We froze OKX's top 50 and watched for seven weeks. The scoreboard says +1.20%; the actual return, −75 basis points

On 4 July we froze a list: the 50 lead traders with the most assets under management on OKX's copy-trading leaderboard, exactly as they stood that day. We wrote the rule before looking at anything — «top 50 by AUM in the first snapshot, untouched even if they drop off the ranking» — and recorded in the same document which test we would run and when: «re-run the persistence test with forward-only data, at 6-8 weeks». Today makes seven. This is that test.

Why freezing the list matters. Back in July we had already analysed the history the exchange publishes, and the numbers looked spectacular: nine statistically significant traders, win rates from 76% to 99%, averages of +466 to +1,192 basis points per trade. But that analysis had a flaw we declared ourselves: the cohort had been picked by assets under management **at the end** of the window. And money arrives after the hot streak. Selecting the ones who had already won and then confirming that they won proves nothing.

The test that does separate skill from luck is persistence: split the window in half and ask whether the good performers in the first half are still the good performers in the second. In July it gave a rank correlation of +0.19 with p=0.215 across 18 traders. It failed the threshold, but with that sample you couldn't claim anything either. That's why we put a date on the rematch.

Today's result, with 4,142 positions opened after the cut-off across 32 traders: **a correlation of +0.042 with p=0.435**. It doesn't fall short. It's gone. With more power and no selection bias, knowing who did well in July tells you nothing whatsoever about who will do well in August.

And one detail deserves emphasis because it's easy to miss: over those seven weeks **bitcoin rose 21.6%**, from $63,151 to $76,801. The cohort was 69% long. Even so, the average directional return — unleveraged, net of 10 basis points of round-trip fees — was **−75 basis points per position**. Being mostly long in a market that rises 21.6% and still finishing negative is not bad luck.

What the copier sees says the opposite. The metric the exchange publishes, leveraged profit ratio, averages **+1.20%** over exactly the same trades. Same people, same sample, two numbers with opposite signs. The difference is leverage, which multiplies the outcome but doesn't create skill: multiply a coin flip by ten and you still have a coin flip.

There's a third thing only your own capture can see, and it may be the most revealing. The exchange's history lists only **closed** positions. We photograph the open ones every hour too. Right now the cohort holds 231 live positions opened after the cut-off: **66% are underwater, with a median unrealised loss of 10.35%**, and the oldest has been open for 47 days. It's the classic pattern: close the winners, which inflates the win rate, and ride the losers, which appear in no statistic until they're realised. Including them, the average moves from −75 to −91 basis points.

It's also worth saying that a single person holds up the average. Without the best trader in the sample, it drops from −75 to −221 basis points. And the median across all positions is **+29 basis points**: a great many small gains and a tail of large losses that eats them. That is exactly what closing fast in the green and holding on in the red produces.

Now the caveats, which go here and not in a footnote. First: our return measure **does not subtract the market**. A trader who is permanently long in a rising window is helped by the tailwind, not by skill. We left it that way because that's how it was written in advance, and because it works against our conclusion rather than for it: even with that help, the average is negative.

Second: positions overlap. With a median duration of about 18 hours and several open at once, the per-position average is good for comparing traders against each other — which is exactly what the persistence test does — but it isn't what someone copying them would have earned. Third: 9.3% of the rows arrive from the exchange with an empty close time on positions its own system lists as closed; they have a valid closing price so we count them, and excluding them makes the result worse (−93 instead of −75), so it doesn't change the reading. And fourth: this is seven weeks of a single, rising regime. What we cannot say is that these traders always lose. What we can: that in seven weeks of clean capture there isn't a trace of persistence.

We won't be chasing the five traders who did come out individually significant. They are precisely what the bias we declared produces, and the persistence test exists to tell a streak from a skill. If we look again in two months and they're still there, that will be a different conversation.

The capture continues, hour by hour, over the same frozen list. Every week adds power and costs nothing. We'll publish the number again whatever it says: the point of writing the test before running it is that you don't get to decide afterwards whether you like the answer. Full methodology, July's pre-registration and the raw output are in the project documentation. None of this is investment advice; our own strategies remain in simulation.