How much does activity score have to change before it stops being noise? The threshold from 711 nights

For a change in activity score to count as more than measurement noise, it has to clear 21.14% when compared against a 28-night window, or 14.95% against a 56-night window. That threshold comes from 711 nights of recording on a single body, with a coefficient of variation (CV) of 11.56% for the metric itself.

What the coefficient of variation of activity score means

CV is standard deviation divided by the mean, expressed as a percentage. At 11.56% for activity_score, the metric moves within a relatively narrow band from one night to the next — that band is exactly what separates ordinary variation from a real change. A low CV means the metric is inherently stable; a higher CV would mean a bigger swing is needed before anything can be called a change at all.

That 11.56% figure is not a short-sample estimate or a theoretical benchmark. It comes from 711 consecutive nights of recording on a single body, with a median of 84 on the metric's own scale. That is the full source row: 711 nights, median 84, CV 11.56%.

That number on its own already answers a question that usually gets skipped: is this metric noisy by nature, or is it stable? At a CV of 11.56%, activity_score sits in a middle zone — not so volatile that comparing any two nights is pointless, and not so flat that a small shift already counts as a finding. That middle zone is exactly why an explicit threshold is needed instead of relying on a gut read of "today looks different from yesterday." Activity_score falls in the band where the threshold actually does work — where it separates something from nothing instead of either catching everything or catching nothing.

Why the threshold shifts depending on the comparison window

"How much does activity score need to move before it's worth looking at?" doesn't have one fixed answer — it depends on what it's being compared against. Comparing tonight against the trailing 28-night average is not the same question as comparing it against the trailing 168-night average. The more nights that go into the reference window, the more individual nightly noise gets averaged out, and the lower the bar drops for calling a change real.

That's why the source row carries four separate thresholds, one per window:

Comparison window (nights)	Threshold to stop being noise
28	21.14%
56	14.95%
84	12.21%
168	8.63%

Read together: against a short 28-night window, a change has to clear 21.14% before it can be called signal. Against the longest window on record, 168 nights, that bar drops to 8.63% — less than half. Same metric, same body, four different numbers depending on how much history sits behind the comparison. A short window barely averages anything, so most of that scatter is still visible and the bar to clear stays high. A long window smooths much more of it away, so a smaller shift is enough to stand out from what's left.

The row also carries two figures that describe how the metric behaves over time. The first is rho, the night-to-night autocorrelation, at 0.848: a value that high means one night's activity_score is strongly tied to the previous night's — the metric doesn't jump independently night after night, it carries memory forward. The second is rcv, the relative CV across windows, at 32.04%: it describes how much the noise threshold itself shifts depending on which window is chosen, and it's the reason the table above needs four rows instead of one.

How it was measured: 711 nights on a single body

Every one of these numbers comes from the same series: 711 consecutive nights of nightly activity_score recording on a single body. It is not an average across multiple people, and it is not a figure published by a wearable manufacturer — it is the CV, the median, the autocorrelation, and the four thresholds calculated directly on this one series. The median of 84 marks the center of the full distribution across those 711 nights; the 11.56% CV describes how spread out that distribution is around the median.

That the four thresholds (21.14%, 14.95%, 12.21%, 8.63%) and the two behavior figures (rho 0.848, rcv 32.04%) all come from one continuous series matters for reading the table correctly. They are not four independent measurements later averaged together — they are four ways of looking at the same 711-night series, each through a different reference window. Changing the window doesn't change the body or the metric; it changes how much individual nightly noise gets absorbed by the average that tonight's value is compared against.

The limit

This data describes the measurement noise of activity_score on a single body across 711 nights. It does not say whether a 21.14% or an 8.63% shift is good or bad for health — it only marks the magnitude past which a change stops being explainable by the metric's ordinary night-to-night variation. It also does not say whether these exact thresholds apply to another body: coming from a single subject, they are not a population standard, they are the measured behavior of this particular series. A different threshold in someone else would be consistent with this data, not a contradiction of it. And the 168-night threshold (8.63%) requires that much history to already be on record — it isn't available from night one, it's built window by window as the series grows. The same is true, in miniature, of the 84-night threshold: any window still needs its full length of nights behind it before it can be reported.

Mental Work — all protocols collects the rest of the cells measured on this same body. The related file magnesium-for-sleep documents a concrete sleep intervention on the same series. And open science — the dataset explains how every source row cited in this piece is built and verified.