Research note 05
Reading Sleep and Recovery Data Honestly
Wearable sleep tracking has improved considerably at telling you when someone slept, and remains weak at telling you how. Knowing which claims to keep and which to discard changes what the data is worth.
Sleep is one of the few recovery interventions with a large, consistent and uncontroversial evidence base behind it. Restricting an athlete's sleep degrades reaction time, mood, glucose handling and immune function; extending it in habitually short sleepers improves performance measures. None of that is in dispute.
What is in dispute is how much of it your wearables can measure. This note separates the claims that hold up from the ones that do not, and suggests what a performance department should actually do with sleep data.
How a wrist device estimates sleep
Consumer and prosumer sleep trackers combine actigraphy — movement measured by an accelerometer — with photoplethysmography, an optical heart-rate measurement, and sometimes skin temperature. Algorithms classify epochs, typically thirty seconds, into wake or sleep, and then subdivide sleep into stages.
The two tasks are not equally hard. Distinguishing sleep from wake using movement plus heart rate is a comparatively tractable problem, and modern devices do it reasonably well. Classifying sleep stages — light, deep, REM — is a substantially harder problem, because the physiological definitions of those stages come from electroencephalography, and a wrist device has no access to cortical activity at all. It is inferring a brain state from peripheral proxies.
What holds up
| Output | Trust | Use it for |
|---|---|---|
| Sleep onset and offset time | High | Schedule conflicts, travel disruption, late-fixture recovery windows |
| Total sleep duration | Good | The single most actionable output; track it against a per-athlete target |
| Sleep regularity night to night | Good | Arguably as important as duration and much less often used |
| Wake episodes after onset | Moderate | Trend within an individual only |
| Stage proportions (deep / REM) | Low | Not a basis for any decision |
| Composite “sleep score” | Very low | Undocumented, vendor-specific and not comparable across devices |
The bottom two rows are where most of the value is claimed and almost none of it exists. Stage classification against polysomnography — the clinical reference — shows agreement that is modest for deep and REM sleep even in the better research-grade devices, and vendor algorithms are proprietary and change between firmware versions without notice. A deep-sleep percentage that shifts because a device updated overnight is not a physiological finding.
The measurement problem nobody mentions
Compliance. A sleep metric only exists if the athlete wore the device, and athletes are less consistent about wearing wrist devices at night than any monitoring plan assumes. Missing nights are not missing at random: they cluster around travel, late fixtures, social events and exactly the disrupted periods you most wanted to observe.
This produces a bias that is easy to miss. Your dataset systematically over-represents the nights when things were normal. An athlete whose sleep data looks stable may simply be an athlete who takes the device off when their routine breaks.
What we recommend instead
Three practical positions, all of which reduce the amount of data collected and increase how much of it is worth acting on.
1. Track duration and regularity, and stop there
Set a per-athlete duration target based on their own habitual need rather than a population figure, and track the seven-day mean plus the night-to-night standard deviation. Regularity is under-used: an athlete averaging eight hours with a three-hour night-to-night swing is in a materially worse position than one averaging seven and a half consistently.
2. Pair objective data with a two-question subjective check
Perceived sleep quality and morning alertness, captured in the same app as session-RPE, add information that no wrist device provides — and provide a cross-check when the device data looks implausible. Two questions is the maximum a squad will complete daily over a season; more than that and the compliance curve collapses within weeks.
3. Use sleep data to change environments, not to change sessions
The strongest returns from sleep monitoring in professional sport come from structural interventions: fixture and travel scheduling, hotel room allocation, blackout provision, meal timing, and moving morning sessions later after night fixtures. Those changes are informed perfectly well by duration and timing data. Adjusting an individual's training on the basis of last night's deep-sleep percentage is a decision built on the weakest number in the dataset.
How it fits our platform
We do not manufacture a sleep device and we do not intend to. What the analytics suite does is ingest sleep duration, onset, offset and regularity from the device your club already uses, place them on the athlete's timeline alongside load and readiness, and carry a completeness indicator so a run of missing nights is visible rather than invisible.
We deliberately do not ingest vendor sleep scores or stage proportions into any modelled output. They can be displayed if a client wants them, marked as vendor-supplied, but they do not feed the readiness model or the injury risk models. Feeding an undocumented, unstable input into a model whose output influences decisions about people is not a defensible engineering choice.
The summary
Sleep matters enormously and your wearables measure a useful subset of it. Duration, timing and regularity are worth collecting, worth tracking against the individual, and worth acting on structurally. Stage proportions and composite scores are worth ignoring. The department that acts on the first list and ignores the second will get more out of sleep monitoring than one that treats every field on the dashboard as equally real.
