Twenty-eight strategy families, four venues, and the fee function that reorders all of them

Prediction markets look inefficient from outside. Retail-dominated flow, thin books, contracts whose fair value is a probability anyone can reason about, and hard no-arbitrage constraints (YES and NO must sum to a dollar, a ladder of thresholds must be monotone) that are trivially checkable and visibly violated on screen.

That invites a systematic screen rather than a hunch, so we enumerated the space first and measured second. Twenty-eight strategy families, grouped by what the edge depends on rather than by what it is called: structural no-arbitrage, liquidity provision, statistical prediction, term structure, resolution mechanics, cross-venue structures, and flow analysis. Four venues: Polymarket, Limitless, Opinion, Predict Fun.

The useful output of a screen is not the survivor. It is the reason most candidates die, because that reason is usually one thing wearing many costumes. Here it is the cost function, and it reorders the entire ranking before a single hypothesis is tested.

Start with the fee, because it decides the ordering

Polymarket charges takers and pays makers nothing. That is visible on chain rather than inferred from documentation: within a single fill, the two OrderFilled events carry different fees, maker fee = 0 and taker fee non-zero, and the collateral ledger agrees, with the taker short by notional plus fee and the maker short by notional alone.

The fee is not linear in notional. It is shaped:

Read the last two rows together. A single marketable leg costs more than the entire round-trip spread. The venue's economics are that resting is free and crossing is taxed beyond almost any edge that a screen of this kind can find. We derived that fee structure and its consequences for arbitrage in more detail in an earlier note on the fee wall; what follows is what it does to the whole family list.

The reordering rule is simple enough to state in one line, and it is the single most useful thing in this post:

If a strategy can express itself by resting, its cost floor is zero. If it must cross, it needs more than 300 bps of mid-book edge before anything else is considered.

Almost every classical prediction-market idea is a crossing strategy. That is why the screen looks brutal.

The screen

Status per family. STRUCT-0 marks a structural zero, an accounting identity rather than a measurement. TODO means not examined, never "available". Polymarket is measured throughout; Limitless is measured for the A and B groups; Opinion and Predict Fun are reconnaissance only, meaning read paths were proven and no strategy was tested, so their columns are omitted rather than filled with an implied verdict.

Group A: the structural ideas mostly are not ideas

The screen's first genuine surprise was that its most attractive family is not a measurement at all.

Complete-set arbitrage looks like the cleanest trade in the space: buy YES and NO for less than a dollar, redeem the pair, keep the difference. Across three independent samples taken hours apart on two machines (1,755, 2,416 and 926 markets), ask(YES) + ask(NO) had a minimum of 100.10c and a median of 101.00c, with zero observations below par in 2,416 markets. The floor sits one tick above a dollar.

The temptation is to write that up as evidence about venue efficiency. It is not. Both venues quote one logical book, and the test could not have found a violation. Polymarket serves two endpoints that are exact mirrors, verified to the cent at |1 − ask_YES − bid_NO| = 0.0000; Limitless serves one book and you fold it. Either presentation gives ask_YES + ask_NO = 1 + spread, which is greater than or equal to one by construction. Three replications of an accounting identity is still an accounting identity.

That is worth stating plainly because it is a screen failure mode rather than a market finding: we ran a test that had no power, three times, and it returned a clean, replicable, entirely uninformative null. The correct action is not to re-run it on the two unexamined venues. It is to check whether their books are genuinely independent, because that would be the finding.

The rest of group A dies to the fee. Ladder monotonicity violations (A3) are frequent and persistent, which is genuinely interesting about how these markets price: on a threshold ladder, P(reach $60k) must be at least P(reach $70k), and it often is not on screen. But the violations are smaller than the cost of the two crossing legs needed to collect them. Frequent, persistent, and below the fee is the modal outcome of this entire screen.

Group C: the statistical families die, and one dies interestingly

Calibration (C1) and favourite / longshot (C2) both died to controls rather than to the fee, which is the more instructive way for a candidate to fail. Momentum (C3) is the one worth reporting, because it did not merely fail: prediction markets in this sample overreact rather than underreact, with reversal monotone in move size. The sign is the opposite of the hypothesis.

Lead / lag (C4) produced the cleanest single sentence in the screen. The only lead we could establish between related markets was arithmetic, one market mechanically implying another, and it was 18x too small to cross. Catalyst convergence (C6) into scheduled sports resolution is real and roughly 2 to 4x too small to trade. Macro prints remain untested and are the one open branch of that family.

Notice the pattern across A3, C4 and C6: the effects are real, replicable, and an order of magnitude under the cost of capturing them. Prediction markets are not efficient in the sense of prices being right. They are efficient in the narrower and more practically important sense that the accessible mispricings are smaller than the toll.

