Polymarket's five-minute Bitcoin makers earn at settlement. Fully 91.4% of their profit arrives there, leaving 8.6% from intraperiod round trips. Zang, Andrade and Nakajima find that almost all of the money comes from buying. Purchases earned roughly +$70K, versus +$2.6K for sales. Purchases between 0.50 and 0.80 contributed +$52.4K, while those between 0.80 and 1.00 added +$24.8K.
Persistent makers finish at +$45.3K across 275 markets and $10.38M of notional. Expand the count to all 451 maker-classified accounts and the total falls to +$29.0K. Retail ends at -$66.7K. Takers pay 1.8%; makers pay nothing, receive no rebate, and have their profits measured after fees. These figures contain no maker subsidy.
The paper turns on that settlement split. Classical dealer economics relies on payoff-uninformative flow to fund market making. Prediction-market shares instead pay 0 or 1 within a few minutes, with no cash flow, consumption value or reference price along the way. The authors argue that makers get paid by quoting beyond the range a signal can reach. Informed traders therefore avoid the book, while makers earn rent by carrying the under-priced side through resolution. Bartlett and O'Hara's 41.6 million Kalshi trades receive the same interpretation: behavioral overbetting covers the maker's adverse-selection losses.
What the model puts in place
The model has three periods. At t=0, an automated market maker (LMSR) commits depth L. CLOB makers post two-sided quotes and incur posting cost kappa. Neither venue can revise its commitment. A common shock arrives at t=1, after which informed traders move price to the post-shock posterior. Tail-demand traders follow. They cannot value the shock and instead trade from a behavioral wedge around the price they encounter. Settlement comes at t=2.
The pickoff result carries the model. Once stated, it is nearly immediate. Informed traders certainly lift any pre-shock quote inside the posterior band. A surviving ask must therefore lie at least sigma above the prior, and a surviving bid at least sigma below it. The informed route to the LMSR. This reverses the familiar Glosten-Milgrom ordering: the AMM absorbs the cost of price discovery, while the book excludes information and earns from tail demand.
Two imposed routing conventions drive the remaining results. The authors do not derive them. They defend the pair as extremal allocations of path rents, one most favorable to the book and the other least favorable, with intermediate routing falling between them.
The bracket remains wide and untested.
Under firm-quote priority, a tail trader uses the book whenever its committed price is available. L then disappears from the free-entry premium, making depth neutral for the book's spread. Price competition reverses the allocation. Traders consume cheap LMSR units first, so greater depth takes volume from the book.
Maker-side contestability determines the spillover's sign. A contestable book, where another maker can undercut before the shock, shifts rightward along its increasing revenue branch and widens. A single committed maker instead chooses a new peak farther left and compresses. Beyond a threshold, the pool eliminates the book's clientele.
The multi-outcome analysis stands apart. With n>=3, every trade switches exposure by buying one leg and selling another. The transaction price equals the difference between two quotes. No trade ever reaches the common level of the quote vector. Competition fixes the switch premium while leaving that level undetermined. Independently margined binary books require collateral of ntheta, compared with Llog n for one LMSR potential. The ratio is exactly (theta/L)(n/log n).
Classification choices and empirical reach
The maker/retail classification is a footprint heuristic built from two-sidedness of quoting, round-trip frequency, within-market inventory mean reversion, holding horizon and clip size. The paper supplies no thresholds for any of the five. It identifies 451 maker accounts with 24.9% of volume, 12,749 retail accounts with 55.0%, and 1,934 unclassified accounts with 20.1%.
The headline +$45.3K comes from 407 persistent makers. Another 44 maker-classified accounts lose $16.3K and are excluded as structurally similar yet persistently unprofitable. Include all 451 and makers earn +$29.0K. The authors state this directly. They also report that Akey et al. find the same whole-platform pattern by order type: limit-order accounts win and market-order accounts lose.
The regime evidence follows a similar pattern. In the pooled data, the longshot band [0.20,0.40) is overpriced by +0.043, while the favorite band (0.60,0.80] is underpriced by -0.037. Conditioning on the realized price path changes the result. Longshots run at +0.225 in trending markets and -0.071 in flippy markets.
