The honest synthesis
Most "common triggers" lists you'll see on wellness blogs cite the same five-or-so items — chocolate, red wine, cheese, caffeine, cured meats — with shaky evidence. The actual prospective research is more disciplined and surprisingly consistent in what it does and doesn't support.
The key reference point: Holsteen et al. 2020 (Headache) prospectively followed 178 patients with daily logging of common triggers — including caffeine, alcohol, sleep, stress, menstruation, and self-prediction. The resulting within-person C-statistic was 0.56, barely above chance. The paper's verbatim conclusion: "episodic migraine attacks were not predictable based on self-prediction or on self-reported exposure to common trigger factors."
That negative result reframes everything. The honest reading is that population-level trigger associations are much weaker than the folk wisdom suggests. The triggers that do survive prospective testing fall into a small set — and even they are modest predictors, not deterministic causes.
The grading
Strong evidence — what Hermly actually models Strong
Four triggers have consistent prospective evidence across multiple cohorts and feed Hermly's prediction model directly. Each one adds 2–5 AUC points; in combination they explain the bulk of the personalised forecast.
Barometric pressure migraine: what research shows
StrongPressure drops are one of the few weather signals tied to migraine in peer-reviewed research. What's solid, what isn't, and how Hermly uses it.
Menstrual migraine: the hormone-cycle connection
StrongRoughly half of women with migraine see attacks cluster around their period. The estrogen-withdrawal mechanism, cycle-day risk windows, what Hermly tracks.
Sleep and migraine: too little, too much, or just irregular
StrongBoth short and disrupted sleep predict next-day migraine in published cohorts. The cited threshold: sleep below 75% of your personal baseline.
Stress and migraine: the strongest single self-report signal
StrongPerceived stress has the most consistent evidence of any self-reported migraine trigger. Houle 2017 reached AUC 0.65 using just stress + current state.
Mixed evidence — real but personal Mixed
Two triggers show paradoxical behaviour at the population level — sometimes protective, sometimes triggering, depending on dose and timing. Where they matter, they matter at the individual level; the cohort signal is muddy.
Caffeine and migraine: trigger, treatment, or both
MixedCaffeine has a paradoxical relationship with migraine — it can abort an attack at low doses and trigger one at high doses or withdrawal.
Exercise and migraine: trigger, treatment, or both
MixedExercise is mildly preventive on average and acutely triggering in some — primarily exertional migraine. The dose-response is real and personal.
Weak evidence — folklore-heavy, prediction-poor Weak
Two of the most-cited triggers consistently underperform in prospective testing. They aren't zero at the individual level for some patients, but population-level associations are modest at best, and self-report is heavily confounded by recall bias and direction-of-causation problems (cravings during prodrome look like food triggers).
Migraine food triggers: why the lists are mostly wrong
WeakMost 'top food triggers' lists cite the same foods with weak evidence. Holsteen 2020 tested them all — within-person C-statistic only 0.56.
Dehydration and migraine: what's solid and what's folklore
WeakDehydration is a real but smaller trigger than weather, sleep, or cycle. What controlled studies show, and what 'drink more water' doesn't solve.
How Hermly uses this grading
The trigger pages double as a transparency document for Hermly's model design. We deliberately read the strong-evidence triggers from Apple Health, WeatherKit, and your attack history. We deliberately don't add daily prompts for food, hydration, or caffeine logging — the AUC gain doesn't justify the friction.
That self-imposed restraint reflects a broader product principle: more features ≠ more accuracy. See the methodology page for the 26-feature breakdown.
Common questions
How many triggers does a typical person actually have?
Far fewer than the "common triggers" lists suggest. Across prospective cohorts, most patients with identifiable personal triggers have 2–4 reliable ones. The rest of the long list usually doesn't survive personal n-of-1 testing.
Why does my "trigger" not show up in Hermly?
Hermly models the population-validated triggers (sleep, cycle, weather, stress, recent attack history). If your personal trigger is something else — a specific food, a screen-time pattern, a specific environment — Hermly may underweight it. The Doctor Report's per-factor associations surface what the data does see.
Should I try an elimination diet?
Not as a default. See our food trigger page for the discussion. Structured elimination under a registered dietitian's guidance can occasionally reveal a true trigger; self-directed restriction usually trades cost for noise.
Where can I go deeper on the methodology?
The methodology page lays out the 26 features, the cohort + per-user model architecture, and the published papers Hermly builds on. The AUC-ROC entry explains what the accuracy numbers mean.