The one post-fee-positive structure, and why it exists

Exactly one structure in this screen was positive after fees, and it exists because of a detail that only appears once you look at more than one venue: the fee curves have different shapes.

Three shapes means the venue that is expensive for one leg is cheap for the other. At a low probability, Predict Fun charges 0.100c where Polymarket charges 0.238c; at a high probability the ordering inverts completely, 1.900c against 0.238c. Routing each leg to the venue whose curve is cheap for that leg is the entire mechanism behind V1 and V2, and it was the only thing in this screen that survived its own costs.

It is also the family that got retired, and not for a research reason. Cross-venue structures require accounts on both venues. The largest documented source of edge in this space is cross-venue market making against a US-regulated venue, and that venue will not onboard from this jurisdiction. Its liquidity incentive programme excludes international users and its designated market-maker programme requires an executed agreement, so solving the account problem alone would not open it. The constraint is jurisdictional, not intellectual, and it is worth naming precisely because a screen that quietly drops such families overstates how much of the space is genuinely closed.

Group B: where the screen actually spent its time

Liquidity provision is the family that survives the fee reordering by construction, because a maker never crosses. It is therefore where a screen like this should concentrate, and it is where we went.

The short version, measured over two independent overnight runs, 1,395 marked fills and 35,198 shares, with no orders placed: unsubsidised spread capture at retail size loses about 6.7 cents a share, on every exit path a maker actually has. We capture the spread reliably. Adverse selection takes about 1.35x what we capture. That result, its mechanism, and the several things we tried that did not fix it are the subject of a companion post, because it deserves more room than a screen entry.

What belongs here is the structural point. Two of the B families, B1 and E2, are not skill strategies at all. They pay for resting depth out of a subsidy pool, pro rata. The terms are public, and the arithmetic of pro rata is unforgiving at small size:

A retail-scale quote across the twelve richest reward markets earns $3.13 a day. Nothing about that is a skill result. It divides by depth, so it rewards balance sheet and nothing else. A reward-driven strategy is a capital story, not a skill story.

Note also where the pools sit: Iran, the Fed, the Strait of Hormuz. Deep, heavily quoted geopolitics and macro. These are not the thin, wide-spread markets a quotability screen selects. Chasing the subsidy means quoting into queues of several hundred thousand shares where a retail order never reaches the front.

What the screen actually concluded

Not "market making does not work on prediction markets". It demonstrably does: the venue runs a gated rewards pool reported at $100k/day for combination markets, five-figure daily pools for sports, and is hiring to bring institutional market makers on. Firms are doing this profitably, on multiple venues at once.

The measured claim is narrower and more useful:

Unsubsidised spread capture, at retail size, without a reference price, without cancellation speed, and on one venue, loses about 6.7 cents a share. The subsidy available on those terms was $3.13 a day, against a gated pool three orders of magnitude larger.

That is a statement about market structure, not about skill. The economics of providing liquidity here are access-conditional: they depend on subsidy tier and on multi-venue reach, and the gap between the two regimes is measurable rather than rhetorical. A screen that enumerates the space and then prices the boundary is more useful than one that finds a strategy, because the boundary is what tells you which side of it a given desk is on.

The three mechanisms that explain the gap are worth stating because they are all linear or better in size. Cross-venue quoting does not absorb adverse selection, it transfers it: the informed trader who runs you over on one venue hands you a position you offset on another, so the drift we measured is a warehousing cost a hedged desk never pays. Coverage changes the statistics: adverse selection averages across thousands of uncorrelated events in a way it cannot across dozens. And subsidy divides by depth. None of those can be out-thought at retail size, which is the honest conclusion of the whole screen.

What is open

Three things are unmeasured rather than closed, and we are not claiming otherwise. Reference-anchored quoting, which is being paid for having a better price than the book, was never built. Cancellation speed has one negative result against it and remains under-tested. Hyperliquid HIP-4 presents a genuinely different book profile and its adverse selection is unmeasured. Each is named here so the screen is not read as more complete than it is.

What this does not measure

The B-group numbers come from marked fills, not live orders: a recorder publishes the universe it subscribed to, a quoting engine decides against exactly that universe, and a marker scores the decisions against the same tape. No orders were placed and no capital was at risk. That design removes execution slippage and rejection from the results and cannot capture the effect our own resting size would have had on the book.

It is a single-venue quoting measurement over two overnight runs, one of them a weekend, in a window where US same-day sport resolves. Weekday flow runs on a different schedule. The screen's A and C group results come from larger cross-sectional samples but are still one venue and one period. Opinion and Predict Fun are reconnaissance only; nothing on either was tested, and their absence from the verdict column is deliberate.

A negative result survives those limits comfortably, which is the one genuine advantage of screening for what does not work. A positive one would not, and none of the positives here are claimed as tradeable.