A fixed misunderstanding of probabilities cannot explain a bias whose sign changes. The authors use that result to make tail demand state-contingent in the model. Yet the classifier depends on the realized within-window path and becomes available only when the market has finished. It identifies the source of the rent retrospectively.
A second sample, and cleaner checks
The multi-outcome evidence uses the Akey et al. (2026) full Polymarket record. It contains 588 million trades and $67 billion of volume from 11 November 2022 to 29 March 2026. The authors reconstruct 16,036 validated multi-outcome events with outcome counts of 2 to 10. Among them, 9,988 have every leg priced on both sides at least once. The resulting panel contains 3,343,329 price-update rows.
The specification is unusually complete, and the authors have already performed the checks that would be my first three. The set premium slope on log n is +0.0595 (HC3 se 0.0056). Per-leg markup accounts for +0.0573, while the book center contributes only +0.0022 (p=0.52). The NO side reproduces the result at +0.048.
Equal-age comparisons retain 84% to 98% of the slope, ranging from +0.0546 to +0.0461 on the panel where every leg has just traded. A drift correction estimated from roughly 8.4M consecutive-trade pairs changes any slope by at most 0.006. When the authors deliberately impose large-book staleness on clean fast binaries, the estimate becomes -0.0031, the wrong sign.
An execution-size filter nearly removes the relationship. Requiring each leg's last trade to be at least 100 shares reduces the slope to +0.006 and makes it insignificant, compared with +0.074 without the restriction. Bundles of k unrelated binary markets, matched by calendar week, horizon and liquidity, generate +0.062. On a per half-spread basis, that equals +0.031, around half the real markup slope.
The authors treat the bundle result as confirmation because a bundle mechanically accumulates per-leg markup yet cannot manufacture a book-center effect. I agree with that reading. It also shows that plain spread accumulation, absent any event structure, accounts for half the multi-outcome cost.
Order-book snapshots are unavailable. The reconstructed ask is the last taker-buy and the reconstructed bid is the last maker-buy, often observed at different times. About 8% of the 3,343,329 price-update rows consequently have the ask below the bid. The authors argue that averaging across legs and time cancels the error. Probably. Their adversarial refill of unobserved hours can reverse the regression's sign at age thresholds of four hours or less, so they limit the claim to full-life and same-age estimates.
The venue-choice claim remains beyond the data
Neither transaction sample includes an AMM, as the authors state in the first section. The venue-choice result in Section 4 therefore goes untested, along with the coexistence equilibrium and the depth spillover whose sign depends on contestability. The paper sends its one testable prediction to Section 5. That section contains the evidence assessed above: +0.0595 on log n, +0.006 at 100-share executions, and +0.062 from unrelated bundles.
Proposition 3's two branches assign opposite signs to the same intervention. The pair is genuinely falsifiable, yet the paper falsifies neither branch. Aoyagi and Ito (2025), cited by the authors, provide the closest evidence: a positive-spillover result resembling the contestable-book case.
We could not run this ourselves for a simpler reason: no maker/taker identities and no LMSR pool state. Substituting listed assets loses the bounded 0/1 payoff that produces the mechanism. We have previously covered adjacent Polymarket plumbing in oracle adjudication timing.
The regime sign reversal and tail-demand calibration come from one contract type over about twenty-four hours. The authors identify this as a limitation. It is the paper's weakest link.
The settlement-versus-spread split has stronger support. It persists across all Polymarket markets and by order type. The book-center slope is also estimated across time-to-resolution buckets, reaching +0.081 at 7-30 days and +0.060 beyond 30 days. Even so, the 56.8% favorable-resolution rate on carried inventory is a thin edge to defend over longer horizons.
The gauge argument requires no calibration. Trades determine differences between quotes and never their common level, even though that level represents probability. Operating a multi-outcome book as n independent binaries leaves displayed probabilities unpinned. Wherever margin is segregated book by book, collateral grows as n instead of log n. I would act on that result.