Why Baseline Drift Confounds Personal Tracking Data

Anyone tracking a physiological measure over months encounters slow movement in the underlying average. That drift is frequently misattributed, and understanding it changes how tracking data should be read.

Baselines are moving, not fixed

Resting heart rate, variability, sleep duration and body weight all shift gradually with season, training load, ambient temperature and life circumstances.

These influences operate on timescales of weeks to months, which is the same timescale over which most people evaluate an intervention.

The result is that a comparison between two periods captures both the intervention and whatever drift occurred during the same window, with no way to separate them.

Seasonal patterns are larger than expected

Several tracked measures show consistent seasonal variation. Sleep duration, activity levels and skin temperature all move with daylight length and ambient conditions.

Someone beginning a practice in autumn and assessing it in winter is comparing across a seasonal boundary. The measured difference includes that shift.

Running the same comparison in the opposite season would produce a different result, which is one reason personal findings often fail when repeated later.

Devices and algorithms change underneath

Wearable manufacturers revise their processing algorithms through software updates. A revision can shift reported values without any change in the wearer.

These changes are rarely announced in detail, and historical data is not always recalculated, which can leave a visible step in a long record.

Replacing a device introduces the same problem more sharply, since sensor placement and processing differ. Data from before and after a replacement is not a continuous series.

Habituation changes the measurement itself

Early in tracking, people often measure more carefully and at more consistent times. Consistency tends to decline as novelty fades.

A change in measurement conditions produces a change in measurements. Later readings taken at varied times will show more scatter and possibly a different average.

Behaviour also adapts to being observed, and that adaptation weakens over time. Part of any early improvement may reflect attention rather than the intervention.

Designs that reduce the problem

Alternating periods with and without the practice, repeated several times, spreads drift across both conditions rather than loading it onto one.

Keeping measurement conditions fixed, including time of day, position and device, removes a substantial source of variation that is otherwise mistaken for signal.

Recording context alongside numbers is equally valuable. Travel, illness and workload explain more anomalies than most interventions do, and they are invisible in the data itself.