How much does readiness have to change before it stops being noise? The threshold from 650 nights

A readiness score has to move more than 6.14% inside a 28-night window, or more than 4.34% inside a 56-night window, before that movement stops being explainable as background noise. Those are the two thresholds measured across 650 nights of tracking on a single body, with a night-to-night variability (CV) of 10.21% across the full series.

What "noise" means in a sleep number

A number going up or down does not mean anything on its own. A readiness score fluctuates from one night to the next even when nothing meaningful changed in routine, training, or rest: sleeping in a different position, a late dinner with no real consequence, or plain biological variation already move the needle. That background fluctuation is the noise. The practical problem is that, without a threshold, any reading that climbs 2 or 3 points looks like a signal — and it is not.

The only honest way to set that threshold is to measure it, not assume it. This file uses 650 nights of first-party data from a single body to calculate, from the real series, the point at which a change in readiness score stops being explainable by chance.

How it was measured: 650 nights, one body

The source row (ruido.medido.json#readiness_score) summarizes the full series: a median of 83 across the 650 recorded nights, with a coefficient of variation (CV) of 10.21% night to night. That CV is the natural dispersion of the metric — how much the readiness score bounces from one night to the next with nothing having actually happened.

The series also carries a correlation coefficient (rho) of 0.137 between consecutive nights. That is a low correlation: it means one night's value predicts little about the next one, so each night contributes nearly independent information instead of repeating the one before it. This matters because when data points are close to independent, averaging them over a multi-night window actually reduces noise — the averaging is not just arithmetic convenience, it works precisely because one night's noise does not carry over into the next.

The raw relative coefficient of variation for the un-averaged series (rcv) is 28.29%. That is the starting point: the raw, night-by-night dispersion before any grouping. From there, averaging over longer windows brings that initial 28.29% down to the thresholds that are actually useful for deciding whether a change is real.

The table: threshold by night window

Four averaging windows were calculated on the same 650-night series. Each one produces a different threshold — the minimum percentage change, at that window length, needed for a move to stop being noise:

Window (nights)	Noise threshold
28	6.14%
56	4.34%
84	3.54%
168	2.51%

The pattern is the one expected in any series with low autocorrelation: the more nights that go into the average, the smaller the change needed to trust it. At 28 nights, a move of 6.14% is required before it counts as real; at 168 nights, with 6 times more accumulated data, that threshold drops to 2.51%. The 56-night window — half of a 112-night span, which is not part of this series either — lands at 4.34%, roughly midway between the two.

Why this changes how the number should be read

Someone checking their readiness score every morning against the day before is looking at the noisiest window possible: a single night, where the relevant threshold does not even show up in the table above — it would sit higher than the 28-night figure. A single daily reading, on its own, cannot tell a real streak apart from an expected fluctuation given a CV of 10.21%.

Comparing against a 28-night average, by contrast, already requires a 6.14% change before it is worth taking seriously. And if the question is about a medium-term trend — a training quarter, a sustained routine change — the 56-night window with its 4.34% threshold is the one that separates a real improvement from a lucky month.

This is not a stylistic preference for averages over single days. It is what the series itself says: with a night-to-night correlation of just 0.137, each individual reading has little to do with the next one, and only once enough nights accumulate does the noise cancel out enough for a change of a few percentage points to mean something.

The limit

This data point does not say whether a 6.14%-or-larger change over 28 nights is caused by something good or something bad — it only says that such a change is no longer explainable by the measured background variability. The cause has to be found separately, cross-referenced against other variables in the same dataset.

Nor does it claim these thresholds are universal. They are calculated on 650 nights from a single body, with its own routine, its own device, and its own physiology. A CV of 10.21%, a rho of 0.137, and an rcv of 28.29% are numbers that belong to this series, not a biological constant — a different body, with a different sleep routine, may carry a different background dispersion and therefore different thresholds.

It also does not cover variables outside the readiness score: every metric in the protocol (31 of them, measured across up to 712 nights, with 18483 measurements in total) carries its own noise level, and a threshold from one cannot be assumed to transfer to another.

And this is not a sleep diagnosis or a routine recommendation: it is a measurement of how much noise sits inside one numeric series, nothing more.

Related files

This file is part of Mental Work — all protocols, where measurements related to cognitive load and recovery are grouped together. The related file: magnesium-for-sleep documents a concrete intervention on this same sleep series. And anyone who wants to verify these calculations directly against the raw data can do so at open science — the dataset, where the source file lives along with its cryptographic fingerprint.