We built a better prediction model. You won’t get to see it — and that is exactly the news.

Something happened in the mypacing lab this week that we want to tell you about — not despite the fact that it looks unspectacular, but because it does.

For months, mypacing has been working out every night, from the community’s pseudonymized data, what precedes a crash. The model behind it deliberately uses few, well-founded factors: exertion, recovery, sleep, accumulated load. On the technology page you can see at any time how well it separates — and also that, honestly, this figure is not yet good enough to let the learned weighting loose on new people. It then switches itself off. That is how the system is built.

Now we have a new candidate. It does not look at a handful of factors, but at all 30 daily values that a wearable and your own entries provide — from resting heart rate through respiratory rate to humidity — and additionally at the temporal depth of the past few days. In the strict historical test, in which the model is retrained for every single person without that very person and then tested on them alone, it separates markedly better than anything we had before: for the first time, the figure lies with its entire confidence range above the coin flip.

Reason to celebrate? Cautiously. Because when we turned the same model loose retrospectively on the past two weeks of everyday life, the lead was almost gone — the confidence range included chance again. Historical strength and everyday usefulness are two different things, and anyone who publishes only the first figure is telling half the story.

Hence: shadow mode

From now on, the new candidate runs along every night — on real, pseudonymized data, under real conditions. But nobody gets to see its results. No display, no warning, no push notification. It computes in the shadows, and everything it predicts is merely logged.

The decisive trick: every prediction is frozen on the morning of the day it is about — before the day has decided itself, calculated by a model that cannot know the outcome. Weeks later we compare these frozen predictions with what actually happened: with the crashes that people marked themselves. That is the only test that counts. Everything else — including our own fine historical figure — is preparation.

And because a model that never learns anything new goes stale quickly, the candidate retrains itself automatically every night on all the data shared up to that point. But not blindly either: a new version is only adopted if it passes five checks — among other things, it must never be worse than the one currently running, and its confidence range must lie above chance. If one check fails, the old version simply stays in place and the event is logged. A discarded run is not a failure here, but the safety net at work.

Why we are telling you this even though nothing has “come of it” yet

Because this is exactly where trust is built — or squandered. Hardly a week goes by without a headline in which some AI “predicts” something. In an illness where a single episode of overexertion can cost weeks, a premature warning feature would not be bold but irresponsible: a false warning takes a good day away from you. A false all-clear possibly a month.

That is why the order is reversed at mypacing: first the proof under everyday conditions, then a deliberate decision, then — perhaps — a display. And if the prospective test comes out negative, we will write that on the technology page too, just as a discarded “resilience score” is already documented there today, along with its figures.

What you can do to get an answer sooner

The test bench lives on two things only you can contribute: marked crashes and regular data. Every marked crash is a touchstone against which the frozen predictions have to prove themselves. Every synchronized day closes a gap — in our first analysis, a data row was simply missing on a good third of the days. And anyone who uses the symptom tracking feeds exactly the factor that research regards as the strongest single early warning sign and that still rests on too thin a basis with us.

You do not have to learn anything new or change any settings for this. Just carry on — at your own pace. You can find the current state of the shadow test at any time in the section “What we are currently testing in the shadows” on the technology page.

This post is for general information and does not replace an individual medical assessment. mypacing.app is not a medical device.

Published on 21 August 2026. The tests described are ongoing; their current status is on the technology page, not here — so that this post does not quietly go stale.