Can I trust my watch's numbers?
Heart rate, and overnight resting HR or HRV as your own trend on one device: mostly. VO2max (±10 mL/kg/min), calories, sleep stages, body fat and readiness scores are estimates: context, not decisions.
The question it answers
An athlete or the assistant holds a number from a device — a low HRV, "VO2max dropped", 3 000 kcal, body fat +2%, core 38.9 °C, a new watch LTHR, two power meters that disagree. Should it change today's session, a zone, a fuelling plan or a goal?
Short answer: only a reading the sensor measures, compared with the same athlete on the same device and read as a trend, carries enough weight to move something, and even that goes with what the athlete reports; every other device number is context whose error is said out loud.
What this framework owns and what it leaves
| Owns (decided here) | Leaves to |
|---|---|
| The class of a device number (measured, estimated, composite) and how much it may move | The response to a low-readiness day and the subjective-first rule → SYNTH-30 |
| The error to quote for each reading, and the condition it was measured in | Heart-rate anchors and their measured-or-estimated label → SYNTH-38; which test anchors a zone → SYNTH-53; the easy ceiling → SYNTH-51 |
| The same-device, new-baseline and change-smaller-than-the-error rules | VO2max as a range → SYNTH-78; food calories → SYNTH-87; body-fat targets → SYNTH-54; heat-session targets → SYNTH-04; sleep → SYNTH-07; trainer modes → SYNTH-56 |
Default (what to prescribe first)
Not from a study; the base's own practice, built on the agreement data in the tables below:
Name the reading's class before using it. Measured by the sensor: heart rate from a chest strap, overnight resting heart rate, beat-to-beat intervals, power from a strain-gauge meter, body mass, a lactate value.
Estimated by an algorithm from other signals: VO2max, threshold heart rate or pace, calories, sleep and its stages, body-fat %, core temperature from a skin sensor, thresholds from a breathing sensor, running power. Composite: readiness, recovery, strain, "body battery".
This generalises SYNTH-38's rule — record whether a heart-rate anchor was measured or estimated and carry that label everywhere — to every device number. The order follows the agreement data (REF-1248, A-) and an expert hierarchy (REF-397, C+).
Act on a reading only when it is measured, compared with the same athlete on the same device, and read as a trend — and then together with what the athlete reports, never against it without asking (SYNTH-30). A measured reading inside a session (chest-strap heart rate, power) steers that session as it happens.
Treat an algorithm's estimate, a composite score, a comparison across devices or people, and a single reading as context. Give its error once, in one clause, and ask the athlete.
A change smaller than the device's error is not a change (inference from the limits of agreement below): a one-night HRV dip inside the device's single-night limits, a few per cent between two power meters, a few mL/kg/min of watch VO2max, a few points of scale body fat.
Trust table
"Support" gives the entry, its level, the key agreement number with its limits where reported, and the condition it was measured in. "(practice)" marks a call that is not from a study; the base's own practice. Every "good enough for" and "not good enough for" call is practice built on the support cell.
