AI-generated research, checked but not peer-reviewed. It may contain errors.

E710　 KILL　 Below costsTested · October 2026

Out-of-order nested prediction markets cannot be traded

On an on-chain prediction market, 7.9% of nested-question pairs looked out of order at the mid in the first scan, but none ending within 14 days offered an after-fee edge of 1c per share in any of 9 scans.

prediction-markets logical-consistency arbitrage order-books scanner nested-claims

Readable summary Full research record JSON recordFeedback on E710

Some prediction-market events list nested questions, such as "by October 31" and "by December 31", where one answer implies the other. If prices break that order, buying NO on the stronger claim and YES on the weaker one pays at least $1 in every outcome. We scanned all 2,100 active events on live order books, nine times, 27 minutes apart. In the first scan, 7.9% of pairs looked out of order at the midpoint of bid and ask, yet the after-fee profit from pairs with an edge of at least 1c per share that resolve within 14 days was $0 in all 9 scans, against $50 needed.

Executable profit within 14 days$0after fees, edge at least 1c per share; 9 of 9 scans

Pairs out of order at the mid7.9%first scan, inside wide spreads

Required to pass$50per scan, in at least half of the scans; pre-registered

Largest total of any executable violation$9.29any edge and horizon, best pair per event, one scan

01 · The idea being tested

Prediction markets list many related questions, and some of them must be priced out of order, which leaves a risk-free profit.

Why people believe it

An event often holds a ladder of questions that differ in one parameter: "X by October 31" and "X by December 31", "reach 120k" and "reach 110k", "win by more than 3.5" and "win by more than 2.5". If the stronger claim is true, the weaker one must be true. Each question has its own order book and its own few quoters, so people argue the ladder is not kept in order.

Why it would persist The argument, as stated before testing

People argue that the ladders are large and spread over thousands of markets, that each book is thin, and that the profit per pair is too small for professional firms to monitor, so a scan across all of them should find leftovers.

02 · How it was tested

We wrote a scanner that finds families of nested questions inside each event, pairs them so that one claim is stronger than the other, and prices the two-leg trade on the live order books: buy NO on the stronger claim and YES on the weaker one. It deducts the trading fee on each leg and walks down both books while the pair still costs less than $1. The test asked whether this found at least $50 of executable profit per scan on pairs resolving within two weeks.

Pass required

In at least half of the scans, summed walked profit (best pair per event) of violations with edge at least 1c per share and S ending within 14 days is at least $50; at least 3 distinct families across the run; the 10 best pairs survive a rules audit.

Killed if

Otherwise.

Costs

Taker fee on each leg: the market's fee rate times p(1-p) per share. No gas, no capital cost of holding the pair to resolution, no slippage beyond the walked book; these omissions favour the strategy.

Controls

None. The pre-registration also required the ten best pairs of a passing run to be checked against the resolution rules; that step applies only if the gate is met.

03 · What happened

Why it was rejected

In every one of 9 scans the after-fee profit with edge at least 1c and S ending within 14 days was $0, against the $50 gate needed in at least half.

Gate result: summed walked profit of violations with edge at least 1c per share and S ending within 14 days was $0 in 9 of 9 scans; the gate needed $50 in at least half.
First scan: 7.9% of ordered pairs were out of order at the mid-price, inside wide spreads. At executable ask prices after fees, 15 pairs in 8 events broke the order, worth $8.47 in total (best pair per event).
No scan reached $10 even without the 1c or 14-day filters: the total of all executable violations ranged from $3.26 to $9.29 per scan (best pair per event).
A violation of at least 1c per share appeared only in the first four scans, in 1 or 2 events, worth $4.92 to $6.04 per scan. From the fifth scan on there was none. The one in the first scan ends in 104 days; none ended within 14 days or within 24 hours.
Violations that did end within 14 days had edges below 1c per share and were worth at most $0.04 per scan.

From 2,100 events to zero qualifying pairs

First scan, each stage narrower than the last. Many pairs are out of order at the mid, but after fees only 8 events remain, 1 of them has an edge of 1c per share, and it ends in 104 days.

Chart data (events)

Stage	Count
Active events scanned	2,100
Events with several markets that are not a complete set	782
Events with an executable violation after fees	8
... with edge of at least 1c per share	1
... and S ending within 14 days	0

Executable profit never approaches the gate

Walked after-fee profit of all executable violations in each of the 9 scans, with no edge or horizon filter, best pair per event. The gate applied to the narrower subset with edge of at least 1c within 14 days, which was $0 in every scan.

