Asking It Backwards: Find the Banned Accounts First, Then Look at Their Data

How this was made (method & sample) Source: this site's own accumulated telemetry snapshot database — 2,379 accounts and the per-match analysis of 70,578 matches, of which 796 accounts carry an official ban status (Innocent / TemporaryBan / PermanentBan). Every metric uses this site's existing definitions (the same ones on your own pages); nothing was invented for this article. All statistics are non-parametric (Cliff's delta, Mann–Whitney, leave-one-out) because the positive sample is small — too small to claim an accuracy figure, large enough only to talk about how big the gaps are. Captured 2026-08-04. No account is named, no account is listed, and no individual suspicion score is published — reasons in the final sections.

Last time we used raw numbers to flag 13 accounts that looked too good to be human, then checked the official ban record one by one: 0 banned. The conclusion was that a clean banType is not innocence, and catching cheaters with data has a hard limit.

That article asked: pick the suspicious ones first, then see whether they got banned. This time we inverted the whole thing:

Find the accounts the developer has already banned, then look back at what their data looked like before the ban.

Same tools, same database — only the direction of causality flipped. The result is very different.

Why asking it backwards is a completely different question

Starting from a suspicion list runs into two dead ends: a cheater may simply not be caught yet (so there's no ban record to find), and you never learn how many real cheaters your filter threw away. Starting from known banned accounts has neither problem — the official ban record is the answer, and all we have to do is look back at how those people behaved.

The control group matters even more. Our ban statuses are only queried for accounts that have appeared on a leaderboard, so the 778 "not banned" accounts we compare against are all leaderboard-grade players. This isn't the easy comparison of "cheaters vs ordinary players," where any threshold separates them and nothing is learned. It's the hard question:

Can data tell a cheater apart from someone who is genuinely very good?

We nearly got fooled by a trap

The first pass looked great: banned accounts had more kills per match and higher damage. But something was off — the banned group's median "human kill share" was 100%, against 93.3% for the clean group.

What does that mean? Ranked has no bots. The banned group plays noticeably more ranked, and ranked has stronger opponents and denser action, which by itself inflates damage and kill counts. Left alone, what we'd be measuring is "this group plays more ranked."

The obvious fix was to split by mode, but the data wouldn't allow it: the field recording match mode was added later and was never backfilled — 311 of the banned group's 415 snapshots are empty, leaving 3 comparable accounts, which is statistically meaningless.

So we used a proxy instead: "almost no bot kills" ≈ "this account mainly plays ranked", applied identically to both groups before recomputing.

The result is the single most worthwhile thing in this article: after controlling for it, the gaps got bigger, not smaller. The mode mixture was diluting the signal, not manufacturing it.

Result: 11 banned accounts vs 310 clean strong players

All figures below are after mode control. Each account contributes its most recent 14 matches, using exactly the thresholds behind this site's percentile panel.

"Separability" reads simply: pick one banned and one clean account at random; the probability the banned one has the higher number. 0.5 is a coin flip; the closer to 1 or 0, the better that metric distinguishes the two groups.

We also ran a leave-one-out check, recomputing with each banned account removed in turn. Headshot-rate separability moved between 0.770 and 0.846, and the other metrics were similarly stable. That means the gap is a property of the whole group rather than one or two extreme accounts propping it up — the first thing you should ask when a sample is 11.

The most interesting finding runs the other way

The headshot rate and long-range share catch your eye first, but the one that makes you stop is the death-by-third-party rate:

• Clean strong players: 14.3% of deaths come from a third party (you finish a fight at half health and a third squad cleans you up)
• Banned accounts: 5.0% — under a third of that

Third-partying is PUBG's most classic way to die, and it is largely not a mechanical problem but an information problem. Whether you beat the squad in front of you is skill. Whether you know another squad is sitting on the hillside beside you is information.

So the gap says: these accounts rarely get poached. And the most direct explanation for "rarely poached" is not steadier aim — it is seeing what other people can't.

Put differently, the most pointed signal may not be "he shoots too accurately" but "he knows too much." That is the direction we intend to dig next.

The metric with no difference at all is why we trust the others

Knock conversion rate (how often a knock actually becomes a kill) is essentially identical between the groups: 76.9% vs 73.1%, separability 0.565 — basically noise.

That matters, because it shows we didn't cherry-pick a pile of metrics and declare them all significant. Whether a knock converts is a function of your team and the situation: is a teammate nearby, is a third squad pushing, will the victim's teammate revive them. It isn't a question of whether one individual's play is plausible, so it failing to separate the groups is exactly right.

