I wore a Polar Loop and a Coros Apex 4 side by side to decide which one to keep for sleep tracking. The more accurate device lost — and I think it deserved to.
· Polar 18 /21 nights recorded · Coros 21 / 21 nights recorded · 18 nights measured on both ·
01 / The setup: two devices, one wrist too many
I had been running a Polar Loop for sleep and a separate watch for training, and I was tired of it. Two devices, two apps, two batteries. I wanted one.
Rather than guess, I wore both for three weeks in July and compared them night by night. Eighteen nights were recorded by both devices, which is enough to see where they agree, where they don’t, and which disagreements actually matter.
| 18nights on both wrists | 2 minmedian bedtime disagreement | +17Coros sleep-score advantage | 21 / 18nights recorded Coros / Polar |
Both platforms date a night by the morning it ends, so the two exports line up directly with no shifting.
02 / Head to head: Polar Loop vs Coros, 18 parallel nights
They agree almost perfectly on when I slept. Bedtime differs by a median of 2 minutes, wake time by 5. Time asleep correlates at r = 0.99. On one strange 4-hour night, Polar logged 01:03–05:22 and Coros 01:04–05:22.
They disagree on nearly everything about quality.
| Metric | Polar | Coros | Δ |
|---|---|---|---|
| Time asleep | 6h42 | 7h12 | +30 min |
| Awake during night | 29 min | 11 min | −18 min |
| Efficiency | 93.2% | 97.4% | +4.2 |
| Deep sleep | 70 min | 102 min | +32 min |
| REM | 100 min | 106 min | +6 min |
| Sleep score | 73.0 | 89.6 | +16.6 |
The one number I’d now ignore entirely is deep sleep: correlation between the two devices is 0.29, essentially unrelated. On 16 July, Polar said 52 minutes and Coros said 134. Treat stage minutes as device-internal — only ever comparable to themselves.
On awakenings, Polar counts 4 long interruptions a night; Coros counts 0.2. The three nights with the most Polar-logged interruptions are the only nights Coros also flagged an awakening — two independent sensors picking the same worst nights, in the same order. But Coros stayed silent on 10 nights where Polar found three or more long interruptions. Coros is high-specificity, low-sensitivity: when it speaks it’s right, but it’s silent 78% of the time. As a number to track, that’s useless.
03 / The verdict: the more accurate device lost
Let me be clear about this, because it’s the uncomfortable part: the Polar Loop is the better instrument. Finer resolution, better arousal detection, a score with real range. If accuracy were the only axis, it wins outright.
It’s going in a drawer anyway.
Polar Loop
- More accurate. Detects ~4 long interruptions a night against Coros’s 0.2, and its score spreads 51–89 where Coros bunches at 93+.
- Unreliable. Keeps dropping its Bluetooth link. Every time, I re-pair it — and lose nights. It managed 18 of 21 nights; Coros managed 21 of 21.
- Loses the nights that matter. My single most fragmented night of July — 17 long interruptions, 74 minutes awake — is one Coros has no record of.
- Can’t recover a broken session. Walk to the kitchen for water and it ends sleep mode, with no way back in. One full night was recorded as 4h07 and scored 51.
- No nap tracking.
Coros
- It just works. 21 nights out of 21. The one missing night was me not setting it up properly.
- Asks before giving up. Get up in the night and it prompts: stop sleep mode or not? Say no, go back to bed, it carries on recording.
- Naps, properly. Set a 25-minute nap, it buzzes you awake and logs it. On the 4-hour night it caught the 2-hour morning recovery sleep Polar simply discarded.
- One device. Sleep and training on the same wrist.
- Less accurate, and generous. +30 min asleep, +17 on the score, +32 min deep, −18 min awake. It flatters you.
- Awake count is unusable — it only counts awakenings over 5 minutes.
A device that loses whole nights is worse than one that’s slightly less granular. Precision you can’t rely on isn’t precision.
04 / Workarounds: how to switch without losing the thread
Your trend line survives the switch
The metrics my current experiment depends on — bedtime and wake time — are exactly the two the devices agree on (2 and 5 minutes median difference). Sleep score, deep sleep and awakenings all step-change; the timing data doesn’t. Note the switch date in your log so any jump has an obvious explanation.
Calibration offsets for old data
To compare a Coros night against Polar history, subtract roughly: 30 min asleep, 17 score points, 32 min deep, 4 points efficiency, and add back about 18 minutes awake.
Use awake minutes, not awake count
This is the one that matters. Coros’s awake count is near-useless, but the awake minutes field sitting right beside it tracks Polar’s long interruptions at r = 0.91, with a 2–39 minute range and no contamination from night length.
A negative result worth publishing
I tried to reconstruct awakenings from gaps in the Coros HRV time series — the theory being that movement disrupts the optical sensor, so missing samples mark wake periods. On the first 7 nights it looked excellent: r = 0.87. Extending to all 18 nights collapsed it to r = 0.44. Gap count correlates with sleep-window length at 0.68 — it was largely measuring how long the night was. On 21 July the proxy found the most gaps of the month on what was actually my best night.
Classic small-sample overfit. The honest version of this post includes the idea that didn’t work — and the fix turned out to be a field the device already reported, which I’d overlooked because I was busy trying to replace the wrong one.
Keep your notebook
No wrist device reliably catches quiet, motionless wakefulness. If that’s the thing you actually care about, a pen by the bed is a better instrument than either watch. Twenty seconds a morning: wake time, whether it was natural and any night wake worth noting.
Leave a Reply