How much does steps have to change before it stops being noise? The threshold from 712 nights

The minimum measurable change in a 28-night window is 32.5%. Below that, the data cannot distinguish a real shift from the ordinary background variation of a single body. This threshold comes from 712 nights of continuous tracking — one person, one device, one protocol throughout the entire period. If the difference you observe between two same-length blocks of your own record falls short of 32.5%, the system has no basis for calling it signal.

Why the threshold is this high

The answer begins with the coefficient of variation: 54.61%. That is the relative dispersion of daily step counts across the full record. The median over 712 nights was 5973.5 steps per day. On any given day, however, steps could deviate by more than half that typical value without any deliberate change in behavior. A work-from-home day, a travel day, a rainy afternoon can shift the count as much as any intentional intervention.

This is not noise introduced by measurement error. It reflects the nature of the variable itself. Steps are a behavioral metric — responsive to schedule, environment, transportation, and weather, not only to the internal state of the body. High behavioral variability translates directly into a high detection threshold.

The serial correlation between consecutive days (ρ) was 0.127: effectively zero. Today's step count carries almost no information about tomorrow's. There is no day-to-day inertia. That is different from more stable physiological signals — resting heart rate, for instance — where a high value today makes a similar value tomorrow more probable. For steps, the context of each day outweighs the trend of the week before.

A high CV combined with a near-zero ρ produces a relative coefficient of variation (RCV) of 151.37%. This is the input the formula uses to project thresholds across observation windows. The higher the RCV, the larger the change required to rise above the noise floor with confidence.

Thresholds by observation window

Longer windows reduce the threshold because more nights average out the day-to-day noise. The thresholds for this body and this metric, calculated from the 712-night record, are:

Observation window	Minimum detectable change
28 nights	32.5%
56 nights	22.98%
84 nights	18.76%
168 nights	13.27%

The practical reading is direct: if step count averages shift by less than 32.5% between two 28-night blocks, the system cannot separate that result from the normal variation of the record. This is not a verdict on whether an intervention worked or failed — it is the minimum condition for making that assessment at all. Without the threshold, any observed change can be read as signal or as noise depending on which period is examined.

Extending to 56 nights brings the threshold down to 22.98%. At 168 nights — six months of continuous recording — it drops to 13.27%. The signal becomes more readable over time not because the body changes less, but because the accumulated nights reduce the relative weight of any single noisy day.

How the number was built

The record consists of 712 nights of continuous observation on a single subject, using the same device throughout the entire period. The full dataset covers 31 metrics and contains 18483 measurement points. Steps are one of those 31 variables.

The method is identical for every metric: compute the CV of the full series, estimate the serial correlation between consecutive days (ρ), derive the RCV, and project thresholds for the standard observation windows of 28, 56, 84, and 168 nights. The threshold is the minimum relative change that, given the historical noise of that metric in that body, allows a confident statement that something different is occurring.

The ρ value of 0.127 has a direct implication: because steps do not carry forward from one day to the next, the day-to-day noise can mask a real directional change for weeks. This is why even at 28 nights — the minimum window that yields a reliable threshold — the required change is still high: 32.5%.

Why this number comes before any interpretation

The question "did I change something and did it shift my steps?" depends on a prior question: does the observed shift exceed the noise floor of my own record? Without that filter, the interpretation varies with the period chosen, not with the data.

This is the gap the record is designed to close. A population mean from another study — measured with a different device on different people during a different period — cannot serve as a personal baseline. The differences are not separable. A result that looks significant against an external reference may be indistinguishable from noise within the same body's own history.

The 650-night record of the founding subject represents the first demonstration that this personal baseline is buildable with a consumer device, without a laboratory. The 712-night dataset extends that record with continuity. The threshold it produces is the best available number today for that body. As the dataset grows, the threshold will be recalculated.

The limit

This record does not say how many steps are appropriate for any goal. A 32.5% threshold is a methodological minimum for detecting a distinguishable change in steps — not a target for physical activity. Crossing it does not mean the change is beneficial or durable: it means the system can separate signal from background noise with the available data.

The thresholds are specific to this body and this device. The CV of 54.61% and the ρ of 0.127 are the signature of the measured subject during these 712 nights. A second subject tracked with a different device would have a different CV and different thresholds. It is not possible to compare absolute step values between subjects using different devices: the sensor, the capture window, and the algorithm differ, and those differences cannot be disaggregated.

The dataset covers 712 nights and closes in 2026. No second-subject data exists for this metric within this record. The threshold could shift if the subject's behavior changes structurally — a relocation, a job with different physical demands — because that would alter the CV of the full series.

The question this record does not answer: what is the step threshold for a second body tracked under comparable conditions with the same device? That number does not exist yet. It is the inherited gap from this record — and the measurement the next entry closes.

---

Mental Work — all protocols · related file: magnesium-for-sleep · open science — the dataset