How Accurate Is Apple Watch Calorie Tracking? The Error Is About 28%

Short answer: the calorie number is the least reliable thing your watch tells you. A 2025 meta-analysis of 56 studies put the mean absolute percentage error for energy expenditure at 27.96%, against 4.43% for heart rate and 8.17% for step count. An earlier Stanford study across seven wrist devices found no device achieved an energy expenditure error below 20%, with the worst above 90%.

The important part is that this is not a defect. Heart rate is measured. Calories are estimated from that measurement plus a chain of assumptions about you, and each link in that chain adds error.

The three numbers, ranked

Choe and Kang pooled 56 studies comparing Apple Watch output against reference instruments. The ordering of those three figures is stable across the literature and it maps exactly onto how directly each quantity is observed.

Heart rate is measured. A green LED illuminates the tissue, a photodiode reads how much light returns, and blood volume changes with each beat. The device is watching the thing itself. Error accumulates only from motion artefact and poor contact.

Steps are detected. An accelerometer sees a pattern and a classifier decides whether that pattern was a step. This is inference, but the thing being inferred is discrete and has a clear signature.

Calories are modelled. There is no sensor for energy expenditure. The watch takes heart rate, movement and the personal data you entered, and runs it through an algorithm that outputs a number in the same font as the two measured ones. That last part is the design decision worth being annoyed about.

Why the model cannot be much better

Shcherbina and colleagues at Stanford tested seven devices against clinical gold standards: twelve-lead ECG for heart rate, and indirect calorimetry, measuring the gas you actually exhale, for energy expenditure.

Most wrist-worn devices adequately measure HR in laboratory-based activities, but poorly estimate EE, suggesting caution in the use of EE measurements as part of health improvement programs.

Shcherbina, Mattsson, Waggott et al, Journal of Personalized Medicine, 2017

The inputs the algorithm needs and cannot obtain are the reason.

Your resting metabolic rate varies by 10–15% between people of identical height, weight, age and sex. It is a function of lean mass, organ mass and thyroid status, and the watch knows none of these. It uses a population equation.

Your body composition. Two people at 80 kg with different muscle mass burn measurably differently. Your watch has one number, and you typed it in.

Your movement economy. A trained runner uses less oxygen at a given pace than an untrained one. That efficiency is exactly what training builds, so the fitter you get, the more your watch overestimates.

Non-locomotor work. Carrying, lifting, resistance training, cycling uphill. Heart rate rises, the accelerometer sees little, and the model has to guess.

What 28% means at your kitchen table

If your watch says you burned 600 calories, a 28% mean error means the honest interval is roughly 430 to 770. And that is the average error. Individual sessions can be much further out, and the error is not random noise: it is systematically biased for a given person, because your particular metabolic rate and movement economy differ from the population model in a consistent direction.

Which produces a specific practical trap. If you eat back the calories your watch reports and you are one of the people it overestimates, you are eating a few hundred calories a day that were never burned. Over a month that is the entire deficit you were trying to create, and the watch will keep telling you the workouts happened.

This is the single most consequential inaccuracy in consumer fitness technology, and it is worth more attention than the sleep-stage debates that get more coverage.

Active calories, total calories, and the number people misread

Before blaming the algorithm, it is worth checking which number you are looking at, because two of the three most common misreadings are not accuracy problems at all.

Active calories are the estimated expenditure above resting. Total calories add your estimated basal rate, which for most adults is somewhere between 1,400 and 1,900 a day and dominates the total.

The confusion this creates runs in both directions. People compare their watch's 400 active calories with a gym machine's 500, not noticing the machine reports gross expenditure including the basal component. And people trying to manage intake sometimes count total calories as though they were earned by exercise, which double-counts the basal rate their food target already accounts for.

The 28% error applies to the estimate as a whole. But a meaningful slice of everyday disagreement between two devices is this definitional gap rather than either one being wrong.

What about sleep stages

The same measured-versus-modelled pattern shows up in the other metric consumer devices lead with, and it is worth knowing because the failure mode is identical.

Chinoy and colleagues tested seven consumer devices against polysomnography in 34 adults across three laboratory nights. The result splits cleanly in two.

Consumer sleep-tracking devices exhibited high performance in detecting sleep, and most performed equivalent to (or better than) actigraphy in detecting wake. Device sleep stage assessments were inconsistent.

Chinoy, Cuellar, Huwa et al, Sleep, 2021

Whether you were asleep is close to directly observable from movement and heart rate. Which stage you were in is an inference from the same signals about electrical activity in a brain the device cannot see. Same structure as heart rate versus calories, same conclusion: trust the duration, treat the breakdown as a trend at best. We went through what those stage numbers can support in how much deep sleep you actually need.

How to actually use the device

Use calories as a relative measure, never an absolute one. Same watch, same activity, week over week. "Today was 15% more than Tuesday" is information your watch can support. "I burned 612 calories" is not.

Trust heart rate, and use it. At 4.4% error it is the most reliable output you have, and it is the correct basis for pacing effort, structuring intervals and tracking fitness. Note that walking produced higher error than cycling in the Stanford data, because arm swing corrupts the optical signal. For hard intervals, a chest strap remains meaningfully better.

Do not eat back watch calories. If you are managing intake, use a total-energy estimate from bodyweight and activity level and adjust from what the scale actually does over three to four weeks. That feedback loop is measured. The watch's is modelled.

Judge fitness by trend metrics with a real physiological basis. Resting heart rate, and estimated VO2 max as a trend on the same device. We covered what that number does and does not mean in VO2 max by age, and the same "trend not value" logic in HRV by age.

Expect worse accuracy the fitter you get. Improving movement economy is invisible to the model, so the overestimate grows as you train. Counter-intuitive, and it follows directly from the mechanism.

Enter your data accurately and update it. Weight, height, age and sex are the model's only inputs about you. Stale weight is a permanent bias you can fix in thirty seconds.

Is any wearable better

Not meaningfully, on this metric. The Stanford study's most useful finding was that all seven devices failed the same way, across brands and form factors. Chest straps improve heart rate accuracy, and better heart rate feeds a better estimate, but the metabolic assumptions underneath are unchanged.

The only way to know your actual energy expenditure is indirect calorimetry, which means a metabolic cart in a lab and a mask on your face. Some sports clinics offer it. It is genuinely interesting once, and it will not fit on your wrist.

This article covers device accuracy for general fitness purposes. Nothing here is dietary or medical advice; if you are managing weight for a health condition, work from clinical guidance rather than a wearable.