| Reading (class) | Good enough for | Not good enough for | Support |
|---|---|---|---|
| Chest-strap heart rate (measured) | Heart-rate zones and caps, decoupling, reps longer than the heart-rate lag (SYNTH-38), heart rate over the last minute of a fatigue test (SYNTH-41); beat-to-beat intervals for HRV | A recording with many artefacts; DFA a1 at high intensity (row below) | REF-1250 B: Polar H10 vs 12-lead ECG, 25 recreational adults, cycling ramp to exhaustion: heart-rate concordance 1.00, beat-interval limits of agreement +4.3 to −2.8 ms at low and +1.3 to −0.5 ms at high intensity; recordings with more than 5% artefacts excluded; one strap model; no external funding |
| Wrist or arm optical heart rate (measured, movement-sensitive) | Resting heart rate, sleep, steady running — on average | One session's exact value; zones, decoupling and fatigue tests on the bike or in the gym (practice); reps shorter than the heart-rate response (practice) | REF-1249 B+ (abstract): 44 studies, 15 brands, before 2020: mean difference −0.01 bpm at rest, −0.40 in sleep, −0.51 walking to running, −4.55 (95% CI −7.24 to −1.87) cycling, −7.26 (−10.46 to −4.07) resistance training; no limits of agreement given; REF-1248 A-: errors of about ±3% in a narrative review of 29 studies, a Fitbit meta-analysis −3.39 bpm with more under-reading in cycling. Arm-worn optical sensors: no entry |
| DFA a1 (computed from measured beats) | A starting VT1 heart rate from a chest-strap ramp, labelled estimated until a hard effort agrees (SYNTH-38) | VT2; a wrist sensor (inference: the devices studied were straps or ECG) | REF-1231 A-: HRV thresholds agreed with lab thresholds at group level (DFA a1 for the first, SMD 0.16, 6 studies; for the second, SMD 0.06, 5 studies; device type did not change agreement), individual error not pooled; REF-873 B: 154 vs 152 bpm at gas-exchange VT1, 15 people, one treadmill ramp; REF-1250 B: strap-vs-ECG DFA a1 limits +11.6 to −9.9% at low intensity and +58.1 to −40.9% at high; the strap read about 6% higher at 0.75 and 7.5% at 0.5, so both thresholds fell at a higher heart rate (size in bpm not reported) |
| Overnight resting heart rate (measured, averaged over a brand-specific window) | A within-athlete trend over days on one device | Comparing brands, or a Garmin "resting heart rate" (lowest 30-min average of the day) with another brand's night average; a one-night shift of a beat or two | REF-1251 B-: 13 healthy adults, 536 nights at home vs a Polar H10 strap: mean absolute error 0.98-1.78 bpm on Oura Gen 3 and Gen 4, WHOOP 4.0 and Polar Grit X Pro, limits from −2.84/+1.08 (Oura Gen 3) to −4.72/+1.90 bpm (WHOOP; Polar's printed −5.36/+2.06 conflicts with its own bias and SD); Garmin Fenix 6 not comparable; REF-364 B: one night, six devices, heart-rate bias 0.1-5.0 bpm in size (Garmin worst, ICC 0.41) |
| Overnight HRV (measured, averaged over a brand-specific window) | A multi-day direction against the athlete's own baseline on one device (SYNTH-30) | Absolute values; comparing brands or people; reading across a device change; one night | REF-1251 B-: MAPE 5.96% Oura Gen 4, 7.15% Gen 3, 8.17% WHOOP 4.0, 10.52% Garmin Fenix 6, 16.32% Polar Grit X Pro; single-night limits −11.43/+6.43 ms (Oura Gen 3), −11.78/+9.85 ms (Oura Gen 4), −12.50/+10.94 ms (WHOOP 4.0), −15.2/+11.6 ms (Garmin) and widest on the Polar (printed −14.30/+23.60, one sign flipped); error persisted against each person's own baseline; REF-364 B: one night, RMSSD bias −4.5 ms (WHOOP 3.0) to −22.4 ms (Garmin, a 3-min manual test before lights out), limits ±7.6 ms on WHOOP and ±55.2 to ±92.0 ms on the others; REF-1248 A-: Polar devices at rest ICC 0.98-1.00 (one review, 13 studies), a range of devices 0.85-0.99 at rest falling to 0.85 with movement (another review); by the entry's reading Polar and chest-strap evidence, not a test of optical watch HRV |
| Sleep duration and timing (estimated from movement and heart rate) | Bedtime, wake time and the trend in total sleep on one device (SYNTH-07) | An exact total; wake during the night; time to fall asleep | REF-1252 B+: 24 studies, 798 mostly non-athlete adults vs polysomnography: total sleep about 17 min high on average (16.9, 95% CI 7.4-26.3) and wake after sleep onset about 13 min low (direction read from the forest plot; the paper states no sign), I² 58-84%, single studies from about 105 min too much sleep to 86 min too little (printed as −105 to +86); REF-364 B: sleep-wake agreement 86-89%, total-sleep bias from −12.2 (WHOOP) to +43.8 min (Garmin), wake detected 26-57% |
