2026-11-22 · Yimaji
Open most period apps and the home screen says “in 5 days”. That number looks like a conclusion. It is an average, plus an assumption that your cycles are regular.
The trouble is that few people are exactly 28 days every month. 28, 31, 26, 29, 33 — if the app only hands you a date, it has hidden the uncertainty. Yimaji does the opposite: a range, a confidence level, and one sentence saying why that range.
To predict the next period you have to answer three questions:
Only with those three can you give a range. Answer the first one and print a date, and you have simply ignored the other two.
Yimaji takes the most recent 12 completed cycles and keeps only those between 15 and 60 days (outside that range it is usually a missed or mistyped entry).
Then it takes a weighted median: recent cycles count for more, with weights decaying by 0.85 per cycle of age. The last cycle weighs 1, the one before it 0.85, the one before that 0.72, and so on.
Why a median rather than a mean? Because a single outlier drags a mean. Suppose your cycles ran 28, 29, 28, 45 (that trip with too little sleep). The mean is pulled to 32.5; the median stays at 29. A median only cares about the value in the middle; one extreme cannot move it. Weighting then lets your body's recent state speak louder than the state it was in two years ago.
With fewer than two cycles on record, the app can only use the “about how many days” you entered when you started, with a spread of 4 days — the range at that stage is wide, because genuinely nothing is known yet.
Textbooks measure spread with standard deviation, which has the same weakness: one outlier inflates it.
So Yimaji uses the median absolute deviation (MAD): take each cycle's distance from the median, drop the signs, and take the median of those distances. Then multiply by 1.4826 to put it on the same scale as a standard deviation. That constant is not arbitrary — it is what makes MAD match the standard deviation for normally distributed data.
The result: one 45-day cycle does not suddenly make your “spread” 8 days. But if the last six months really have been swinging between 26 and 33, the number honestly grows.
The window runs ±1.28 times the spread, which under a normal approximation covers roughly 80%; two boundaries apply:
Whatever the arithmetic says, the screen says. That is why the range can differ from month to month — it is not jitter, it is your own record talking.
When the spread exceeds 4 days the app calls the cycle irregular, deliberately widens the window, and says so on the home screen. That sentence is not reassurance, it is an explanation: the prediction is not broken; the data simply has no stable centre at the moment.
There are three levels, on plain criteria:
The home screen does not bandy the word “confidence” about either. It translates it: “based on your last 7 cycles, varying by about 2 days”. Cycle count and spread are things you can check; a score is not.
For further-out predictions (the second and third cycles ahead), the window widens with the square root of the distance, up to 21 days; from the fourth cycle on, everything is marked low confidence — the further out a prediction goes, the more it should be read as a hint than a plan.
Ovulation is not observed directly; it is inferred by counting back: 14 days before the predicted start of the next period (the luteal phase length is adjustable between 9 and 17 days). Ovulation gets a range of ±2 days, widening to ±3 when cycles are irregular; the fertile window runs from 5 days before ovulation to 1 day after.
This is the standard calendar method, and precisely for that reason it cannot be used as contraception and cannot confirm ovulation — both statements are written into the interface. For something closer to real evidence you need basal body temperature or ovulation tests; the app only puts those records next to the prediction so you can look at them together.
The temperature reading is retrospective as well: the highest of the previous 6 readings plus 0.05 °C sets a baseline, 3 consecutive days above it count, and ovulation is placed on the day before the rise. It needs at least two cycles and five readings before it says anything; otherwise it draws the chart and draws no conclusion.
It is the same app, but a different goal means the home screen should be saying different things:
The mode can be switched at any time; history is kept.
When you change phones, the record exports to a .ymbk file. Inside is data encrypted with AES-GCM, with the key derived from your passphrase by PBKDF2-HMAC-SHA256 at 600,000 iterations and a 16-byte random salt.
The whole file header — format marker, version, iteration count, salt — is treated as additional authenticated data: change one byte and decryption fails. The app stores no passphrase and offers no recovery, so a wrong passphrase and a damaged file return the same error. That is not laziness; it is what makes “only you know it” true.
Predictions are never written to Apple Health. Only periods you actually recorded as having started go there. Write a prediction into Health and it becomes indistinguishable from a real record; after that, nobody — including you — can tell them apart.
No account, no cloud sync. Records live on the device. No server means nothing to analyse and nothing to leak.
No diagnosis, no contraception advice. A prediction is an estimate, not medical advice. Unusual bleeding, severe pain or fainting: see a doctor.
The only network communication is Apple's in-app purchase. No ads, no third-party analytics.
App Store: id6826955008 · iPhone / iPad, iOS 18 or later; Apple Watch app requires watchOS 11 · Free download, Pro comes with a free 3-month trial, then a one-time purchase
Predictions are built on your own record and are for information only, not a method of contraception.