Attendance Count Accuracy: What MAPE and Confidence Intervals Really Mean

Every people-counting vendor waves an accuracy number — "98%," "99.5%" — and every one of them is telling you less than it looks. To judge an attendance count accuracy claim you need three ideas that the marketing banners quietly skip: MAPE, confidence intervals, and the difference between panel data and ground truth. Understand those and you can read any vendor's claim, including ours, for what it is worth. This is a plain-language primer written for event organizers.

This is shared engine-trust content. The full accuracy and methodology handbook lives once on our hub, StreetProof — this page frames it for attendance counting and points you there for the deep detail, which is why it is canonicalised to StreetProof.

Key takeaways

• MAPE (mean absolute percentage error) is a real accuracy measure, but it is an average — it hides how a counter behaves in the crush at your gate.
• A single headline accuracy percentage is a red flag; honest accuracy is reported as per-interval error bars plus a confidence tier.
• Panel (mobile-device) estimates go blind on a specific gate, a free-entry event, and short windows; ground-truth counting sees the entrance you filmed.
• Ask how accuracy is measured and how it holds in dense, dark conditions — not just for one best-case number.

Attendance count accuracy is not one number

The instinct is to want a single figure: "how accurate is it?" But a count's accuracy is not a fixed property of the software. It moves with the conditions. The same engine that nails a steady 3pm market gate will struggle at a packed 9pm festival entrance where bodies overlap three deep. Quoting one number for both is like quoting one travel time for a road regardless of traffic. It is technically a number and practically a fiction.

That is why the honest way to talk about accuracy is conditional: strong here, weaker there, and always with the uncertainty attached. Which brings us to the two tools that make that possible.

MAPE: a real measure, honestly limited

MAPE — mean absolute percentage error — is the standard way to score a counting method against known ground truth. You count a scene where you already know the true figure (say, a controlled gate where every crossing was hand-verified), compare the machine's counts, and express the average gap as a percentage. A MAPE of 4% means the counts were, on average, 4% off the truth.

MAPE is genuinely useful, and we test against it. But read the word mean. It is an average across many conditions, so a flattering MAPE can hide terrible performance in exactly the moment you care about — the peak crush — averaged out by lots of easy, quiet frames. A single MAPE quoted with no conditions is not a lie, but it is not the whole answer either. Always ask: measured against what ground truth, and in what conditions?

Confidence intervals: where the number probably sits

If MAPE describes a method in general, a confidence interval describes your count in particular. Instead of "35,000," an honest count says "34,000–36,500," meaning the true figure very likely sits in that band. The interval widens when the count is noisier — a dense gate, a short sample — and tightens when it is clean. It is the count admitting, precisely, how sure it is.

This is why every GateProof Attendance Report shows error bars per interval rather than a single confident total, and why dense-crowd stretches get wider bars, flagged plainly. A report that gives your crush hour the same false precision as your quiet open is hiding the truth; one that widens the band is telling it.

Two different uncertainties

There is a subtlety worth holding onto, because it trips up almost everyone. There are two separate uncertainties in a projected attendance figure:

1. Counting noise — natural variation in the count. This is what the confidence interval captures.
2. Representativeness — whether the footage you observed actually stood in for the whole event. A confidence interval cannot fix this, so it is reported separately as a confidence tier.

A count can therefore have a tight interval and a "low confidence — indicative" label: the arithmetic was precise, but the sample was too short to speak for the day. We keep these apart on purpose; blurring them is how short clips get sold as settled totals. The mechanics are in how sampling turns a few hours into a full-event estimate.

Panel data vs ground truth — where mobile estimates go blind

You will also meet a very different kind of "attendance" figure: panel data, sold by mobile-location providers. They sample a slice of phones, then scale up with modelling assumptions to estimate how many people were in an area. For wide, regional trends this is powerful. For your event it has three blind spots:

• The specific gate. Panels estimate an area, not the entrance you filmed. They cannot tell your main gate from your service road.
• Free-entry and non-phone crowds. Anyone with location off, a flat battery, or no smartphone is invisible, and the scaling assumptions vary by venue.
• Short windows. Panels smooth over hours; they are weak at "how many crossed between 6 and 7pm."

Ground-truth counting is the opposite: it does not model a region, it counts actual crossings at the gate you pointed a camera at. That is narrower — and, for proving your attendance to your sponsor, far more defensible. It is also why an unverified panel figure deserves the same scrutiny as any other; see why you should never trust an unverified number.

How to read any accuracy claim

Next time a vendor — us included — makes an accuracy claim, run it through four questions: How is accuracy measured? Against what ground truth? How does it hold in dense crowds and low light? And does the report disclose per-interval error bars and a confidence tier, or just one headline number? A method that answers all four is measuring. One that offers a single blanket percentage is marketing.

For the counting methods themselves, start at the pillar, how to count event attendance. When you want to see honest error bars on your own event, count your gates with a $199 Event Report at your next event — the report shows you exactly where it is certain and where it is not. On privacy specifically, our approach — counting silhouettes, not people — is documented in full on the StreetProof methodology hub.