What a watch actually observes
A wearable may detect acceleration, changes in reflected light associated with the pulse, and other signals. It can combine those signals with information such as age or body size. The calorie number is the result of an estimation model connecting those inputs to energy expenditure. The watch does not directly observe every physiological process contributing to that expenditure.
This distinction explains why a convincing display can still be uncertain. Showing 347 rather than approximately 350 calories is a formatting choice. It does not mean the underlying estimate is accurate to a single calorie. The same applies to a smooth graph: frequent estimates can create visual continuity without removing measurement error.
What validation studies have found
O’Driscoll and colleagues’ systematic review found that accuracy varied by activity type and that combining physiological sensors with movement data could improve estimates. Fuller and colleagues’ review similarly distinguished the performance of step counts, heart rate, and energy expenditure. Good performance on one measurement did not establish good performance on all of them.
A separate review of combined-sensing Fitbits examined both accuracy and precision against reference methods. That distinction matters because an estimate can be close on average while individual observations vary substantially. These reviews cover particular devices and study periods. They support caution about treating calorie estimates as exact; they do not provide a current ranking of every product on the market.
Average accuracy can conceal individual error
Consider an invented test involving four exercise sessions, each using 300 calories according to the reference method. A device estimates 200, 250, 350, and 400. Its average estimate is exactly 300. Its average signed error is zero. Yet every session estimate is wrong by at least 50 calories.
This is why a statement such as “accurate on average” needs clarification. Positive and negative errors can cancel. Absolute errors, the spread of differences, and the conditions tested help describe whether individual readings are close enough for the intended purpose.
There is a second problem with averages across people. A model may perform differently for individuals who were sparsely represented in its development or validation data. A population-level result does not establish that the error is identical for everyone. It also does not tell a particular person the direction of their own error.
The comparison method has a scope too
Validation can use different reference approaches. Indirect calorimetry estimates energy expenditure from respiratory gas exchange during a measurement period. Doubly labeled water can estimate total energy expenditure over a longer free-living period. They answer related but different questions, and neither comparison should be described without its timeframe and setting.
Comparing two consumer watches is a different exercise. Agreement between them does not show that either is accurate. They could share similar assumptions or errors. Disagreement does establish that the outputs differ, but it does not identify which is closer to the underlying quantity.
Active calories and total calories are not the same output
A display may report estimated expenditure associated with activity, total daily expenditure including resting needs, or a workout total using a manufacturer-specific definition. Comparing those numbers without checking the labels can create an apparent disagreement that is partly definitional.
For example, a fictional app might show 400 active calories and 2,100 total calories. Those numbers need not conflict. The first is one component of the second under that app’s model. A different app may allocate resting expenditure during workouts differently, so even similarly named totals may not be directly comparable.
Are trends more useful than single readings?
Sometimes, but that is a conditional claim. If the device’s error stays reasonably stable under similar conditions, repeated estimates may help describe a pattern. If the activity changes from walking to cycling, the device fit changes, or the algorithm is updated, the error may also change. A trend in the output can therefore reflect both real behavior and changes in estimation.
A wearable can still be useful for noticing that a day involved more movement or that a routine became more consistent. That use asks less of the calorie estimate than treating it as a precise physiological budget. The relevant question is how much accuracy a particular interpretation requires, and whether the available validation actually supports that interpretation.
