AI-generated research, checked but not peer-reviewed.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

E180　 KILL　 Below costsTested · October 2026

Copying skilled prediction-market wallets loses 1.3 to 2.6c a share

On a prediction market, copying top-ranked wallets' buys five minutes later lost 1.3c per share (interval -4.5 to +2.0); 82 wallets picked after the fact made +1.6c at their own prices, which a follower could not get.

prediction-markets copy-trading wallet-skill persistence latency sports-contracts

Readable summary Full research record JSON recordFeedback on E180

A common idea says to find prediction-market traders with a proven record and copy their buys. We ranked wallets on early data, then in later months copied the top 200 five minutes after each buy of $100 or more, holding to resolution. The pre-registered copy lost 1.3 cents per share (95% interval -4.5 to +2.0), and only 17 of the 200 wallets had a qualifying buy. Copying 1,005 wallets with strong statistical evidence of skill lost 2.1 cents, and copying everyone lost 2.3. A group of 82 wallets, chosen after the fact, earned +1.6 cents at their own prices, but within seconds the price moves to where a follower enters.

Pre-registered copy, 5 min delay-1.3cper share, 95% interval -4.5 to +2.0; 546 copies in 424 events from 17 wallets

Copy 1,005 well-evidenced wallets-2.1cper share, 17,967 copies; copying every wallet lost 2.3c

Post hoc top group, own fills+1.6c82 wallets (36 traded in test), 9,291 trades, t = 4.5

Follower entering 5 min later-2.2csame group; entering at the next print (about 2 s) gave -0.5c

01 · The idea being tested

Some traders on prediction markets are consistently better than the crowd, so copying their buys after they trade captures their skill.

Why people believe it

Trades on an on-chain market are public and tied to a wallet. People argue that past results show who has an information or modelling advantage, and that a follower who buys the same outcome shortly after can share the gain without doing the research.

Why it would persist The argument, as stated before testing

People argue that skill is scarce and slow to copy at scale, and that the best wallets trade in sizes too small for the market to move on, so a delayed follower would still capture most of the edge.

02 · How it was tested

We collected the public trade tape, with the buyer's wallet, from 48 hours before to 8 hours after the scheduled start of each of 51,479 yes/no markets on match winners (soccer, esports, tennis, US leagues) and on words said in speeches, with scheduled ends from 2026-01 to 2026-10. Each wallet was scored on trades from the first period only, inside that window. In the later period, every time a selected wallet made its first buy of an outcome worth at least $100, we simulated buying one share of the same outcome 1, 5 or 15 minutes later at the last visible price plus slippage, and held to resolution. Wallets are stored as anonymous hashed numbers.

Pass required

Primary cell (selection B, $100, 5 min delay, decision-time pricing) needs all of: at least 150 independent events; net mean of at least +1.0c per share with event-clustered 95% interval above 0; positive mean in each of the three test thirds; mean above the 95th percentile of the random-wallet control; positive with +0.5c extra slippage; still positive at a 15 minute delay. Any other cell needs |t| of at least 3.4 (Bonferroni over about 60 cells).

Killed if

Any of the pass conditions failing. The primary cell failed: net mean of at least +1.0c with interval above 0, mean above the 95th percentile of the random-wallet control, positive after extra slippage, positive at a 15 minute delay, and positive in each test third (first third -1.7c).

Costs

Venue fee = the market's fee rate x p x (1 - p) per share (0.05 for sports and esports, 0.04 for spoken-word markets when none was listed), 1.25c at p = 0.5, plus 0.5c slippage; stress case +0.5c more. No order-book depth was modelled.

Controls

Same copy rule applied to all wallets' qualifying buys; 200 random draws of 200 wallets with at least 20 training events; the fade of selection B's trades; event-clustered bootstrap (2,000 resamples); test window split 60/20/20 by time. A pre-game-only variant estimated skill only on trades before the scheduled start and copied only before the start.

03 · What happened

Why it was rejected

The primary cell's mean (-1.3c, interval -4.5 to +2.0) was below the +1.0c threshold and below the random-wallet control's 95th percentile (+0.1c); it was also negative with extra slippage (-1.8c), at a 15 minute delay (-1.4c) and in the first test third (-1.7c).