| Sleep stages (estimated) | Curiosity | Any decision; "deep sleep" or REM minutes as a target | REF-364 B: 50-65% agreement with polysomnography, κ 0.20-0.52, one night, 53 healthy adults; REF-1252 did not analyse stages |
| Readiness, recovery, strain, "body battery" (composite) | Opening a conversation | Any decision on its own | REF-1248 A-: the authors state that these scores (Fitbit, Garmin, Oura, Whoop) have not been formally validated; REF-1242 B+: no readiness marker, subjective or device, validated against the next session; REF-397 C+: composites trusted least (expert hierarchy, a Gatorade Sports Science Institute publication) |
| Watch VO2max (estimated) | A long-term direction on the same watch from similar runs (inference: test-retest reliability barely studied) | A fitness verdict; comparing athletes; calling a change of a few mL/kg/min progress; zones; an estimate made from cycling (none was tested) | REF-1253 A-: 14 studies, 403 healthy active adults aged about 25, every exercise-based estimate from running: from runs (Garmin, Fitbit) bias −0.09 mL/kg/min, limits −9.92 to 9.74; from a resting test (Polar) +2.17, limits −13.07 to 17.41; I² 77-88%, 11 of 14 studies at high risk of bias, method confounded with brand; partly funded by Huawei (no Huawei device evaluated); the lab criterion carries error too (REF-169, B) |
| Watch LTHR and LT pace (estimated) | A starting estimate, labelled as such | Zones without a field effort or a lab test (SYNTH-38, SYNTH-53) | REF-1232 B: 100 recreational runners (Huawei; 23 also Garmin, 17 COROS), the watches' own test: LTHR about 9-11 bpm off the lab on average, r = 0.13-0.67, no value in 22-53% of tests, LT pace about 1-2 km/h too fast; Huawei supplied the equipment. A value auto-detected from ordinary runs: no entry |
| Watch calories (estimated) | Order of magnitude; the same watch on the same kind of session week to week (inference) | Fuelling or deficit arithmetic; comparing brands | REF-1254 B-: 20 young recreationally active adults, 2-km outdoor walk and run vs portable gas analysis: Apple Watch Series 6 and Garmin Fenix 6 about 17-29% high on average (MAPE 19.8-32.0%), Huawei Watch GT 2e unbiased on average but single bouts off by up to about ±30% (these percentages are the base's arithmetic from the paper's kcal means and SDs); funded by Huawei; REF-1248 A-: a Fitbit meta-analysis about 3 kcal/min low (limits −13 to +7 kcal/min). No cycling. On the bike the power meter's kJ is the better number: about 1 kcal per kJ at ~24% gross efficiency (arithmetic, practice; absolute gross efficiency differs between energy-expenditure equations, REF-489, B) |
| Cycling power meter (measured) | FTP or CP, zones, load and pacing — on that meter | Comparing two meters, or carrying zones to another meter, without a check; a 2-3% change measured on different meters | REF-1255 B- (abstract): 54 riders' own meters, 9 makers, before 2017, vs a physics model of treadmill cycling: −0.9 ± 3.2% (mean ± SD), 6 of 54 off by more than ±5%, repeat precision CV 1.2 ± 0.9%, Stages more variable than SRM and PowerTap; REF-846 B: a 1998 bicycle-mounted meter validated against a laboratory ergometer, road model R² = .97, SE 2.7 W |
| Smart-trainer power (measured or modelled by the trainer) | Indoor sessions on that trainer once its offset from the rider's own meter is known (practice, SYNTH-56) | Trusting it unchecked; reading trainer and outdoor watts as one scale | No entry tests trainer accuracy (gap); REF-244 C+ (commentary: questionable data accuracy on virtual platforms); the "known" meter it is checked against is itself off by a typical ±3% (REF-1255) |
| Running power (modelled from accelerations, speed and mass) | Pacing by watts within one athlete on a repeatable device, on a treadmill or track | Comparing brands or with cycling watts; Garmin, Polar or RunScribe power as a pacing target; trails and steep descents | REF-1256 B- (abstract): 12 endurance-trained men, treadmill and track, devices before 2020: Stryd SEM ≤12.5 W, CV ≤4.3%, ICC ≥0.980, r with VO2 ≥0.911; Garmin, Polar and RunScribe power repeated poorly; validity judged by correlation with VO2, not by measured mechanical power; REF-1241 B: its authors note that Stryd's stiffness model is not validated for steep uphill running |
