Running power accuracy, device by device

Why running power is a different kind of number from cycling power

A bicycle power meter contains strain gauges. It measures the force you apply to a crank or pedal and the angular velocity of that crank, multiplies them, and reports watts. The measurement is physical and the unit is unambiguous.

Running power is not that. No consumer running device measures the force you apply to the ground. What they do is infer it - from accelerometry at the wrist or the foot, from pace, from estimated vertical oscillation and ground contact time, from barometric gradient - and then feed those inputs into a proprietary model that outputs a number labeled "watts". Different manufacturers use different inputs and different models, which is the reason two devices on the same run report different wattages and neither can be declared wrong: they are not measuring the same thing, and neither is measuring the thing the unit implies.

This matters for how you should use the number, not just whether to believe it. A quantity that is internally consistent but not comparable between devices is still useful - as a pacing index relative to your own history. It is not useful as an absolute physiological figure, and it cannot be compared with a training partner's. What the metric is supposed to represent is covered in running power explained; this page is about what has actually been measured.

The one validation that exists, and its limits

Imbach et al. (2020) is the only peer-reviewed validation of a consumer running power meter in our library. Six recreational runners ran an incremental test from 8 to 20 km/h on a 200 m indoor track, measured against 500 Hz force platforms, a Cosmed K4b2 metabolic analyzer and 400 Hz motion capture - a genuinely rigorous reference setup.

The authors' own stated limits are as important as the result, and we quote them rather than summarize them away: six participants, a correction function estimated from that same small dataset, no maximal or supramaximal speeds, no slopes or changes of direction, and proprietary algorithms the authors could not inspect. A correction derived from the data it is then applied to is the methodological point to be most careful about - it is not independent validation in the sense a reader would assume from the word.

The second relevant paper is often cited as if it were a second validation, and it is not. Berzosa et al. (2024) put a Stryd foot pod against a Garmin fenix 7S Solar with an HRM-PRO on a 2.5 km trail course with 195 m of climbing, run twice a week apart by five participants. There was no gold standard in that study. It measures how repeatable each device is and how far the two disagree with each other - it cannot say which one is closer to the truth. Anyone citing it as evidence that one brand is more accurate than the other has misread it.

Beyond those two there is the 2025 systematic review of 13 studies and 273 participants covering wearable VO2 max and threshold estimation, which includes foot-worn inertial units among its devices and reports the same structural problem: undisclosed algorithms prevent cross-device comparison, and none of the studies tested the latest models. That review was screened and extracted by a single reviewer, which the authors state.

What the watches on this page actually publish

All three devices below report running power natively, from the wrist, with no foot pod required. None of those figures has been validated by anyone.

Garmin Forerunner 970. Garmin's specification gives it running dynamics, multi-band GNSS with SatIQ, 26 stated GPS hours and 15 smartwatch days. Running power on this watch is derived on the wrist, and adding Garmin's own HRM 600 chest strap adds the stride metrics the model uses - which is an input change, not a validation. It costs $28.85 per stated GPS hour.

COROS PACE 4. COROS states 41 hours of All-Systems GPS and 19 days of daily use in a 32-gram case. COROS reports running power from the wrist too, on its own model, and the brand's wider caveat applies: the only independent test of a COROS optical sensor flagged divergence during intensity changes, and heart rate is an input to several of the derived figures.

Polar Vantage V3. The most sensor-dense watch of the three - dual-frequency GNSS, skin temperature, on-wrist ECG and offline maps - which means its power model has the most inputs available to it. More inputs is not the same as more accuracy, and Polar publishes no per-model validation of its power figure either.

How to use a number nobody has checked

Three rules that follow from the record rather than from opinion.

Compare it only with itself.Your watch's running power is internally consistent enough to pace a tempo effort or hold an even output up a hill. Across brands, or against a friend's watch, it means nothing.

Do not set training zones from it alone. The 2025 threshold study of 100 runners found that even a watch-derived lactate threshold - a much better-studied output - returned a usable result only 65.2% of the time on the best device tested, and that protocol used a chest strap rather than the wrist. A power-derived zone inherits that uncertainty and adds a model nobody has inspected.

Expect it to change under your feet. These models are updated by firmware. A number that moves because the algorithm moved, with your physiology unchanged, is the clearest demonstration that it is a model output rather than a measurement - and it is why the 2024 umbrella review lists undisclosed, periodically-updated proprietary algorithms as a structural obstacle to validating consumer wearables at all. That review found roughly 34 of 310 identified consumer wearables have ever been validated for anything.

If what you actually want is a measurement rather than a model, the one upgrade that delivers it is a heart-rate sensor off the wrist - see better heart rate accuracy. The related model outputs are at VO2 max accuracy by device and training load explained.

What to buy if running power is the reason

• If you want the one validated device: it is a foot pod, not a watch, and the validation is six runners with a self-derived correction. We do not stock a pick for it, because we will not put a buy button on the back of a study that thin.
• If you want a usable pacing index: any of the three watches here reports one natively. Pick the watch on everything else - battery, positioning, platform - because the power figure is unvalidated on all of them equally.
• If you want a real measurement to train on: buy a chest strap and train by heart rate and pace. Both are measured; running power is modeled.