Chart data (USD)

	All executable violations, any edge and horizon
Scan 1	8.47
Scan 2	7.59
Scan 3	9.29
Scan 4	8.29
Scan 5	3.26
Scan 6	3.63
Scan 7	6.67
Scan 8	4.66
Scan 9	4.38

04 · What it means

Why it failed Interpretation · not established

One reading, not tested here: many mid-price violations sit inside wide spreads, where a buyer cannot trade the gap, and quoters may price each rung off its neighbours so that the fee and the spread absorb what is left. The scans show the outcome, not who enforces it or whether the books were wide.

Lesson

Counting order violations at mid-prices measures how wide the books are, not how wrong the prices are. A mispricing test needs ask prices on both legs, fees and depth. Here the same ladder looks 7.9% inconsistent at the mid and at most $9.29 per scan executable at the ask.

Scope of this result

Covers nested families inside single events (deadline, threshold, spread and total ladders) on one on-chain prediction market, in nine scans on one day, with the parser's keyword rules for direction. It does not cover relations between different events, other venues, or other days.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

Active events, their markets and fee schedules — an on-chain prediction market, public event listing API (anonymous); one snapshot per scan; 2026-10; public | Full order books (bids and asks with sizes) for every market in a detected family — an on-chain prediction market, public order-book API (anonymous); one snapshot per scan; 2026-10; public; only scan summaries and violating pairs stored

Universe

All active events at scan time (2,100), of which 782 had several markets that are not a complete mutually exclusive set. Per scan the parser produced about 2,340 to 2,465 candidate sub-families and 47,000 to 51,000 ordered pairs (47,600 in the first scan).

Signal

Family: at least two open markets with an order book in one event, whose questions are identical once digits, month names and currency tokens are removed (same question, only the parameter differs), split into sub-families by the rest of the title. Parameter: a date or a number. Direction by keyword: for dates, "by" or "before" makes the later date weaker, while "through", "until" or "continues" makes it stronger; for numbers, "reach", "above", "over" or "at least" makes the higher value stronger and "dip", "below" or "under" makes it weaker; over/under markets use "over", and spread markets use the named team, with the larger handicap stronger. "On a date" families are disjoint days, not nested, and were excluded. A family with fewer than three tight-book mids is not checked for direction. If more than 60% of the pairs with tight-book mids (spread at most 10c) fall the wrong way, the family is treated as misread and dropped. About 150 to 170 candidate sub-families per scan were skipped for lacking a direction keyword and about 190 to 220 events failed the question match. The parser rules were fixed before the verdict scans.

Execution

For every ordered pair (S stronger, W weaker): buy NO(S) at the ask and YES(W) at the ask. The pair pays at least $1 in every outcome. Profit is realised when S resolves, so the horizon is the time to S's end date. Depth: walk both ask ladders while the marginal set costs less than $1 after fees. Per event only the best pair is counted, because legs overlap across pairs.

Costs

Controls

None. The pre-registration also required the ten best pairs of a passing run to be checked against the resolution rules; that step applies only if the gate is met.

Sample split

Nine scans at intervals of about 27 minutes on one day, after the parser was fixed; they are the scans behind the verdict. The pre-registration planned about 6 hours of scans. Three later scans also returned $0 under the gate and were not part of the verdict. The same events largely recur, so the scans are not independent samples.

Pass criterion

Kill criterion

Otherwise.

Power

A scan covers the full list of active events, not a sample, so it sees every nested family the parser detects at that moment, but the skipped sub-families above were not checked. It does not see mispricings that exist only at other times, and the nine scans span under four hours.

2. Result

Observations

Instruments

—

Net mean

95% interval

Not reported

Median

Worst

—

Finding 1

Gate result: summed walked profit of violations with edge at least 1c per share and S ending within 14 days was $0 in 9 of 9 scans; the gate needed $50 in at least half.

Finding 2

First scan: 7.9% of ordered pairs were out of order at the mid-price, inside wide spreads. At executable ask prices after fees, 15 pairs in 8 events broke the order, worth $8.47 in total (best pair per event).

Finding 3

No scan reached $10 even without the 1c or 14-day filters: the total of all executable violations ranged from $3.26 to $9.29 per scan (best pair per event).

Finding 4

A violation of at least 1c per share appeared only in the first four scans, in 1 or 2 events, worth $4.92 to $6.04 per scan. From the fifth scan on there was none. The one in the first scan ends in 104 days; none ended within 14 days or within 24 hours.

Finding 5

Violations that did end within 14 days had edges below 1c per share and were worth at most $0.04 per scan.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download all active events with their markets and fee schedules. 2) Within each event, group markets whose questions are identical after stripping digits, months and currency tokens, and parse a date or number parameter from each market title; infer direction with the keyword rules and drop families whose tight-book mids contradict it in more than 60% of pairs. 3) Fetch the order books of all markets in the families. 4) For each ordered pair, cost = ask NO(S) + ask YES(W) + both taker fees; walk the ladders while cost is below $1. 5) Keep the best pair per event, repeat every 27 minutes, and apply the gate.

Grade

COMPLETE