| Scale weight (measured) | The body-mass trend; a sweat-rate test on the same scale (SYNTH-68, SYNTH-87) | One morning's change read as fat | No entry tests scale weight (gap); practice |
| Scale body-fat % (estimated from impedance by an equation) | Nothing to act on; at most a slow trend on one device in the same conditions | A target or race weight (SYNTH-54); a single reading; a change of a few points over months | REF-1257 B: 15 bioimpedance devices vs a four-compartment model, 73 healthy non-athletes (37 retested at 12-16 weeks): constant error −3.5 to +11.7 points; a single reading within about ±6.5 points on the best models and ±11-15 on most foot-to-foot scales, a change within about ±3.5-5 points against a true change of SD 2.9 (limits reconstructed from the paper's tables); "athlete" modes not tested; the principal investigator has had in-kind equipment from InBody among other makers |
| CORE core temperature (estimated from skin heat flux) | A within-session direction on the same sensor (inference; not tested) | A 38.5 or 39 °C threshold, a safety stop or a heat-dose count (SYNTH-04) | REF-1258 B- (abstract): 19 elite female hockey players, 4 hot training sessions before the 2021 Tokyo Games, the CORE version of that time vs an ingested pill: concordance "poor" in every session, CORE 0.06-0.34 °C low on average (intervals ±0.35-0.56 °C), under-reading grew as core temperature rose, 41-60% of samples off by more than ±0.3 °C; funding not given in the abstract |
| Tymewear thresholds and breathing (estimated from chest stretch) | VT1 as a rough estimate (close on average); a change between two identical tests on the same device (inference) | VT2 without a field or lab check; the ventilation value itself, or a ventilation value at VT1 or VT2 stored as a zone boundary (untested); the strap; cycling | REF-1259 B-: 15 recreationally active adults, treadmill ramps at 2,348 m, the Tyme Wear shirt of about 2018-2021: test-retest ICC 0.91-0.98 for time and speed; against gas exchange VT1 0.2-0.3 mph (about 0.3-0.5 km/h) and about 4 bpm low, VT2 0.5-0.8 mph (about 0.8-1.3 km/h) and 5-8 bpm low, VT2 heart-rate limits up to −18.7 to +35.1 bpm, the fittest runners under-read most; 4 of 19 excluded for equipment error or data loss; protocol developed by Tyme Wear, no external funding. Breathing rate: no accuracy entry; why it matters, REF-1083 B |
| Wrist SpO2 (estimated optically) | Nothing decision-grade in the base | Altitude, illness or acclimatisation decisions (SYNTH-06) | REF-1248 A-: Apple Watch Series 6 only, one low-quality review: mean absolute difference up to 2.0% vs a pulse oximeter, limits as wide as −5.8% to +5.9% in some studies; no entry tests altitude or exercise (gap) |
| Handheld lactate (measured, with a unit offset) | The same meter and protocol, compared with itself (SYNTH-53) | Another meter, a lab value or a published mmol/L table | REF-1081 B: four handheld meters −7%, +7%, −10% and +42% vs a Biosen reference, 95% prediction interval 0.60-0.87 mM wide at 3 mM; REF-030 B: six portable analysers read low, repeatability generally within 0.5 mM between about 1 and 10 mM |
Decision table
| Situation | Do this | Support and level | How strong |
|---|---|---|---|
| "My HRV is low this morning" | One night is context: read the 3-day direction against the 28-day baseline on the same device, check the usual causes (a late hard session, alcohol, a short night) and ask how they feel; the response ladder is SYNTH-30 | REF-1251 B- (single-night limits about −12 to +10 ms on the best device tested, −15 to +12 ms on the Garmin Fenix 6); REF-1193 B+ (a maximal session ending 2 h before habitual sleep onset went with HRV 14.1% lower that night than a light one; WHOOP cohort, observational); REF-361 B+ (averaged trend) | Moderate on the device error; the response is practice |
| "My watch says my VO2max dropped" | Say it is an estimate with about ±10 mL/kg/min of individual error from runs, more from a resting test, so a drop of a few points is inside it; read the same effort repeated instead (best 5-min power, pace or power at a fixed heart rate; SYNTH-78); no zone change | REF-1253 A- | Strong on the error; the response is practice |
