How Accurate Are Fitness Tracker Calories?

The watch measures motion and a pulse, then models the calories#

How accurate are fitness trackers? At counting steps and reading heart rate, genuinely good. At telling you how many calories you burned, they are the weakest instrument on your wrist — and the sensor is not what's failing. In the largest synthesis of the question, covering 60 validation studies, 1,946 adults and 40 wrist- and arm-worn devices measured against indirect calorimetry and doubly labeled water, the authors' verdict on the entire category was that they "would be hesitant to consider any device sufficiently accurate" for energy expenditure¹.

The more useful finding is what kind of wrong it is. The folk rule on Fitbit calorie accuracy and Apple Watch calories-burned accuracy is that trackers flatter you, so you should knock something off the total. The evidence does not support that move. Errors run high on some devices and low on others, high in one activity and low in the next, and the current review of the Apple Watch specifically found energy-expenditure error "inconsistent and frequently large," with no distinct directional pattern across studies³. You cannot subtract a bias whose sign you don't know. This article is about where that number comes from and why its sign stays unknowable; whether the calories are worth spending at all is the pillar's question, in does exercise burn as many calories as you think.

Nothing on your wrist can measure a calorie#

A calorie is a unit of heat. Measuring energy expenditure means capturing that heat directly, or inferring it from the oxygen you consume and the carbon dioxide you produce — a mask, a sealed metabolic chamber, or, out in ordinary life, doubly labeled water. None of that fits on a wrist, and none of it ever will.

So the device measures what is cheap to measure: acceleration at the wrist, and a pulse read optically through the skin. Then it pushes those proxies through a model built on other people. The load-bearing assumption is the relationship between heart rate and oxygen uptake, which is real — within one person, across a calibrated range, pulse tracks oxygen consumption closely enough to be useful. The catch is in the review that mapped the method's limits: using the heart-rate-to-oxygen-uptake relationship to estimate energy expenditure in field conditions "provides a satisfactory estimate of energy expenditure on a group level, but is not very accurate for individual estimations"⁴.

That sentence is the whole article. The method works on averages. Your wrist is not an average.

Worse, heart rate moves for reasons that have nothing to do with the work you are doing. Dehydration and ambient temperature "can have a profound effect" on that relationship, and heart rate drifts steadily upward over a bout even when the intensity never changes⁴. A hot room, a short night, a coffee, the last mile of a long run: your pulse rises, the model reads the rise as work, and the calories tick up. The mechanics of turning heart rate into calories have their own article.

Steps are fine. Heart rate is fine. Calories are the failure.#

The most important thing to understand about your tracker is that it is not uniformly unreliable. A systematic review of 158 publications across nine brands scored every laboratory comparison against a ±3 percent band, and the three headline metrics came apart sharply²:

Measure	Lab comparisons landing within ±3%	Studies
Heart rate	56.5%	32
Step count	45.2%	126
Energy expenditure	9.2%	43

Nine percent. And no brand of wearable was within ±3 percent for energy expenditure more than 13 percent of the time. The device is not junk — it is good at the two things it physically measures and unreliable at the one thing it computes on top of them.

Your tracker is accurate at the two things it measures and unreliable at the one thing it calculates. The step count earns the trust that the calorie number then spends.

The error has no fixed direction#

Here is where the popular correction falls apart. Of 305 laboratory comparisons of energy expenditure in that review, 54.1 percent came in low, 36.7 percent came in high, and 9.2 percent landed inside the band. Split those by brand and the majority direction reverses depending on whose device is on your wrist²:

Brand	Which way most of its lab errors pointed
Withings	Under-read in 74% of comparisons (34/46)
Garmin	Under-read in 69% (37/51)
Polar	Over-read in 69% (9/13)
Apple	Over-read in 58% (18/31)
Fitbit	Split — 48.4% under, 39.5% over (157 comparisons)

The sign flips inside a single device, too, depending on what you are doing. During walking and stair climbing the Fitbit Charge HR over-read (effect size 0.78, 95% CI 0.27 to 1.29) while the Garmin Vivofit under-read the same activity (−1.24, 95% CI −1.86 to −0.62). And the most accurate device in the entire meta-analysis, the SenseWear Armband Mini, still produced study-level errors spanning −21.27 percent to +14.76 percent¹. That spread is not two studies disagreeing about a fact — it is one device behaving differently on different bodies doing different things, which is exactly what a population model does when you point it at an individual.

