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

Sedentary minutes fluctuate substantially from one day to the next. The question is not whether they changed. The question is whether the change is large enough to stand out from the fluctuation the body produces on its own, with nothing different happening.

The direct answer: over a 28-night window, sedentary_min has to shift at least 18.13% before the move is likely signal rather than noise. Over 56 nights, that threshold drops to 12.82%. Anything below those figures lives inside the metric's own natural variance. These numbers come from 712 nights of a single body, measured with the same device throughout, with no change of instrument.

Why sedentary_min is hard to read from day to day

The coefficient of variation (CV) for this metric is 28.64%.

CV measures how much a variable oscillates around its own median. The recorded median for this subject is 454 minutes of sedentary time per day. A CV of 28.64% means the number moves considerably even when nothing in the lifestyle changes. A week heavy on remote calls, a travel day, a Tuesday spent entirely at a writing desk — sedentary_min rises and falls without any identifiable underlying cause.

This matters in practice. A day with more seated meetings, a commute by car instead of on foot, an afternoon at an unfamiliar desk — all of these move the number without any sustained change in underlying behavior. The variable responds to ordinary daily variation with the same intensity it responds to a genuine shift in how you move.

Reading today's number against yesterday's is reading noise against noise.

The autocorrelation (rho) for this variable is 0.187. A value near zero means yesterday's reading barely predicts today's. This metric has almost no short-term memory. Each day is statistically close to independent from the one before it.

That also means today's reading does not usefully predict tomorrow's. Each data point is, in practice, nearly independent: there is no accumulated momentum connecting one day's reading to the next.

How the thresholds were calculated

The reference change value (RCV) answers a specific question: how much does a single reading have to move before that movement exceeds the combined margin of the instrument and the subject's own biological variability?

For a single night of sedentary_min, that value is 79.4%. A single reading needs to sit 79.4% away from the median before the move is distinguishable from normal fluctuation. This is not a model assumption. It is a calculation from 712 nights of one subject, one device, applied with the same method to all 31 variables in the dataset.

Aggregate over longer windows and the noise averages out. At 28 nights — the shortest window tracked — the threshold is 18.13%. At 56 nights it drops to 12.82%. At 84 nights it reaches 10.47%. At 168 nights, 7.4%.

The relationship is direct: longer window, lower threshold, cleaner signal.

The source row

Indicator	Value
Nights recorded	712
Daily median	454 min
Coefficient of variation (CV)	28.64%
Autocorrelation (rho)	0.187
Reference change value (RCV)	79.4%
Threshold at 28 nights	18.13%
Threshold at 56 nights	12.82%
Threshold at 84 nights	10.47%
Threshold at 168 nights	7.4%

Source: ruido.medido.json#sedentary_min. One subject. One device. 712 consecutive nights.

What these thresholds mean in practice

Most wearables show the day's number. Some add a weekly average. None tell you how much that average has to shift before the shift is real rather than noise.

Sedentary_min has a CV of 28.64%. That is a natural oscillation built into the metric — generated by the body and by ordinary daily variation. A move below the 18.13% threshold over 28 nights could be a Tuesday with more standing calls, not the result of any real change in how you move through the day.

The threshold does not say what level of sedentary minutes is good or bad. It says when the data carries enough signal to be worth reading. That distinction is one most wearable apps do not make.

A straightforward way to apply this: compare the median of your last 28 nights to the median of the 28 before that. If the difference does not reach 18.13%, the shift has not cleared the noise floor. If it reaches 12.82% over 56 nights, the signal is more robust.

The median is also less sensitive to outlier days than an arithmetic average. One travel day or sick day pulls the arithmetic mean for the entire window, but it carries no more weight than any other day in the median. That makes the window-to-window median comparison more resistant to single-day anomalies than any comparison of simple averages.

The number that matters is not the daily reading. It is the sustained shift in the median across time.

The limit

This dataset covers 712 nights of one body, measured with one device.

The CV of 28.64% belongs to that subject in that period. There is no basis for assuming another body produces the same coefficient of variation for this metric. It could be higher. It could be lower. The thresholds here are not universal norms. They are the calibration of a single baseline.

The autocorrelation of 0.187 is also specific to that subject and period. If the underlying routine changes in a sustained way — different work, different environment, different schedule — the oscillation pattern may shift with it. These thresholds are not permanent.

The RCV of 79.4% means a single reading cannot separate signal from noise for this variable. That is a property of the data, not a limitation of the device.

There is one more boundary worth noting: the dataset does not include any change of device. If the wearable is replaced, the instrument error may shift, and thresholds calibrated on the previous device do not carry over automatically. The calibration would need to be recalculated from the new baseline.

One question this piece does not answer: what happens when sedentary_min stays above the threshold for 56 nights or longer? If there is a real underlying shift, which other variables in the 18,483-measurement dataset move alongside it? That is what the open record tracks.

---

Dataset source: hash published in corpus.huella.json · 18,483 measurements · 31 variables · 712 nights · one subject.

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