Heart rate variability has become one of the more discussed wearable metrics over the last few years, mostly in the context of recovery and stress monitoring. What gets less attention is how HRV tracks across the menstrual cycle, a pattern that is real, measurable, and useful for cycle phase inference, but subtle enough that it is easy to misread if you are not modeling it carefully.
This post is about what that signal actually looks like and why extracting it requires more than reading a daily HRV number off a device app.
What drives HRV variation across the cycle
HRV reflects the balance between sympathetic and parasympathetic branches of the autonomic nervous system. Higher RMSSD (root mean square of successive RR-interval differences) generally indicates greater parasympathetic dominance, associated with states of rest and recovery. Lower RMSSD indicates more sympathetic activation.
The menstrual cycle modulates autonomic tone through two main hormonal pathways. Estrogen broadly increases parasympathetic activity, which tends to raise HRV in the follicular phase as estrogen rises toward its pre-ovulatory peak. After ovulation, progesterone's rise is associated with increased sympathetic tone, which tends to suppress HRV in the mid-to-late luteal phase. The progesterone effect also drives the well-known resting heart rate elevation in the luteal phase, typically 2-5 beats per minute above follicular baseline in regularly cycling individuals.
So in theory, you would expect a pattern where HRV is relatively elevated in the follicular phase, peaks or plateaus around ovulation, and then declines through the luteal phase before recovering as menstruation begins. In practice, you see this in individual data, but it is buried under noise from daily life. The hormonal HRV signal is on the order of 5-15 milliseconds RMSSD range across the cycle for most individuals. One stressful afternoon, a bad night of sleep, or a moderate workout can produce a same-magnitude perturbation. The challenge is not that the cycle signal does not exist; it is that it is small relative to the day-to-day noise floor.
Why the segment you sample matters enormously
Consumer wearables typically report a single daily HRV value. The methodology varies by device: some compute it from a brief morning measurement, some aggregate overnight data, some take a 5-minute resting window. This is a bigger source of variance than most cycle-tracking discussions acknowledge.
For our purposes, the overnight sleeping RMSSD, specifically the period roughly between 2 AM and 5 AM when the autonomic system is least disrupted by daytime activity, is the most stable and most hormone-correlated signal. A 2 AM reading is not contaminated by the prior day's exercise afterburn, which can suppress HRV for 12-18 hours after moderate intensity training. It is not contaminated by the pre-sleep anxiety or the first-morning cortisol pulse.
In our internal pilot data, the cycle-phase signal in overnight-segmented HRV was substantially cleaner than the same participants' reported daily HRV aggregate. The correlation between cycle phase and overnight RMSSD was consistent enough to contribute meaningfully to phase prediction. The all-day aggregate often added more noise than signal in the same participants.
This is one reason device selection matters for the kind of inference we are doing. A device that reports overnight HRV computed from all sleep-stage segments is giving us more signal than a device that only reports a single morning spot measurement. The raw data access also matters: ideally we want the underlying RR interval stream, not just the derived RMSSD number, so we can apply our own segmentation and outlier filtering. Not all consumer wearables expose that.
Baseline modeling: you cannot skip it
The cycle-correlated HRV pattern is visible most clearly when it is expressed as a deviation from each person's own baseline, not as an absolute RMSSD number. A person whose typical overnight RMSSD is 85 ms and who drops to 72 ms in the luteal phase is showing a meaningful phase-correlated shift. A different person whose typical overnight RMSSD is 35 ms would show a different absolute range with a different percentage change. Comparing these two people's raw numbers to a reference table would obscure the pattern in both cases.
We build a per-user baseline using the first complete follicular phase we have data for, then compute each day's HRV as a normalized deviation from that reference period. This approach works well for regularly cycling individuals with enough data to establish the follicular reference. It breaks down for users with highly irregular cycles, where there is no stable follicular period to anchor against, and for new users in the first one or two cycles of data collection, where the baseline is still being established.
We want to be clear about the initialization period: for the first 30-45 days of data collection, the HRV contribution to phase inference is weighted lower because the personalized baseline is not yet reliable. During that period, the temperature signal carries more of the inference weight. As the baseline stabilizes, HRV becomes progressively more informative and its weight in the model increases accordingly.
The confounders you have to account for
Even with proper baseline modeling and overnight segmentation, three confounders consistently degrade HRV-based phase inference when not handled explicitly.
Illness. Acute illness, including subclinical viral illness, reliably drops HRV before other symptoms become obvious. An HRV drop during what the model expects to be the follicular phase can look like an early luteal shift. We flag anomalously large HRV drops that do not coincide with expected phase-transition timing and hold their phase-prediction contribution with reduced confidence rather than treating them as confirmatory signal.
Acute training load. High-volume or high-intensity training within 24 hours suppresses overnight HRV in ways that overlap with the luteal suppression pattern. We incorporate the movement intensity channel to detect training-load days and adjust the HRV feature weighting on those days. A training athlete wearing a biosensor needs different HRV normalization than a sedentary user.
Alcohol. Even moderate alcohol consumption reliably suppresses overnight HRV in most people. A Saturday night that produces a Sunday morning HRV drop does not indicate a phase shift. We cannot detect alcohol consumption directly, but the overnight heart rate pattern that accompanies it (elevated overnight HR plus suppressed HRV with normalized movement) is distinctive enough that we can partially identify these nights and weight them down.
What HRV contributes to the full model
By itself, HRV would not be a reliable cycle phase predictor. Combined with wrist temperature and resting heart rate in the multi-channel LSTM, it improves phase boundary precision, particularly around the follicular-to-periovulatory transition where temperature alone is still in a relatively flat region. The HRV drop toward the end of the follicular phase, as sympathetic tone begins to rise in advance of the LH surge, provides a few days of earlier signal than temperature alone in many individuals.
The signal is not this clear in everyone. Some users show very little HRV variation across the cycle, either because their hormonal modulation of autonomic tone is less pronounced or because their lifestyle noise floor is high enough to mask it. For those users, the model's phase inference relies more heavily on the temperature channel, and the HRV input becomes less informative. The model learns this per-user over time: when HRV and temperature predictions consistently diverge for a given user, the model learns to down-weight HRV for that individual's inference.
That is the honest account of what the signal shows: genuinely informative, non-trivially noisy, and only useful when modeled with a per-user baseline and careful segment selection. Anyone claiming that daily HRV alone reliably tracks cycle phase is either working with unusually clean data or not accounting for the confounders carefully enough.