Step Counting Accuracy: Phone vs Wrist

Over three days of ordinary life, measured against an ankle-worn reference device, a smartphone missed 27% of its wearer’s steps. A consumer wrist tracker on the same people over the same days missed 22 to 23% [1].

That gap is four percentage points. Your users experience it as proof that one device works and the other does not, and your support queue treats the watch as ground truth. Neither position survives the data.

What free-living validation actually found

Every device tested undercounted, and the spread between them was narrower than the error inside any one of them.

The study put 29 adults through three consecutive days of normal life wearing several devices simultaneously, with an ankle-mounted StepWatch as the criterion [1]. The participants were active, averaging close to 9,700 steps a day, so this is not an edge case of sedentary behaviour.

Two things in that table matter more than the headline.

The first is that the research-grade wrist device is the only one that comes in appreciably better, and it does so while overestimating during sedentary periods. It is not simply more accurate; it fails in a different direction, and it is not something any of your users owns.

The second is that hip placement, the classic pedometer position, did no better than a phone in a pocket. Placement folklore does not survive contact with free-living data.

Why everything misses, and it is the same reason

The error concentrates in short, slow bursts of walking, which is what most daily movement consists of.

The authors attribute the high error rates to undercounting during brief, intermittent bouts of ambulation [1]. Ten steps to the kettle. The walk from desk to printer. Crossing a room. These are not marginal contributors to a daily total, they are the bulk of it for anyone who is not deliberately exercising.

A separate laboratory study makes the mechanism explicit. CADENCE-adults tested 21 wearable technologies across 258 adults on a treadmill, with steps directly observed and hand-tallied as the criterion [2]. Accuracy tracked speed almost perfectly:

• Slow walking, 0.8 to 3.2 km/h: mean absolute percentage error of 40%
• Normal walking, 4.0 to 6.4 km/h: mean absolute percentage error of 7%

The authors state that accuracy was compromised across slow speeds for all wearable technologies tested. Not the cheap ones. All of them.

That single pair of numbers explains most of what your users report. A step detection algorithm looks for a periodic acceleration signature. Three steps at 1.5 km/h to reach the fridge does not produce a periodic anything. The signal genuinely is not there, on any limb.

The gap between the quoted figure and the lived one

Laboratory accuracy at normal walking speed is roughly four times better than free-living accuracy, and the laboratory number is the one that gets published.

Put the two studies side by side. A wearable at a normal treadmill pace averages 7% error. The same class of device across three days of real life averages 22 to 29%. Nothing is wrong with either measurement. They are answering different questions, and only one of them resembles a user’s day.

This is worth holding onto when a vendor quotes an accuracy figure, and worth holding onto when your own QA walks a hundred steps down a corridor to check the integration. That test will pass. It is a normal-speed continuous bout, which is the condition every device handles well and no user spends their day in.

What this breaks in a product

Three specific failures follow, and all of them are common.

Summing sources. A user with a phone and a watch generates two overlapping records of the same walk. Adding them does not improve the estimate, it roughly doubles it. This is the most frequent cause of implausible step totals in health apps, and it requires provenance-aware deduplication per time window rather than arithmetic. Provenance metadata is the only thing that makes this tractable, and it is the part most integrations skip.

Absolute targets. A 10,000 step goal is not a fixed thing. On a source undercounting by 27% it is a materially harder goal than on one undercounting by 22%, and harder again for a user whose walking is slow and fragmented, which correlates with age and with several clinical conditions. The target silently changes difficulty depending on who the user is and what they carry.

Cross-source comparison. Leaderboards, team challenges and any social feature that ranks users against each other will rank them partly by device. A colleague with a research-grade tracker and a colleague with a phone in a bag are not competing on the same measurement.

This is the same shape as sleep stages without a wearable, where the architecture number is inference and the timing is solid. The pattern is consistent: features that depend on an absolute count inherit the full error, while features that depend on relative change within one consistent source largely do not. A user walking 20% more than their own four-week baseline is a real finding, even if both numbers are 25% low, because the error is broadly systematic for that person and that device.

What to do instead

Pick one source per user per time window and stay with it, rather than blending. Consistency matters more than picking the theoretically better device, because a change of source mid-history produces a step in the data that looks like a behaviour change and is not.

Frame features around change rather than level. Trends, streaks measured against personal baselines, and week-over-week comparison all survive a systematic 25% undercount. Absolute thresholds and cross-user comparison do not.

Show the provenance when the number is prominent. A user who can see that the figure came from their phone rather than their watch has an explanation for the discrepancy. A user who cannot has a bug report.

And treat any step count as an estimate with a confidence interval you cannot display but should reason about. The honest internal framing is not “the user walked 6,214 steps” but “the user probably walked somewhere between six and nine thousand steps, and did more walking than yesterday.”