Copy-edge

Prediction markets settle in cash against a public outcome, and every fill is visible: wallet, price, size, timestamp. That is a rare thing. In most markets you cannot see who traded, only that a trade happened.

So there is an obvious question. If some wallets are informed, or simply good, and the whole trade history is public, it should be possible to find them from past behavior and ride their future bets. The idea is not new. It is the same logic as following disclosed institutional filings. What is different here is that the record is complete and free.

The point was to find out whether that could be turned into something worth running, or killed cheaply. A validation experiment with a kill switch, not a mandate to make it work. That distinction is the reason it cost a week.

Closing line value measures whether an entry was good at the moment it was made, independently of how the market happened to resolve. For a long on the chosen outcome it is the last traded price before resolution minus the entry price, sign-adjusted for side.

CLV = P_close - P_entry      per entry, sign-adjusted for side and outcome

A wallet qualified only if it cleared every gate. The qualifying set was then ranked on a composite score and cut to a top decile.

1  Track record   traded >= 10 resolved markets AND >= $5,000 notional
2  Closing line   mean CLV >= +0.03
3  Early entry    >= 50% of winning entries placed before the market
                    moved 0.15 toward the eventual outcome
4  Rank           composite score, keep the top decile

One independent copied bet is a unique combination of market, outcome and side that at least one selected wallet entered during the holdout window. That is the count the thresholds are measured on.

Net ROI per bet    >= 0.05     kill if <= 0
Significance       p < 0.05     kill if not significant
Independent bets   N >= 200     kill if unreachable
Mean CLV           > 0

The root cause was not that the selected wallets lost. It was that the rule was too selective to be testable at all. The funnel across 720 training markets:

74,000  wallets in the universe
 1,370  traded >= 10 resolved markets
   391  and >= $5,000 notional
    60  and mean CLV >= +0.03
     4  and cleared the early-entry gate
     1  after the top-decile cut

One wallet. With a sample of one, the return and the p-value are not measurements of anything, and the few bets it did produce lost money. The early-entry gate was the binding constraint, and it stayed binding under four times the data.

That is a narrower conclusion than it looks, and it was recorded as such at the time. Version 1 was dead by its own criteria, but a failure driven by over-selectivity is not a powered disproof of the general idea. So the idea got five more attempts, each pre-registered, each on data the previous version had not touched.

Version 4 is the one worth reading twice. Copying proven specialists returned plus 34 percent, which looks like a finding until you see it rests on 40 bets at p = 0.20. The high-conviction subset, the bets the signal was most sure about, returned minus 25 percent. The signal was inverted exactly where it should have been strongest.

The structural reason it cannot work is that proven winners are episodic. A trader who reliably fades one kind of market is tied to one class of event. A corpus of 1,413 markets yielded roughly 40 copyable future bets. The sample never reaches significance, not because the edge is absent but because the people who have it do not trade often enough to follow.

Version 6 found something. Fading cheap longshots on a corpus of geopolitical markets returned plus 8.4 percent at p = 0.02. Significant, in the right direction, and large enough to act on.

It was re-run, same frozen rule, no retuning, on 1,566 markets from a different regime.

One conflict, retail-heavy      +8.4%    p = 0.02    significant
Other conflicts, 1,566 markets  -1.1%    p = 0.69    gone

The favorite-longshot bias was real and it was local to one emotional, retail-flooded episode. In ordinary markets the favorites were priced fairly. A 75 percent win rate that is still net negative is what fair pricing looks like from the inside.

Without that check this would have been the finding. It cleared every other gate. The check caught a false positive before any money was committed, which is the entire reason the check exists.

On accessible prediction-market data there is no reliable, scalable, prospectively profitable copy or detection edge. What looks like edge is consistently skill, strategy, bots or noise, and the rare genuine winners do not generate enough future volume to follow. The one positive result was an artifact of a single market regime.

That also kills the product behind it. A service promising winning trade ideas, built on an edge that does not generalize, is a tout service. The honest version would be analytics and transparency with no profit promise, which is a different business and not the one being tested.

The cost was compute and about a week. The alternative cost, which is the whole point of running it this way, was learning the same thing with a funded position.