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

For hrv_avg_ms, you need a change of at least 11.85% inside a 28-night window before you can call it real. That number comes from 650 nights measured on a single body: the night-to-night noise of this variable has a coefficient of variation (CV) of 14.94%. Below the threshold, the movement you see in your average HRV is indistinguishable from ordinary physiological drift.

What does a CV of 14.94% actually mean?

The CV describes how much your hrv_avg_ms moves from one night to the next with nothing of interest happening — no routine change, no intervention, no event behind it. It is the ruler everything else gets measured against. Any change you want to read as a signal — an intervention, a new habit, a stretch of bad sleep — has to clear that background noise first. If it doesn't clear it, you're looking at the same kind of variation that shows up even when nothing different occurred.

Reading the CV as a percentage makes it comparable across variables, regardless of the unit each one is measured in. A high CV means the variable swings a lot night to night for reasons unrelated to any real intervention; a low CV means the variable is inherently more stable, so a smaller change already deserves attention.

How was this measured?

The number comes from 650 nights recorded on a single body, the same subject behind this entire record. The median hrv_avg_ms across those 650 nights is 89 ms. The day-to-day correlation of this variable (rho) is 0.392: a moderate value, not so high that one bad night drags the whole trend with it, and not so low that the metric behaves like pure randomness. That moderate rho is exactly why a threshold is needed in the first place — the variable carries partial memory from one day to the next, but not enough to read a single point and trust it on its own.

The CV and the rho measure different things, which is why both matter together. The CV describes how much the variable disperses around its own average; the rho describes whether today's value gives any clue about tomorrow's. A variable can carry a lot of dispersion (high CV) while still holding some day-to-day memory (moderate rho) — exactly the case for hrv_avg_ms: it swings considerably night to night, but that swing isn't pure structureless noise.

The table: the threshold by how many nights you average

How much change you need before trusting a difference depends on how many nights you average before comparing. This table comes straight from the hrv_avg_ms source row, unrounded:

Window (nights)	Real-change threshold
28	11.85%
56	8.38%
84	6.84%
168	4.84%

The pattern holds across all four rows: the more nights you average, the smaller the change needed to count as signal. At 28 nights the threshold is 11.85%. At 56 nights it drops to 8.38%. At 84 nights, to 6.84%. At 168 nights — roughly half a year of data — it drops to 4.84%. The ratio between the raw noise of a single night and the tightest threshold reached at the longest window measured, the rcv, is 41.42.

Why does the threshold drop as the window grows?

Averaging more nights doesn't remove the noise from any individual night, but it does reduce the noise of the average: upward and downward deviations tend to cancel each other out once more points enter the calculation. That's why 4.84% at 168 nights and 11.85% at 28 nights describe the same underlying phenomenon — the same 14.94% CV — seen through different windows. The body isn't becoming more stable over time; the average of more nights is simply damping the swing of any single night more effectively.

Applying it: comparing two HRV numbers

If you compare last night's HRV against the night before, that comparison doesn't match any of the four windows in this table — it's a one-night window, noisier than any of them. A change of 11.85% or more between two single nights can be fully explained by this variable's 14.94% CV, with nothing different actually going on. When your sleep app shows last night's HRV against the night before that, it's showing you exactly this kind of short-window comparison.

If instead you average 28 nights against the previous 28 and the change clears 11.85%, that difference goes beyond what background noise produces on its own. And with 168 nights on each side, the threshold you need drops to 4.84% — with more data behind it, a smaller movement already counts as signal.

Which window to pick also determines what kind of question you can answer. A 28-night window suits a month-scale routine change: a new habit, a shift in sleep timing. A 168-night window suits a longer-horizon question, where what matters is whether something moved in a sustained way across roughly half a year. There's no "correct" window in the abstract — the right one matches the timescale of the change you're actually trying to detect.

The limit

This threshold is specific to hrv_avg_ms and to this body: it doesn't transfer automatically to another variable or another body, because every variable has its own CV and every body its own variance. The 650 nights come from a single subject (n=1), so this data doesn't say what threshold a different physiology would produce. It also doesn't establish causation: clearing 11.85%, 8.38%, 6.84%, or 4.84% confirms a change is statistically distinguishable from noise, not what produced it. And it isn't a prediction tool — the threshold describes noise already observed across 650 past nights; it doesn't forecast how much your HRV will move tomorrow night or next week.

Keep reading

This threshold is one row inside Mental Work — all protocols, the section that collects the measured thresholds for this body, variable by variable. The related file: magnesium-for-sleep documents a concrete sleep intervention with its own measured data. And the full dataset, with verifiable provenance, is published at open science — the dataset.