| "My watch says I burned 3 000 kcal" | Order of magnitude only: do not eat back to the number or build a fuelling plan on it; on the bike use the power meter's kJ; judge intake by the body-mass trend, sessions and recovery (SYNTH-87) | REF-1254 B- (Huawei-funded); REF-1248 A- | Moderate that it is imprecise; its direction belongs to the device |
| "My scale says body fat is up 2%" | Not a change on any scale tested; look at the weight trend and how sessions and recovery go; never a body-fat target (SYNTH-54) | REF-1257 B (change limits about ±3.5-5 points, reconstructed from its tables) | Moderate (one careful study, non-athletes) |
| "CORE shows 38.9 °C" | A direction, not a threshold, and the true core may be higher; stop on how the athlete feels and functions (confusion, staggering, nausea, unusual weakness), not on the number (practice); do not count heat dose in CORE minutes (SYNTH-04) | REF-1258 B- (abstract) | Thin (one abstract, one sensor version, a team sport) |
| "My watch found a new LTHR or threshold pace" | Label it estimated; keep the zones until a field effort or a lab test agrees (SYNTH-38, SYNTH-53) | REF-1232 B (Huawei supplied the equipment) | Moderate |
| "My Tymewear VT2 is lower than my zones" | Treat the device's VT2 as a floor until a field effort or lab test agrees; its VT1 is usable as an estimate, with easy work kept comfortably below it (SYNTH-51) | REF-1259 B- | Thin (n = 15, one shirt generation, treadmill at altitude) |
| "My pedals and my trainer disagree" | Expected: choose one meter for tests and zones, ride once with both recording to learn the offset (practice), retest after a meter change; neither is shown to be right (SYNTH-56) | REF-1255 B- (abstract); trainers: no entry | Low-moderate; the offset ride is practice |
| "My tracker says I slept badly, I feel fine" | Believe the athlete; use the device for bedtime, wake time and the total-sleep trend; ignore stage minutes (SYNTH-07) | REF-1252 B+; REF-364 B | Moderate |
| "My readiness score is red" | Ask what happened and how they feel; a composite that was never validated cannot decide on its own (SYNTH-30) | REF-1248 A- (authors' statement); REF-1242 B+ | Moderate that it is unvalidated; the response is practice |
| "I changed watch or ring" | New device, new baseline: rebuild the 28-day baseline, or wear both for a few weeks, before reading trends; do not compare absolute HRV or resting heart rate across the change (practice) | REF-1251 B- (brands average different windows; HRV error 6-16% by model); REF-364 B | Practice built on agreement data |
| "Wrist heart rate on the bike looks wrong" | Use a chest strap for heart-rate zones, decoupling and fatigue tests on the bike (practice) | REF-1249 B+ (−4.55 bpm cycling, abstract); REF-1250 B | Moderate on the direction |
Key Findings — what the evidence shows
Measured beats estimated, and composites are untested. A living umbrella review of 24 systematic reviews (249 validation studies) found heart rate within about ±3% (narrative), resting HRV close to ECG for Polar devices (ICC 0.98-1.00) and worse with movement, watch VO2max with individual limits of about ±10 mL/kg/min after an exercise test and ±15 after a resting test, energy expenditure under-read by about 3 kcal/min with limits of −13 to +7 kcal/min (Fitbit), and total sleep over-read; only 34 of 310 devices (11%) had any validation, half the reviews were rated low quality, nothing could be pooled, and the authors state that readiness, stress and "body battery" scores have never been formally validated (REF-1248, A-).
The same order — what the sensor measures, then what an algorithm estimates, then composites — is an expert's hierarchy too (REF-397, C+).
The condition sets the error of the same sensor. Wrist optical heart rate was close on average at rest, in sleep and running and differed, probably reading low (the abstract states no sign convention), in cycling (−4.55 bpm) and lifting (−7.26 bpm) (44 studies, devices before 2020; REF-1249, B+, abstract, no limits of agreement).
A chest strap matched a 12-lead ECG from rest to exhaustion (heart-rate concordance 1.00), while its DFA a1 limits widened from +11.6/−9.9% at low intensity to +58.1/−40.9% at high (REF-1250, B, 25 recreational adults).