Primary cell: 546 copies in 424 events from 17 wallets returned -1.3c per share (interval -4.5 to +2.0, t -0.8). Test thirds were -1.7c, +0.3c and +0.5c. With 0.5c extra slippage it was -1.8c; at a 15 minute delay -1.4c. The random-wallet control's 5th to 95th percentile was -5.0c to +0.1c.
Only 17 of the 200 selected wallets had at least one qualifying trigger in the test period (first buy of $100 or more, price 0.05-0.95, inside the window), and 1 of the top 50. The top ranks were low-activity wallets with training returns of +50% per dollar (top 200) and +87% (top 50) over a median of 26 to 28 events.
Broader selections lost with high significance: 1,005 well-evidenced wallets -2.1c at 5 minutes (17,967 copies, 4,865 events, t -7.1); top 200 by raw dollar profit -1.8c (37,968 copies, t -10.0); copying every wallet -2.3c (363,902 copies, t -6.1).
Fading 2,433 negative-skill wallets (buying the opposite outcome) also lost: -2.0c (8,751 trades, t -4.3). Copy and fade both lose because the fee, spread and slippage cost more than any difference in skill is worth.
No pre-registered copy cell with at least 50 copies was positive. A pre-game-only variant, where skill is estimated and copied only before the start, lost 2.6c (1,570 copies, t -2.3) and its wallet-skill rank correlation between periods was 0.02 (p = 0.14, 3,800 wallets).
Across 5,967 wallets active in both periods the training-to-test rank correlation was 0.070 (p = 5e-8). In a post hoc cut, the 82 wallets with a training z-score above 3 (36 of them traded in the test period) earned +1.6c per share on their own test fills (9,291 trades, t 4.5, 93% in-play), with no slippage added.
A follower of that group loses the edge as the delay grows: entering at the next print (median 2 seconds later) -0.5c, within 10 seconds -0.7c, within 30 seconds -1.2c, after 1 minute -1.7c, after 5 minutes -2.2c. Signs are the follower's net gain per share bought, after fee and 0.5c slippage; the own-fill figure includes no slippage.
A post hoc re-estimate that floors each wallet's variance at the payoff variance implied by traded prices found no measurable spread in skill across wallets (Q/(k-1) = 0.98 and tau^2 = 0).

Every way of choosing whom to copy lost money

Net return of copying a buy 5 minutes later, per share held to resolution, with event-clustered 95% intervals. The pre-registered selection is the widest and includes zero; the large selections are clearly negative, and so is copying everyone. The dashed pass line is +1.0c.

Chart data (cents per share)

	Mean	95% low	95% high
Top 200 by skill (primary, 546 copies)	-1.3	-4.5	2
1,005 well-evidenced wallets (17,967)	-2.1	-2.7	-1.5
Top 200 by raw dollar profit (37,968)	-1.8	-2.2	-1.5
Pre-game skill only, top 200 (1,570)	-2.6	-4.8	-0.4
Fade 2,433 negative-skill wallets (8,751)	-2	-2.9	-1.1
Copy every wallet (363,902)	-2.3	-3.1	-1.6

The skilled wallets' edge is gone within seconds

The 82 wallets with a training z-score above 3 (post hoc group). The first bar is what they earned on their own fills, with no slippage; the others are what a follower earns entering later, after fee and 0.5c slippage. The edge turns negative at the next print, about 2 seconds later.

Chart data (cents per share)

	Net per share
Their own fill	1.6
Next print (about 2 s)	-0.5
Within 10 s	-0.7
Within 30 s	-1.2
After 1 min	-1.7
After 5 min	-2.2

The top-ranked wallets mostly did not come back

Wallets with at least 20 training events, the 200 selected as top skill, and how many of those had a qualifying trigger in the test period. The pre-registered rule selected low-activity wallets.

Chart data (wallets)

Stage	Count
Wallets with at least 20 training events	41,053
Selected as top 200 by skill	200
With a qualifying trigger in the test period	17

04 · What it means

Why it failed Interpretation · not established

As interpretation, the numbers fit skill that sits in fast, in-play trading: 93% of the skilled wallets' trades were after the start, and their information appears in the price within seconds, so a follower entering after one to five minutes pays more than the edge is worth. Copy and fade both lose about 2c, which points to trading costs dominating any selection effect. The pre-registered ranking may also have rewarded short winning streaks, with a variance estimate too small after all-win records; that is a hypothesis, and the highest-ranked wallets mostly had no qualifying trigger in the test period.

Lesson

A wallet's profit on its own fills is not what a follower earns, because the follower enters at a later price and pays the same costs. For any trader-ranking method, the variance used to shrink skill has to be floored at what the payoffs imply, or the ranking may select lucky short records, and the delay a follower faces has to be measured against how fast the information reaches the price.

Scope of this result

Covers copying first taker buys of at least $100 in match-winner and spoken-word markets on one on-chain prediction market, entered 1 to 15 minutes later (down to the next print for 82 wallets), held to resolution, one share per copy. Entry prices come from the public trade tape without order-book depth.

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

Taker trade tape (48 hours before to 8 hours after the scheduled start) with the taker's wallet id, for match-winner and spoken-word markets on an on-chain prediction market — the venue's public trades API and public market metadata; tick (every taker order); market end dates 2026-01..2026-10; 51,479 markets in 24,691 events; 21.2M trades; 524,000 wallets; public API, no login

Universe

Match-level moneyline markets of esports, soccer, ATP/WTA tennis, MLB, NBA, NFL, NHL and college football, sampled by scheduled end-date day (every 2nd day for esports, every 3rd day for soccer, tennis and the US leagues) plus every spoken-word market in the same window, which was not day-sampled. The sampling was fixed before any tape was read. Training: markets starting 2026-01 to 2026-06 (26,059 markets). Test: 2026-07 to 2026-10 (23,951 markets), after a 7-day gap.

