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

Across 650 nights on a single body, night temperature carries the highest coefficient of variation in the entire dataset: 1670.16%. To clear the statistical noise floor at 28 nights, a shift must reach 1073.38%. No other metric in the 31-variable set reaches this level of dispersion. Those numbers come from one source: the expedition's own data. No other page on the internet has them.

The finding

The ring does not measure absolute temperature. It measures the nightly deviation from each person's personal baseline. The number in your app is not a thermometer reading. It is a distance from your own zero.

Across 650 nights of continuous tracking, the median of that deviation is 0. Half the nights land exactly at baseline. The rest scatter. Some shift a little. Some shift a lot. The distribution is not smooth.

That structure explains the coefficient of variation: 1670.16%. When the center of a distribution sits at zero and the tails extend far, the CV climbs without bound. A CV of 1670.16% is not a data error. It is an accurate description of how this metric behaves.

The autocorrelation coefficient is 0.202. That is low. A high-deviation night does not predict the next night well. Temperature does not carry momentum. Each night begins, statistically, close to a fresh start.

Metric	Nights	Median	CV	rho	RCV	u28	u56	u84	u168
temperature_delta	650	0	1670.16%	0.202	4629.45%	1073.38%	759%	619.72%	438.21%

The RCV (Relative Coefficient of Variation) is 4629.45%. That composite index combines variability and autocorrelation. It is the highest value in the 31-metric expedition. Nothing else in the dataset reaches it.

The gap

Sleep science has studied temperature for decades. Large studies with thousands of participants link ambient temperature to sleep onset and depth. Recovery protocols lower the thermostat before bed. Guidelines recommend an optimal room temperature range.

The gap is H3: real group-level effect, without individualizing. Those studies show what happens on average. They do not say how much YOUR night temperature needs to shift before the change is distinguishable from YOUR body's natural variation in YOUR data.

A statistically significant group effect does not automatically translate to a readable signal on your wearable.

The transfer

Suppose you adjust something tonight. You lower the thermostat two degrees. You change your bedding. You take a cold shower. Your wearable logs a temperature deviation for that night.

The question that follows: is that shift the intervention, or the usual scatter?

The expedition has the numerical answer. With 28 nights of data, the threshold to exit the noise band is 1073.38%. At 56 nights: 759%. At 84: 619.72%. At 168: 438.21%.

The threshold falls as nights accumulate. More observations compress the noise floor. But even at 168 nights, the threshold sits at 438.21%.

This is the gap transferred to your week. Not that the thermostat is useless. But that the sensor, in its current form, produces a signal with enough dispersion that a single-night variation on screen is not sufficient evidence of anything. Neither for nor against the intervention.

Most self-experimentation frameworks use short observation windows. For night temperature, 28 nights carries a threshold of 1073.38%. The gap between human evaluation pace and statistical signal pace is the hole moved into your schedule.

The instrument

The dataset holds 650 nights. 18483 total measurements across 31 metrics on a single body, tracked with Oura. The device matters. Oura and WHOOP sensors do not produce equivalent values even when they use the same variable name. Different sensor, different window, different algorithm. The thresholds in the table belong to that subject with that device. They do not transfer directly to another body or another sensor.

What the instrument offers is individual and specific: the magnitude a temperature shift must reach before it is distinguishable from that body's own noise floor. Not a population average. Not an estimate from large-scale studies. The exact threshold, derived from 650 nights of measured variability on one person.

That number decreases over time. At 28 nights: 1073.38%. At 168 nights: 438.21%. Longer observation windows compress the threshold. The signal required to trust a change shrinks as the window grows. If you plan to test an intervention on night temperature and want to read the result, the observation window has to match the corresponding threshold. Otherwise the reading is noise by definition.

The limit

This data does not say temperature is irrelevant to sleep.

It says the temperature signal — across 650 nights on a single body with Oura — carries enough dispersion that it demands high change thresholds to exit the statistical noise floor in short observation windows.

What this data does NOT say:

The thresholds are not universal. Each body has its own variability profile. A second subject, Lucca, is in the expedition, tracked with WHOOP. His temperature thresholds do not compare directly with the case zero values. The sensor differs. The algorithm differs. Absolute values across different devices cannot be separated — any comparison mixes instrument and body.

This data does not say no intervention can move temperature systematically. The expedition has no controlled intervention data on this variable. There is no structured experiment of the form: I lowered the thermostat for 28 nights and measured the effect on my temperature delta. That experiment is not in the vault.

It does not say temperature is the right proxy for what you are trying to track. If the real target is recovery quality, or sleep depth, or something else, a more stable metric in the 31-variable set may capture the same outcome with less noise. The expedition has published thresholds for all 31.

The median is 0 across 650 nights. That does not mean the body holds temperature flat. It means regulation is stable enough that the median deviation is zero. When shifts occur, they are large. The CV of 1670.16% captures those shifts, not the resting state.

The question this piece does not answer: does any real intervention produce a systematic temperature shift that consistently clears the 1073.38% u28 threshold across repeated 28-night windows? That experiment is what the next piece inherits.

---

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