← All articles

How Ferrabelle predicts your next period (and how I check it)

By Chris Ayers 6 min read

Most cycle apps describe their predictions with one unexplained number, or no number at all. I work with data for a living, and that has always bothered me. If I am going to tell my wife her period is due Thursday, I want to know how often that kind of statement has been right, measured on data I did not get to pick.

So here is how the prediction works, and how I check it. Every accuracy number below is copied from the accuracy page at ferrabelle.com/accuracy, which is the page that changes whenever the engine’s published numbers change.

Where the first guess comes from

Before you have logged anything, the app has to start somewhere. It starts from assumptions grounded in published population research covering over 600,000 cycles, conditioned on age.

One piece of that is a 2019 paper in npj Digital Medicine that analyzed 612,613 ovulatory cycles from 124,648 people. Mean cycle length was 29.3 days, not the 28 everyone quotes. The mean follicular phase was 16.9 days and the mean luteal phase 12.4 days, and cycle length fell by about 0.18 days for every year of age between 25 and 45. Another is the Apple Women’s Health Study, published in 2023 in the same journal, which looked at 165,668 cycles from 12,608 US participants and found that cycle variability was lowest at ages 35 to 39, about 46% higher under 20, 45% higher at 45 to 49, and 200% higher over 50.

That is why age matters to the starting guess. A 22-year-old and a 47-year-old should not get the same first prediction, and when Ferrabelle knows your age, they do not.

How it learns you

From your first logged cycle onward, the engine personalizes to you, and your own data always outweighs the population as it accumulates. Once a week it recalibrates to your own history, right on your phone. Nothing is pooled across users, because the Android app has no server to send your data to.

Two more facts about what it is. The predictions come from a transparent statistical engine. No AI is involved in them. The optional AI assistant only explains and answers questions when you turn it on, and it never sees raw Health Connect values.

The test

Claims are cheap, so I replayed the real production prediction engine against a public research dataset: the Fehring/Marquette menstrual cycle study, 159 women and 1,665 cycles, published by Marquette University. For each woman the engine predicted every period as of several days before it actually arrived, using only the data that existed at that moment. The method is called leave-future-out walk-forward validation. 120 women and 3,282 predictions qualified for scoring. There is no way for the engine to peek at the answer.

Two things make the result honest. First, the engine was never trained or tuned on this dataset. It saw each cycle cold, the same way it sees yours. Second, the evaluation is deterministic and repeatable. The harness lives in the codebase and reruns on demand, so the accuracy page can always be regenerated from scratch.

The results, for the next period:

The calendar line is the one I watch, because counting forward from your usual cycle length is what most people do in their head. On the same women, the same cycles and the same rules, the engine’s mean error is 1.82 days against the calendar method’s 2.08.

Where it is weaker

I would rather you hear this from me than find it out the hard way.

Long and irregular cycles are harder. On the same public dataset, cycles in the regular range predicted at 1.79 days mean error; long cycles (over 35 days) at about 2.8 days. If your cycles are irregular, expect wider, honest confidence windows rather than falsely precise dates.

The public dataset has no wearable temperature data, so the numbers above reflect the engine without its strongest signal. Wearable temperature is where ovulation detection sharpens most, and the wearable-anchored numbers will be published when there is a hormone-confirmed validation set to measure them on. Not before.

Hormonal birth control changes what prediction means. If you are on hormonal birth control, ovulation-based prediction does not apply the same way. The app has a cycle tracking mode for that which turns predictions and phases off and keeps the log for bleeds and how you feel, rather than pretending the math still works.

Check it yourself

You do not have to trust the page. On your own phone, Settings > Check prediction accuracy replays 300 test cases and confirms your phone gives the same answers as the engine measured on the accuracy page. If your phone’s numbers ever disagree with the published ones, something is wrong, and I want to know.

What these numbers are not

Ferrabelle is not a medical device and not a contraceptive, and its predictions must never be used to prevent pregnancy or to diagnose or treat any condition. Prediction error is a population statistic. Your cycles are your own, and any app can be wrong about them. If your cycle changes unexpectedly or worries you, talk to a clinician.

A 1-day median error means at least half the predictions were off by a day or less (56.8% were within 1 day), and the rest were off by more. I think that is a good number, and I also think you should know it is the number.

This is general information, not medical advice. Talk to your doctor about your own health.

Ferrabelle is in a closed test on Google Play right now. If you want to try it, you can join at ferrabelle.com/testers.

Sources