Two wearables worn simultaneously report different step counts, calorie figures and sleep durations. The disagreement is expected, because almost everything they display is inferred rather than measured directly.
Steps are inferred from movement patterns
An accelerometer records motion, and software decides which patterns constitute a step. That decision involves thresholds chosen by the manufacturer and never published.
A device tuned to avoid missing steps will count arm movements that were not walking, while one tuned conservatively will miss slow or irregular walking.
Placement compounds this, since a wrist and a pocket experience entirely different motion during the same walk.
Heart rate at the wrist is an optical estimate
Optical sensors shine light into the skin and measure changes in reflection caused by blood flow, then extract a rate from that signal.
Motion disturbs the signal, so accuracy is best at rest and worst during activity involving wrist movement, which is precisely when readings are most wanted.
Fit, skin tone, tattoos and temperature all affect the optical path, which is why the same device performs differently across people and across sessions.
Energy expenditure is a model, not a measurement
Calorie figures are calculated from movement, heart rate and entered personal details using a model of metabolism, which is an estimate at every stage.
Different manufacturers use different models, so figures diverge substantially even where the underlying sensor data broadly agrees.
The value is therefore in relative comparison over time on one device, rather than in the absolute number or in comparison with anyone else's device.
Sleep staging is inferred from proxies
Clinical sleep staging uses brain activity measurement. A wearable infers stages from movement, heart rate and its variability, which correlate with sleep depth without determining it.
Total sleep duration is generally more reliable than the breakdown into stages, since detecting stillness is easier than distinguishing what kind of sleep it represents.
Algorithms change under the same name
Firmware updates revise these calculations, so a device can report differently for identical behaviour before and after an update, sometimes retroactively.
Because the methods are proprietary, there is usually no notice of what changed or by how much.
The consistent reading is that these devices are useful for tracking personal trends and unsuitable as measuring instruments, which is a distinction their presentation rarely makes clear.