So the common fix — "my watch overstates, I'll take 20 percent off" — is not a conservative adjustment. It is a second guess stacked on the first. That 20 percent usually traces to the Stanford evaluation of seven wrist devices, which found that no device achieved an error in energy expenditure below 20 percent⁵. Read it again: that is a statement about the size of the error, not its direction. Subtracting 20 percent from a device that was already reading low doesn't remove the error. It doubles it.

Out of the lab, it gets worse#

Every number above came from validation studies where researchers followed the manufacturer's setup instructions, checked that the device sat in the right place, and entered the correct height, weight, sex and age. Your morning is not that. O'Driscoll's team make the point themselves: in free-living conditions, the absence of a researcher could yield greater error than their analysis observed.

It does. When devices were compared against doubly labeled water over whole days — the only reference that works outside a lab — they significantly under-read total energy expenditure (effect size −0.68, 95% CI −1.15 to −0.21, across 16 comparisons)¹. In free-living comparisons in the other review, just 18 percent landed within ±10 percent — a band three times wider than the lab one, and most estimates still missed it².

One honest complication, since it cuts against the internet's assumption rather than for it: pooled across all 104 comparisons, devices under-estimated on average (effect size −0.23, 95% CI −0.44 to −0.04). If there is a house tendency at all, it is to read low, not high. But that pooled figure came with heterogeneity of I² = 92.18 percent — a formal way of saying the studies scatter so widely that the average is a poor description of any of them. The mean of a set of errors pointing in opposite directions is not a correction factor. It is an artifact.

What the number is actually good for#

Not nothing. Three things survive the evidence.

Trends, not totals. For one person wearing one device doing one familiar activity, much of the error is likely to repeat itself week to week. That makes this Tuesday's number worth comparing against last Tuesday's, even though neither is worth reading as calories. The comparison carries information the absolute value doesn't.

A heart-rate sensor beats a bare accelerometer. In the sensor breakdown, devices using accelerometry alone significantly under-read (−0.36, 95% CI −0.55 to −0.17), while devices combining accelerometry with heart rate were statistically indistinguishable from the criterion (0.06, 95% CI −0.18 to 0.31)¹. Better. Not good.

Never as a budget line. Don't eat the number back — that conclusion belongs to the pillar, and it survives compensation as well as measurement error. The same problem in a different costume shows up on the intake side of the ledger, where tracking apps have their own accuracy ceiling and calorie counts are ranges rather than facts. A figure that hides its own error bar is not more useful than one that admits it; it is less.

The wrist is a superb place to count steps and a poor place to account for energy. Read the step count as data. Read the calorie total as a rumor with a decimal point.

FAQ#

Does my Apple Watch overestimate the calories I burn?#

Sometimes, and sometimes the reverse — which is why no correction factor works. Across nine brands, Apple devices over-read in 58 percent of laboratory comparisons, so the tendency exists, but it is a tendency, not a rule. The current review of the Apple Watch pooled 8 energy-expenditure studies in 270 people and found error "inconsistent and frequently large," ranging from 9.71 percent to 151.66 percent depending on the activity, with no distinct directional pattern.

Why are my tracker's step counts accurate when its calorie numbers aren't?#

Because steps are measured and calories are modeled. A wrist accelerometer detects the physical event of a stride; 45.2 percent of laboratory step comparisons landed within ±3 percent, and heart rate did better still at 56.5 percent. Energy expenditure has no physical event to detect — it is inferred from motion and pulse using a relationship calibrated on groups, which is why only 9.2 percent of those estimates fell in the same band.

Does a heart-rate sensor make a tracker's calorie estimate better?#

Measurably, yes, though not enough to trust the output. In a meta-analysis of 40 devices, those using accelerometry alone significantly under-read energy expenditure, while those combining accelerometry with heart rate were statistically indistinguishable from clinical reference measures. That is a real improvement in the average. It doesn't fix the per-person error, because the underlying heart-rate method was never accurate for individuals in the first place.

Sources#

This article was researched and drafted with AI assistance and reviewed for accuracy by the BurnWeek team. It is general information, not medical advice. How we research and correct our articles →