The promise we tested
Every product in this category sells the same promise. Find the wallets that win, put their trades in front of you, and let their skill become your returns. It is an easy promise to make because the first half is testable and the second half never gets tested.
So we tested both.
We pulled 741,322 settled fills from public prediction market data across three categories, measured whether the traders who did well in one set of events also did well in a completely different set, and then measured what happened to someone who simply copied them.
The answers point in opposite directions, and the gap between them is the entire problem with how following is sold today.
First, the part everyone gets right
Skill is real. It persists across events. It is not an illusion produced by a few lucky months.
To check this we needed a design that could not flatter itself. If you rank traders on a set of results and then measure them on the same results, you have proven that winners won. That is a tautology dressed as a finding.
Instead we split each category’s events into two halves. Call them set A and set B. A trader had to have traded in both, with at least five fills in each, to be counted at all. We ranked everyone on set A only, then went and looked at what those same addresses did in set B, which had no part in the ranking.
If skill is real, the set A ranking should predict set B results. If it is noise, it should predict nothing.
Sports covers 400,529 fills across 29 separate events, from 63,298 addresses, of which 838 met the both-halves requirement. Politics covers 274,747 fills across 65 events and 363 qualifying addresses. Crypto covers 66,046 fills across 47 events and 126 qualifying addresses.
A correlation of 0.712 between two disjoint sets of events is not subtle. Restrict it to addresses with twenty or more trades and sports rises to 0.754 and politics to 0.797. More evidence per trader, cleaner signal, exactly as you would expect if the thing being measured is real.
We also removed the World Cup from the sports sample, because a single tournament with long-dated positions can manufacture correlation on its own. Correlation went up, from 0.440 to 0.712. The effect was not a scheduling artifact. It was being hidden by one.
There is a second finding buried in that table that almost nobody talks about.
Being bad is more persistent than being good. In crypto the bottom group went from -77.4% to -76.1%. They barely regressed at all. The top group fell from 55.3% to 20.4%. Losing money is a more reliable trait than making it, which is uncomfortable and also useful, and we will come back to it.
Now the part that gets sold and never checked
If skill persists, copying it should work. That is the whole pitch.
We took the sports data, ranked addresses on set A, and selected the top 64 by ROI. Their average set A return was 56.1%. By any reasonable definition these are the traders a follow product would put in front of you.
Then we followed them. Not on set A, where they were selected. On set B, which had nothing to do with the ranking.
Those 64 addresses produced 10,841 fills in set B. Of those, 860 were buys at prices that could actually be followed, meaning not already pinned against 0 or 1 where nothing is left to capture. We simulated buying each of those 860 positions at the price available zero, five, fifteen, thirty, sixty, and one hundred eighty minutes after the original fill.
Read the first row again. At zero delay the follower makes -0.10%.
Zero delay means you saw the fill and got the exact same price, instantly, with no slippage and no competition. It is a physically impossible advantage and it still does not work. Which means the problem was never latency. Every article about copy trading that ends with “you need to be faster” is solving a problem that is not there.
Both findings are true, and here is how
This looks like a contradiction. The traders are demonstrably skilled. Copying their trades returns nothing. Pick one.
You do not have to. The two measurements count different things.
The correlation test measures ROI per address. It asks whether a person who returned well on one set of events returns well on another. The follow test measures ROI per fill. It asks whether an individual trade, entered at an observable price, made money.
Those come apart when returns are concentrated. If a trader’s year is made by a small number of large, well-timed positions and paid for by a long tail of small losing ones, their per-address ROI can be excellent while their average fill sits near zero. Copying every fill buys the tail along with the wins, at prices that have already moved.
Which points at what the skill actually is.
Their edge is in which markets they enter and when they leave, not in the entry price you can see. By the time a fill is public it has already done its work on the order book. The information that made the trade good was available before the trade, not inside it. Copying the fill copies the residue.
This is also why the bottom group persists so stubbornly. A bad trader is not making one large mistake you could avoid by not copying one trade. They are making a structural error in market selection, repeatedly, and it shows up in every fill they place.
What this means if you are actually trading
Three things follow from this, and none of them is “stop paying attention to other traders.”
A track record is evidence about a person, not an instruction about a trade. The correlation is real. A trader with a strong record across separate events is genuinely more likely to be right than a random address. That fact is worth knowing. It just does not convert into “buy what they bought” without losing everything in the conversion.
Timing signals from fills are worth roughly nothing. We measured this directly at six delays and the best number in the table is -0.06%. If a product’s entire value proposition is showing you fills faster, the data says the speed is not the constraint.
Avoiding bad traders is more reliable than following good ones. This is the finding we did not expect and the one we would act on first. The bottom group’s persistence is higher than the top group’s. If you are going to use other people’s records for anything, use them to filter out the noise before you use them to find signal.
What we are building from it
We run SmartX because of the gap in the middle of this data, not in spite of it.
The category sells the fill. The fill is the part with no edge in it. What has edge is everything around the fill: which markets a trader is actually good at, how large they size when they have conviction, whether their record survives being split in half, and whether any of that matches how you trade.
That is the layer we are building. Not a faster feed of what someone bought, which we have now measured and can tell you does not work. A record you can interrogate, attached to a real position, matched to the way you actually take risk.
A screenshot is not a track record. A fill is not a strategy. And the honest version of “follow smart money” is a great deal more complicated, and a great deal more useful, than a green number in a feed.
Method notes
Fill data comes from public prediction market APIs. Sports covers markets ending between June and July 2026, politics and crypto from September 2025 onward. Addresses needed at least five fills in each half to qualify, and at least ten in each half for the follow test. Only buys were followed; following exits is a different strategy and a different test. Follow prices use the first observed trade at or after the target timestamp, which is a close approximation of what a follower would actually pay. ROI is profit over volume at entry price, not over bankroll.
LIVE ON POLYMARKET