
·
back per 100
·
median session
·
memecoin wallets ahead
·
Simulated in the browser with equal stakes. The paper's H2 test instead replays each wallet's own stake sequence, so the shares here move with the number of bets while the paper's figures do not.
Half of all memecoin retail wallets (49.1%) lost nine tenths or more of what they put in. Restricted to wallets that staked at least 0.1 SOL in total, 19.9% end ahead, closer to the quarter reported by Bitquery's count of Solana memecoin buyers, which stops two weeks after each launch and covers other launchpads too.
A slot machine without the jackpot
The usual defence of memecoins says buyers expect most trades to fail and pay for a small chance at a hundredfold gain, the way a slot player pays for the jackpot. A separate registered study tested that directly by placing five bets of $100 on every token launched in the census, each entering at a random time in the token's first day and holding for anywhere between a minute and a week.
Over 5,246,772 random bets the RTP is 0.979 and the median 0.975, which is what a round trip through pump.fun's fees leaves, because most random bets land on tokens where nothing trades between entry and exit. Only 0.2% of their total payout comes from multiples of 100 or more, against 9.5% at Plinko, so the registered ruling fails on the jackpot condition: under the paper's own rule a random memecoin bet is not a slot.
Interactive · waiting for a 100× win
Set a number of plays and read the chance of at least one payout of 100 times the stake or more, 1 − (1 − p)ⁿ, from each source's measured or exact odds.
Odds per play: random memecoin bets and retail positions measured in the paper, Plinko exact from its published multiplier table, Sweet Bonanza from 34,968,204 demo spins tracked by 101rtp.net.
Actual retail positions have a fatter tail than random bets because traders pick tokens that are moving, yet a 100× position still comes up once in about 115,000, some 60 times rarer than at Plinko. Their median position returns 0.679, which puts the losses in many positions that each give back a large share of the stake rather than in a lottery ticket that rarely pays.
Skill exists, and it is not where it pays retail
On venues where a few positions return a hundred times the stake, a lucky wallet and a skilled one look alike. The null therefore reshuffles payout multiples among positions taken on the same day in the same stake decile, then counts the wallets whose return beats the 95th percentile of their own reshuffled distribution, where luck alone puts the share at 5%.
H4 · wallets beating their own null
H5 · top decile in the later period
Memecoins give the strongest evidence of skill, with 10.6% of wallets beating their null and the first half's top decile getting back 0.980 in the second half while everyone else got back 0.716. The venue that is worst for retail on average is also the one where a minority most reliably beats the rest.
Hyperliquid's skilled minority is real but small per unit of exposure, and its persistence is the most fragile result in the paper: the top decile clears its null in only 2 of 5 weekly folds. Polymarket wallets keep their rankings across halves (Spearman 0.173 against a null of 0.027), yet the first half's top decile does worse than everyone else afterwards, which points to stable trading styles such as backing favourites rather than skill.
Who takes the money
Wallets removed from the retail sample by rule did far better than retail. Token creators got back 1.56 per unit staked (launch-day interval 1.47 to 1.66) for a net gain of 1.52 million SOL, and launch-block buyers got back 1.49 for 1.49 million SOL, whereas atomic bots (0.76) and high-frequency wallets (0.85) did about as badly as retail.
Positions that sold tokens they had received by transfer show a multiple of 2.09, but another wallet paid part of their cost, so that figure tracks where tokens went rather than trading skill. Within retail, the 24,415 wallets in the first half's memecoin top decile enter a median 2 minutes after launch and hold a median 62 seconds (5 minutes and 3 minutes for everyone else), a profile that fits automated or semi-automated trading.
The 2026 result belongs to its window
The registered window covers six weeks, so the retail computation was repeated without registration on Dune's decoded events for two windows of the same shape in 2024, one starting 8 August and one starting 8 November near that year's memecoin peak. Graduated tokens then moved to Raydium rather than PumpSwap and the curve charged 1% a side with no creator fee, so the rules were adapted rather than copied.
* In Nov–Dec 2024 all classes together took out more than they put in (1.011), so bags are overvalued there and the retail figure reads as a range from 0.921 to 0.984.
Memecoin retail got back 0.971 in August to September 2024 and between 0.921 and 0.984 at the late-2024 peak, against 0.754 in 2026, so the shortfall describes that market at that time rather than memecoins in general. In the 2026 window, positions that traded in a PumpSwap pool after graduation returned 0.710 and carry 91% of retail losses, while positions that stayed on the bonding curve returned 0.901.
What it takes to move the number
Retail filters are rules rather than identities, and the direction of their error is known: any insider left in the sample raises measured retail returns. Every variant the paper reports keeps memecoin retail well below the casino band, the closest being positions opened more than an hour after launch at 0.814.
The Dune rebuild is the strongest of these checks: it starts from a second data source with no collection gaps, holds 1,165,879 launches against 1,108,160 in the main data and still gives 0.754 (95% interval 0.724 to 0.780) with 14.9% of wallets ahead.
Limitations
- Six weeks. Memecoin and prediction-market activity is seasonal, and the 2024 memecoin windows paid retail between about 0.92 and 0.98 per unit.
- Wallets are not people. One person can run many wallets and a wallet can be shared, so the share of people ending ahead could differ in either direction.
- Stake on perpetual futures. Margin is not observed on chain, so peak notional exposure is the stake; per unit of margin there would be far more dispersion and fewer winners, and 30.5% of retail wallets were liquidated at least once.
- End-of-window values. Open memecoin positions are valued by selling into the final pool state, and Polymarket positions in unresolved markets at the last trade price.
- Sweet Bonanza. Its replay uses an approximate payout law built from measured demo spins, while its RTP comparison uses the published 96.5%.
- RTP is not harm. A casino-level RTP on Hyperliquid or Polymarket does not make those venues harmless, any more than a high RTP does for a slot.
Method in brief
- Window: 8 August to 18 September 2026 (UTC); pump.fun tokens launched up to 10 September, so every token has at least eight days of trading.
- Data: pump.fun curve and PumpSwap events from a public Hugging Face corpus and Dune; Hyperliquid node fill records with hourly funding; Polymarket exchange, conditional-token and negative-risk adapter events on Polygon.
- Games: European and American roulette, the craps pass line and the eight-deck baccarat banker bet exact from their rules; Plinko (16 rows, high risk) exact from its published table; Sweet Bonanza at its published 96.5%.
- Tests: H1 day-cluster bootstrap with Holm correction across six games; H2 stake-matched replay, 200 per wallet; H4 and H5 against 200 permutations within day and stake decile.
- Registration: the memecoin studies were registered and hashed on 23 September 2026 and the three-venue study on 10 October 2026, before any Hyperliquid or Polymarket outcome; every later change was registered, hashed and timestamped on the Bitcoin blockchain before it was applied.
Reproducibility
Every script used for the data pulls, positions, tests and figures is in the replication package, along with the aggregated results behind each number in the paper, the registrations with their OpenTimestamps proofs and a synthetic smoke test that runs in about a minute. Every figure on this page is drawn in the browser from those aggregated results.
Cite this paper
See also
Two other papers measure crypto trading from the trader's side: a market maker's adverse selection on Bybit, Binance and Hyperliquid, and whether DEX-only tokens can be traded at all.
Keep this free to read
Running the lab costs about $400 a month, out of my own pocket. Sponsorship is what covers it.