Before you can build a model that infers cycle phase from wrist physiology, you have to understand what else is driving those same physiological signals. Wrist temperature and heart rate variability, the two primary inputs we use, are both strongly regulated by circadian rhythms. A 24-hour oscillation in wrist temperature spans roughly 0.4 to 0.6 degrees Celsius. A similar oscillation appears in autonomic tone, reflected in HRV patterns that peak during nighttime sleep and trough in the early afternoon. These circadian oscillations are substantially larger in magnitude than the menstrual cycle signals they overlap with. Getting cycle phase inference right requires understanding the circadian layer first, because otherwise you are trying to detect a small signal inside a large one you have not fully characterized.
The Circadian Temperature Arc
Core body temperature follows a predictable 24-hour cycle driven by the suprachiasmatic nucleus (SCN), the primary circadian pacemaker located in the hypothalamus. Temperature rises through the morning, peaks in the late afternoon, then drops progressively through the evening in preparation for sleep. The overnight trough, typically occurring in the early morning hours before waking, is the lowest point in this daily arc.
At the wrist, the same arc is present but modulated by peripheral blood flow. When the body initiates the nocturnal temperature drop, it does so in part through peripheral vasodilation: routing blood to the skin surface increases heat dissipation to the environment and accelerates core cooling. This means wrist skin temperature actually rises briefly in the early evening even while core temperature begins to fall, because vasodilation brings more blood to the skin. The wrist temperature then drops through the overnight period as the core temperature trough is reached.
This peripheral-core decoupling is relevant because it means the thermal signal at the wrist is not a simple attenuated copy of core temperature. It is a composite of the core temperature arc and the peripheral vasomotor response, both of which are under circadian control. Any model that treats wrist temperature as a direct proxy for core temperature misses this complexity.
Menstrual Hormones and Circadian Timing
Estrogen and progesterone interact with the circadian clock at multiple levels. Estrogen receptors are expressed in the SCN, and estrogen modulates the amplitude and period of circadian gene expression in clock cells. In practical terms, this means that the circadian temperature arc is not identical across cycle phases. The amplitude of the daily temperature oscillation tends to be slightly larger in the follicular phase compared to the luteal phase. The phase of the oscillation (when the trough occurs) can also shift by an hour or more across the cycle in some individuals.
Progesterone raises the thermoregulatory set point, which shifts the entire temperature arc upward during the luteal phase. This is the physiological mechanism behind the basal body temperature shift that cycle trackers have used for decades. At the wrist, this manifests as an elevated overnight trough: the lowest point of the daily arc is higher by roughly 0.15 to 0.30 degrees Celsius on average during the luteal phase compared to the mid-follicular phase, in individuals with good wrist sensor data coverage.
The LH surge and ovulation are also timed to the circadian clock. LH secretion from the pituitary has a circadian pattern, with the surge most commonly occurring in the morning to midday hours in people with typical sleep-wake schedules. Chronotype, the individual variation in preferred sleep timing, interacts with this: evening-preference individuals may show different LH surge timing than morning-preference individuals. This is one reason why menstrual cycle and circadian biology are not separable when studying either one from a time-series perspective.
HRV and the Circadian Autonomic Pattern
The autonomic nervous system, which drives HRV, is under strong circadian regulation. Parasympathetic (vagal) tone is highest during nighttime sleep, when the body is in a recovery and conservation state. Sympathetic tone is higher during waking hours. This drives a characteristic HRV pattern: high-frequency HRV (which reflects vagal modulation) peaks during the first half of the night and declines toward morning as sleep stages shift and cortisol begins its pre-awakening rise.
Menstrual hormones modulate this baseline autonomic pattern. Estrogen is generally associated with greater vagal tone, which tends to produce higher HRV during the follicular phase. Progesterone in the luteal phase is associated with a modest reduction in vagal tone, producing slightly lower HRV in the luteal phase compared to follicular baseline in some individuals. The effect size is individual-specific and often modest relative to the day-to-day variability driven by sleep quality, stress, and physical activity.
The practical challenge for cycle inference from HRV is exactly this: the menstrual hormone effect on HRV is a secondary modulation on top of the circadian pattern, which is itself modulated by sleep quality, exercise, and individual stress state. Separating the hormonal signal from these overlapping sources requires robust baseline modeling and is inherently more uncertain than the temperature-based inference in most individuals.
How We Handle the Circadian Layer in Our Models
We do not treat circadian variation as noise to be suppressed. We treat it as structure to be used. The circadian arc contains information about the individual's sleep timing, sleep regularity, and overall physiological state that is relevant for interpreting the cycle signal correctly.
For temperature, our processing extracts the overnight trough value after alignment on inferred sleep onset. This normalization removes the bulk of the circadian arc variation while preserving the night-to-night differences in trough level that reflect the hormonal state. The trough extraction also serves as an implicit quality filter: nights with unusual temperature profiles (e.g., abrupt mid-night rises from waking episodes) produce trough estimates with higher variance, which the uncertainty quantification layer reflects.
For HRV, we work with sleep-period HRV features rather than full-day values, for similar reasons. Daytime HRV is more contaminated by behavioral and environmental variation. Nocturnal HRV, extracted from the period between sleep onset and final waking, has a more stable circadian context and is more closely tied to underlying physiological state than spot measurements taken at arbitrary times of day.
The model does not receive raw wrist temperature values as input. It receives features derived from the overnight temperature segment after circadian normalization. This is not an attempt to remove circadian information entirely. It is an attempt to represent the temperature signal in a form where the cycle-relevant variation is more clearly separated from the day-to-day circadian timing variation that would otherwise confound the phase prediction.
When Circadian Disruption Affects Cycle Inference
Shift work, jet lag, and irregular sleep schedules all disrupt circadian rhythms in ways that propagate into wrist physiology. A person who works night shifts has a circadian temperature arc that may be inverted relative to their clock time, meaning the "overnight trough" in clock time is actually occurring during their active waking period. Applying a standard overnight extraction window based on clock time to this person's data would capture the wrong part of the circadian arc.
We use the accelerometer channel to infer actual sleep windows rather than assuming fixed overnight periods defined by clock time. This makes the processing more robust to atypical sleep schedules, though not perfectly so. Highly irregular sleep (someone who sleeps at different times every night) makes reliable sleep onset detection difficult, and the downstream temperature features are less reliable as a result.
Circadian disruption also interacts with the menstrual cycle itself. There is evidence that chronic circadian disruption, as experienced by long-term shift workers, is associated with longer and more irregular cycles and with increased menstrual irregularity. This creates a specific challenge for our models: individuals whose circadian rhythms are disrupted may also have atypical cycle patterns, so the two sources of complexity are correlated rather than independent.
The Layer Structure Matters for Understanding Limits
The reason we spend time on circadian biology in this post is that it is impossible to understand the limits of wrist-based cycle inference without understanding the circadian context in which all the relevant physiological signals sit. The menstrual hormone signal that we are trying to detect is small relative to the circadian variation it is embedded in. Our ability to detect it depends on accurate characterization of the circadian structure, which in turn depends on consistent sleep behavior from the individual wearing the device.
This is not a temporary limitation that better hardware will solve. It is a structural property of the measurement approach: we are reading indirect hormonal proxies through the lens of a circadian-regulated physiology. The people who get the most informative output from the system are those with regular sleep timing, because regular sleep creates a stable circadian structure against which the subtler hormonal modulations become more visible. This feedback loop shapes our product design: features that support regular wear timing and consistent overnight coverage are not just about data completeness. They are about circadian stability, which is the foundation on which the cycle inference rests.