Overnight: resting heart rate is the steadier number, HRV depends on the model, and each brand reads its own part of the night. Resting heart rate had a mean absolute error of 0.98-1.78 bpm on four devices; HRV MAPE 5.96-16.32% across five; Garmin's resting heart rate is the lowest 30-min average of the day, Polar's values come from the first 4 h of sleep, WHOOP weights toward late slow-wave sleep; errors survived each person's own baseline (REF-1251, B-, 13 healthy adults, 536 nights).
Over one night, WHOOP 3.0's HRV limits were ±7.6 ms and the other five devices' ±55.2 to ±92.0 ms (REF-364, B).
The shifts athletes read are of the same order: a maximal session ending 2 h before habitual sleep onset went with nocturnal heart rate 3.9 bpm higher and HRV 14.1% lower than a light one, averaged over many nights (REF-1193, B+) — for a single night about twice WHOOP 4.0's resting-heart-rate error (mean absolute error 1.78 bpm) and less than twice its HRV error (MAPE 8.17%) (the base's arithmetic on REF-1193 and REF-1251, not a finding of either paper), which is why one night is context.
Algorithm estimates can miss one person by more than the changes athletes look for — watch VO2max (±10 mL/kg/min, REF-1253, A-), watch LTHR (9-11 bpm, REF-1232, B), watch calories (REF-1254, B-), a breathing-shirt VT2 (REF-1259, B-), scale body fat (±6.5 points at best, limits reconstructed from its tables, REF-1257, B) and a skin-sensor core temperature in the heat (REF-1258, B-); numbers in the trust table.
Instruments that measure are repeatable within a unit and offset between units. Power meters repeated within a CV of 1.2% but sat −0.9 ± 3.2% from true power, 6 of 54 beyond ±5% (REF-1255, B-); handheld lactate meters −10% to +42% from a laboratory analyser (REF-1081, B; REF-030, B); Stryd running power repeated within a CV of 4.3% and three other systems did not, with no comparison of absolute watts in the abstract (REF-1256, B-).
These data are what the same-device rule rests on; the rule itself is practice.
The reference has error too. Fifteen laboratory gas-analysis systems erred on VO2 by 1.1-13.3% against a simulator (REF-169, B); gas-exchange thresholds located by eye repeated with CVs of 2.4-8.0% (REF-1259); a pill reads the temperature where it sits in the gut (the base's note on REF-1258's criterion; the abstract does not say how this was handled).
Acting on device numbers is barely tested. HRV-guided training favoured submaximal markers (g = 0.296) but not performance (g = 0.079, not significant) over 8 small studies, mostly with fewer hard sessions in the guided arms (REF-994, A-); no readiness marker, subjective or device, has been validated against the next session (REF-1242, B+).
It does not show the accuracy of 2026 devices and firmware or of most models athletes wear, whether a watch tracks one athlete's change, or whether acting on a device number improves performance; those rows are practice or gaps (Evidence gaps).
Conflicts and how they are resolved
Practical Application — how the assistant says it
Evidence gaps
Limitations
How this entry was checked and changed
(the direction — what a sensor measures at rest agrees with a reference best, an algorithm's estimates worst, composite scores have never been validated, and agreement belongs to a device model, a condition and a firmware — rests on an A- living umbrella review that could pool nothing and takes each number from one included review REF-1248, an A- meta-analysis of 14 small studies, 11 at high risk of bias REF-1253, B+ meta-analyses that report group means only REF-1249 REF-1252, B validation studies REF-1250 REF-1257 REF-364 REF-1232 REF-873 REF-1081 REF-030, six B- validation studies REF-1251 REF-1254 REF-1255 REF-1256 REF-1258 REF-1259 and a C+ expert hierarchy REF-397. Above C+ because the gradient from measured to estimated to composite is what the agreement data show across outcomes; below B because the default — which readings may move a session and which are context — is practice, no study tests whether acting on a device number changes performance except HRV-guided training, which moved submaximal markers and not performance in 8 small studies (REF-994, A-), most validations are one session or one night on one firmware in healthy young non-athletes, and four of the twelve new entries are read from the abstract)