What 90 seconds of game history actually tells a HUD

Most HUD overlays in poker are showing you a window that's roughly 90 seconds wide, and a lot of players treat that window like a verdict. It isn't. What 90 seconds of game history actually tells a HUD is closer to a weather report than a character reference — enough to flag that something's happening, not enough to tell you what kind of player you're sitting with.

The 90-second figure isn't arbitrary. It's roughly the span you get when a HUD defaults to session stats or a rolling recent-hands view on a single table, and at typical online speeds that's maybe 25 to 40 hands depending on how fast the table's running. If you're four-tabling, the same window covers more hands across the board but fewer of them on any one table. Either way, the sample is small enough that a single hot streak or a run of junk cards dominates it. That's the core problem, and it's worth pulling apart properly.

What a 90-second window can legitimately tell you

Short-term stats are good at two things: flagging sudden deviations and describing the current table dynamic. They're bad at almost everything else.

Volume and aggression, not style

Over 25–40 hands, you can reasonably read whether a player is currently involved in a lot of pots or very few, and whether they're raising or calling. That's about it. VPIP and PFR over that sample will swing wildly on card distribution alone. A tight player who picks up three premium hands in 30 deals looks loose; a loose player who gets dealt 7-2, 9-4, J-3 for two orbits looks like a nit. Neither reading is wrong exactly — it's just describing the last 90 seconds, not the player.

What you can trust is the direction of change. If a player's session VPIP was sitting at 22% twenty minutes ago and the recent window shows 60%, that jump is information. The absolute number is noise; the delta is signal.

Table texture

This is where a short window earns its keep. If the last 30 hands have gone multiway to the flop with little preflop raising, that tells you the table is passive right now, and passive tables reward different adjustments than aggressive ones. You're not profiling individuals, you're reading the room. A HUD showing recent table aggression is doing honest work here because it's aggregating across everyone, which smooths out the per-player variance.

Timing and bet-sizing patterns

If your HUD tracks bet sizing or timing tells over recent hands, a 90-second window can catch a live pattern — someone who's suddenly min-betting every flop, or tanking before folding. These are behavioural flags, not statistical conclusions, and they're worth more than the percentage stats at this sample size because they don't require a large denominator to mean something.

Where the 90-second read falls apart

The maths is unforgiving. To distinguish a 25% VPIP player from a 30% VPIP player with any confidence, you need hundreds of hands, not dozens. At 30 hands, the confidence interval around a VPIP estimate is so wide that the two players are statistically indistinguishable. You're not reading a stat; you're reading a coin that's landed heads four times.

The failure mode is confirmation bias. You see a recent window showing 55% VPIP, decide the player is a maniac, and then interpret every subsequent hand through that lens. When they fold three orbits in a row, you don't update — you assume they're trapping. The HUD didn't lie to you; you over-read it.

There's a second failure mode specific to fast-fold and short-handed formats. In Zoom-style pools or 6-max tables that break and reform quickly, 90 seconds might be 15 hands against opponents you'll never see again. Session stats there are nearly meaningless for profiling, and the only defensible use is the table-texture read above.

How to use the window without getting burned

The practical fix is to treat the short window as a change detector and the long window as your baseline. Most HUDs let you display both — session stats alongside a larger sample, or recent-hands stats next to lifetime figures. The 90-second view should only ever answer the question "has something shifted?" It should never answer "what kind of player is this?"

A rough hierarchy that works:

• Under 50 hands: ignore individual stats entirely; use only table-level reads and behavioural tells.
• 50–200 hands: treat stats as weak priors, heavily discounted, and let table dynamics override them.
• 200+ hands: now the numbers start carrying weight, and the recent window becomes useful as a deviation check against that baseline.

The other discipline is noting why a stat looks unusual. A 70% recent VPIP could mean the player is on tilt, could mean they've been dealt good cards, or could mean they've just changed gears deliberately. The HUD can't tell you which. Only the hands themselves can, and that means watching the action rather than the overlay.

The number that should keep you humble

Here's the anchor worth pinning to your monitor: at a 30-hand sample, the 95% confidence interval around an observed 30% VPIP spans roughly 15% to 48%. That's not a typo. Your "loose" read could belong to a player who's actually tighter than average, and the HUD has no way of telling you. You need somewhere north of 300 hands before that interval tightens enough to separate a 25% player from a 30% one with any reliability — and 300 hands at 90 seconds a pop is well over an hour of continuous play against the same opponent, which in most online pools you simply won't get.

Which raises the question the HUD industry hasn't really answered: if the default recent-hands window is statistically useless for the thing most players use it for, why is it the default? There's a case that overlays should ship with the long-sample view front and centre and the 90-second window buried as a diagnostic tool, rather than the reverse. Until they do, the discipline is on you — and the players who profit most from HUDs tend to be the ones who trust the 90-second read least.

And as always, if you're playing long enough to build a 300-hand read on anyone, you're playing long enough to need a break. Set a session limit before you sit down, not after the stats start looking interesting.