Conversely: if even a metric like that came out "significant," you should suspect the analysis itself.

Combining the signals: usable as a filter, not as a verdict

We combined four signals into a single score (headshot rate, long-range kill share, damage per match, plus the inverted third-party death rate) and asked how many banned accounts you'd cover by checking only the top scorers:

Checking just 10% of accounts covers more than half the banned ones. Data works as a cheap filter, shrinking the set worth looking at closely by an order of magnitude.

But note the other half of that sentence: a filter is not a verdict. The vast majority of that top 10% are accounts that were never banned — they are simply good. Which is the next section.

Being honest: what this data proves, and what it doesn't

Every number here was genuinely measured, but they support far narrower conclusions than they appear to:

• The positive sample is 11 accounts. This is a study of how big the gaps are, not an accuracy claim, and nowhere near enough to train any model. Treat any "XX% detection rate" claim built on a sample like this with suspicion.
• The combined score's separability is an optimistic upper bound. It was computed on the same data it was derived from; there is no independent validation set.
• The control group is "not banned yet," not "verified clean." Undetected cheaters are certainly mixed into it — which makes us underestimate the gaps, but also means the labels themselves carry noise.
• The ban reason is unknown. Official ban status doesn't say whether it was cheating, teaming, or verbal abuse, so non-cheating bans may be in the banned group.
• The most fundamental one: these metrics cannot separate "cheating" from "simply better." We showed the two groups genuinely play differently. We did not show that cheating is the cause of the difference. A high headshot rate can be a cheat, or it can be good aim.

That last point is exactly why we will not publish a "suspicion score" against any named account, and don't intend to. A statistical outlier deserves suspicion, but suspicion is not evidence — the same position as last time, held more carefully now that the data behind it is stronger.

We tried the next step and hit a wall — an informative one

"He knows too much" is a direction you can keep measuring. PUBG telemetry records the position time-series of everyone in the match (close to 100 players in normal mode, around 60 in ranked), so we ran this test:

Every time someone lands a shot on an opponent, look back at the period before contact and rank how much they moved toward that specific victim against every other enemy alive at the same moment, at the same distance.

Using same-match, same-moment, same-distance enemies as the control absorbs the "everyone is running toward the safe zone" confound. We analysed 373 matches. The result: no difference between the groups (separability 0.478 — a coin flip).

But that "no difference" has two pitfalls, and we fell into both:

First, the metric itself was broken. The original observation window ran right up to the moment of firing, and 44% of engagements scored a perfect result — because "did you move toward the person you later shot" is true of everyone; it's nearly tautological. Moving the window back to 60 seconds before contact turned it into a real question ("before they ever made contact, were they already heading for that person?"). Fixed, the answer was still no difference.

Second, and more embarrassing: we committed the exact error described in the first half of this article. Those 373 matches were not filtered by mode — a mixture of normal, ranked, team deathmatch, even the training range. Normal mode has close to 100 players and ranked around 60, and the headcount directly determines how many others are available as controls in the same distance band. Worse, the banned group plays proportionally more ranked — the very confound we worked so hard to control in the first half walked straight back into this test.

Rerunning with the mode filter in place:

Control the confound and the sample collapses. None of the three runs showed the banned group scoring higher, but after mode control only 3–5 accounts remain, which supports no conclusion at all. The honest statement is one sentence: there is currently no evidence that banned accounts differ in how they move toward targets — which is not the same as proving they don't.

Even so, setting it beside the earlier table is interesting:

That points to an unproven inference: an information advantage may show up on defence rather than offence. Seeing what others can't doesn't make you hunt differently — but it does mean you rarely get ambushed.

If that holds, the next thing to measure isn't "how they approach you" but "how they avoid you." The same dataset can answer it. That would be the next article — assuming we can gather enough sample.

(One hard limit worth recording: position data is logged only once every 10 seconds. That resolution is enough to see movement trends on a tens-of-seconds scale, and nowhere near enough to analyse turning, pre-aim or reaction time — the details that would actually distinguish a cheat. Those live only in the video, not in anything the official API returns, which is why we believe pure data has a ceiling.)

Want to see your own numbers? Look yourself up on L!NCHPIN — the same metrics used here (headshot rate, long-range share, third-party death rate) are on your own percentile panel.

Data is a single snapshot taken 2026-08-04, drawn from this site's own telemetry snapshot database. Ban status changes with each official ban wave, so these figures do not update live. No account is named and no individual ranking, account information, or suspicion score is disclosed.