Signal

Per wallet and per event, profit and cost of taker trades held to resolution (buy of outcome k at p earns y_k - p per share, cost p; sells are treated as buying the complement; prices in 0.02-0.98; fee deducted). Skill = empirical-Bayes posterior mean of the wallet's return per dollar (theta ~ Normal(mu, tau^2), fitted by method of moments on 41,053 wallets with at least 20 training events; sampling variance clustered by event). Selections: A = top 50 by posterior mean, B = top 200 (primary), C = all 1,005 wallets with posterior P(theta > 0) of at least 0.95, R = top 200 by raw dollar profit, FADE = 2,433 wallets with P(theta < 0) of at least 0.95, whose complement is bought. A and B also required P(theta > 0) of at least 0.9. Selection A produced 8 copies from 1 wallet and is not reported.

Execution

Trigger: the first taker buy by a selected wallet in a market and outcome with notional of at least X ($100 primary, $1,000 tested) and price 0.05-0.95. Our decision time is the trigger plus D (1, 5 or 15 minutes), only if within 24 hours before to 6 hours after the scheduled start. Entry = last ask-side print of that token in the 30 minutes before the decision time, plus 0.5c slippage (fallback: last print of any side plus 1.0c; otherwise no trade). Hold to resolution, one share per copy. A second pricing variant used the first print after the decision time.

Costs

Controls

Sample split

Train 2026-01..2026-06, 7-day gap, test 2026-07..2026-10. Skill estimated only on training trades whose markets resolved before the test window.

Pass criterion

Kill criterion

Power

The primary cell has an interval half-width of about 3.2c (546 copies from 17 wallets), so it could not exclude an effect of a few cents either way. The broad cells (17,967 to 363,902 copies) have half-widths of about 0.4 to 0.8c and exclude any gain.

2. Result

Observations

546

Instruments

424

Net mean

-1.3

95% interval

-4.5 to +2

Median

—

Worst

—

Finding 1

Primary cell: 546 copies in 424 events from 17 wallets returned -1.3c per share (interval -4.5 to +2.0, t -0.8). Test thirds were -1.7c, +0.3c and +0.5c. With 0.5c extra slippage it was -1.8c; at a 15 minute delay -1.4c. The random-wallet control's 5th to 95th percentile was -5.0c to +0.1c.

Finding 2

Only 17 of the 200 selected wallets had at least one qualifying trigger in the test period (first buy of $100 or more, price 0.05-0.95, inside the window), and 1 of the top 50. The top ranks were low-activity wallets with training returns of +50% per dollar (top 200) and +87% (top 50) over a median of 26 to 28 events.

Finding 3

Broader selections lost with high significance: 1,005 well-evidenced wallets -2.1c at 5 minutes (17,967 copies, 4,865 events, t -7.1); top 200 by raw dollar profit -1.8c (37,968 copies, t -10.0); copying every wallet -2.3c (363,902 copies, t -6.1).

Finding 4

Fading 2,433 negative-skill wallets (buying the opposite outcome) also lost: -2.0c (8,751 trades, t -4.3). Copy and fade both lose because the fee, spread and slippage cost more than any difference in skill is worth.

Finding 5

No pre-registered copy cell with at least 50 copies was positive. A pre-game-only variant, where skill is estimated and copied only before the start, lost 2.6c (1,570 copies, t -2.3) and its wallet-skill rank correlation between periods was 0.02 (p = 0.14, 3,800 wallets).

Finding 6

Across 5,967 wallets active in both periods the training-to-test rank correlation was 0.070 (p = 5e-8). In a post hoc cut, the 82 wallets with a training z-score above 3 (36 of them traded in the test period) earned +1.6c per share on their own test fills (9,291 trades, t 4.5, 93% in-play), with no slippage added.

Finding 7

A follower of that group loses the edge as the delay grows: entering at the next print (median 2 seconds later) -0.5c, within 10 seconds -0.7c, within 30 seconds -1.2c, after 1 minute -1.7c, after 5 minutes -2.2c. Signs are the follower's net gain per share bought, after fee and 0.5c slippage; the own-fill figure includes no slippage.

Finding 8

A post hoc re-estimate that floors each wallet's variance at the payoff variance implied by traded prices found no measurable spread in skill across wallets (Q/(k-1) = 0.98 and tau^2 = 0).

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download the full taker trade tape with taker wallet ids and the market metadata for the market set above; hash wallet ids. 2) Per wallet and event, compute profit and cost of taker trades held to resolution with the fee deducted; keep wallets with at least 20 training events. 3) Fit the empirical-Bayes skill model with event-clustered variances, then form selections A, B, C, R and FADE. 4) In the test period, simulate copies of first buys of at least $100 with the stated delay, entry price rule, slippage and fee. 5) Bootstrap by event, compare with 200 random-wallet draws, and run the delay sweep for the highest-ranked wallets.

Grade

COMPLETE