Footfall Accuracy, MAPE and Confidence Intervals: What 'Accurate' Really Means

Ask a people-counting vendor how accurate they are and you will hear "98%", "99.5%", sometimes a confident "99.9%". Ask to see the evidence and the room goes quiet. Footfall accuracy is a real, measurable thing — but only when it is defined against a known ground truth and reported with its method and its margin, and almost none of the headline percentages you will see meet that bar. This is the shared engine-trust explainer for the whole counting engine, written here for landlords and brokers who have to judge a number's trustworthiness. Because it is shared methodology rather than FootfallCert-specific content, the canonical, most detailed version lives on our sister site StreetProof; this page summarises it for the property audience.

Key takeaways

• Accuracy only means something measured against a ground truth you can inspect — a bare percentage is a claim, not a measure.
• MAPE (mean absolute percentage error) is a more honest headline than "accuracy" because it describes typical error, not a best case.
• A confidence interval describes the uncertainty in this estimate and widens honestly when less was observed.
• Ground-truth counting sees the specific doorway a panel model can only infer from a sample.
• The trustworthy signal is not a big number — it is disclosed method, inspectable test data, and visible error bars.

Footfall accuracy is a comparison, not an adjective

A count is "accurate" to the degree that it matches the ground truth — the actual number of people who crossed the line, established by careful human labelling of test footage. That means accuracy is always a comparison: counted-versus-true on a dataset you can examine. Strip away the ground truth and "99% accurate" says nothing, because there is nothing for the 99% to be relative to.

This is why the first question in how to audit a footfall claim is "how do you know?" A provider who can answer shows you labelled footage, the counts their engine produced, and the gap between the two. A provider who cannot is quoting a number with no denominator.

MAPE: a more honest headline than "accuracy"

The counting field's cleaner metric is MAPE — mean absolute percentage error. Instead of a single triumphant accuracy figure (which invites cherry-picking the best clip), MAPE reports the average size of the error between the counted number and the true number, as a percentage, across a test set. If an engine's MAPE is, say, in the single digits on representative footage, that tells you the typical miss is small — and, importantly, it is a number you can reproduce by running the engine against labelled ground-truth video yourself.

Two honesty notes:

• MAPE depends entirely on the conditions of the test set. Error on clean, well-lit, moderate-density footage is not the same as error in a dense night-time crowd. A credible MAPE comes with the conditions attached.
• MAPE is about the counting method, not about whether your particular week was typical weather. Method error and representativeness error are different things, and blending them into one reassuring percentage is how numbers get oversold.

Confidence intervals: the uncertainty in this number

Where MAPE describes the method in general, a confidence interval describes a specific estimate in particular. When a certificate projects observed footage to a full period, the interval is the range the true total is reasonably likely to occupy. It answers "how firm is this figure?", which is exactly what a tenant or bank wants to know.

The mechanics, kept plain: the counting variation in a total scales roughly with the square root of the count, so the engine sizes the interval accordingly (the shared methodology uses a dispersion factor and a standard error of the form two times the square root of the dispersion times the count, then reports a 95% interval). The practical consequences are the two things you actually need to remember:

1. Observe more, and the interval narrows. A well-covered seven-day window projects tightly. A short clip projects wide.
2. The interval models counting noise, not the calendar. It does not, on its own, promise the observed week was a typical week of the year — seasonality is flagged separately, never hidden inside the interval.

That is also why the engine attaches a confidence label: high for a well-observed multi-day window, low for a brief spot reading that it refuses to project into a full day. The label carries the representativeness warning the interval deliberately does not. See how this surfaces on the document in how to read a Footfall Certificate.

Ground truth versus panel data

Much of the "footfall accuracy" argument is really an argument about method. Two families dominate:

• Panel / mobile-location estimates model total traffic from a sample of location-enabled devices. Strong at catchment and zone scale; weak at the resolution of a single unit, where they can struggle to separate one side of a street from the other and where a thin local sample makes the inference wobble. It is, in effect, an estimate built on an estimate.
• Ground-truth video counting watches the actual pavement and counts each crossing. It does not infer the street from phones; it counts the street. That is what lets it answer the per-doorway question a lease turns on.

Neither is "the accurate one" in the abstract — but for certifying the footfall past a specific commercial unit, a direct count is measuring the thing itself while a panel model is estimating it from a proxy. The trade-offs are laid out in full on the canonical StreetProof accuracy and methodology research, and the risk of leaning on an un-checkable estimate is covered in why you should never trust an unverified footfall number.

What actually signals a trustworthy number

Put the vocabulary together and the test is simple. A trustworthy footfall figure shows you:

• the ground truth it was validated against, and lets you inspect it;
• an honest error metric (MAPE with its test conditions), not a lone marketing percentage;
• a confidence interval on the specific estimate that widens when less was observed;
• a confidence label that refuses to over-project; and
• a public verification page so all of the above can be checked, not merely asserted.

Notice what is not on that list: a single, enormous, evidence-free accuracy claim. The paradox of "99.9% accurate" with nothing behind it is that it is itself an unverified number — the exact thing verified footfall exists to replace.

Because this is shared engine science, the deepest version — validation datasets, the full derivation, the GDPR and privacy stance — is maintained once on StreetProof and linked, not duplicated, across the brands. To see how these principles are baked into the property artifact, start from the Footfall Certificate guide; when you want to produce one, the pricing page is the place to begin.

Frequently asked questions

What does footfall accuracy mean? How close a count is to the real number that passed — the ground truth. It is meaningful only when measured against an inspectable ground truth and reported with method and margin.

What is MAPE? Mean absolute percentage error — the average size of the gap between counted and true numbers, as a percentage. It is more honest than "accuracy" because it describes typical error and can be reproduced against labelled footage.

What does a confidence interval add? It expresses the uncertainty in a specific estimate — the range the true value likely sits in given how much was observed — and it widens honestly when less footage was seen.

Why is a 99% claim with no evidence a red flag? Because accuracy only means something relative to a ground truth you can inspect. No test data, no method, no error bars means an unverifiable number.