Where our numbers come from

Between 2026 and 2026 we ran a training-analytics service that synced Garmin Connect data with users' explicit permission. At its peak the database held roughly 81,000 activities and 375,000 daily health summaries, some accounts carrying nine years of history.

That service is being retired. Before the personal data is deleted we computed aggregate statistics from it, and those aggregates are what we publish. We do not retain a personal dataset to mine later, and we do not sell data to anyone.

Our athletes are self-selected. They own a Garmin device, went looking for a third-party analytics tool, and handed over their credentials to use it. That makes them almost certainly fitter, more consistent, and considerably more data-curious than the average watch owner.

So when we report a median, it is the median of that group. It is not a population norm, and we will not present it as one. Sample sizes appear next to every claim we make — if a number came from 39 people, we say 39.

Six accounts are excluded from every study. They were created within a three-day window in May 2026 and carry near-identical multi-year histories under different Garmin accounts — average resting heart rates matching to within 0.03 bpm across roughly 3,700 days, with one pair byte-identical. Whatever produced them, they are not six independent people, so they are not counted as six independent people.

We also drop activities dated before 2010. Those are the FIT-format epoch (31 December 1989) leaking through as a default timestamp — six records in total.

Almost everything we publish is a within-athlete comparison. We measure each person against their own baseline and then report the distribution of those individual results. Pooling everyone into one scatter plot would let fitness differences between athletes swamp the effect we are actually asking about.

Correlations are Spearman rank correlations, which do not assume a straight-line relationship. We report medians and quartiles rather than means, because physiological data is skewed and a couple of outliers should not move a headline.

No cell built from fewer than five people is ever published. When a result falls below that line we mark it suppressed and say so, rather than quietly dropping it.

When we report that something does not predict anything, the obvious objection is that our method is too crude to detect it. So we test the method on a relationship that must exist.

Example: we found Garmin's Training Readiness had almost no relationship with how a run actually went (median correlation 0.056). In the same data, using the same method, readiness tracked the previous day's training load clearly (median −0.272, negative for 96% of athletes) — which it should, because load is an input to the score. The method works. The prediction is what is missing.

Any null result we publish comes with its control. If the control fails, the study does not run.

• Absolute accuracy against lab equipment. Nobody in our cohort has a metabolic-cart VO2 max on file. We can describe how a metric behaves, not whether its number is physiologically true.
• Causation. Everything here is observational. People train, live and sleep as they please; we watch.
• Race-day outcomes at scale. Only a handful of efforts in our data are tagged races with a prediction recorded beforehand. Where we use them, we call it a case series and give the count.
• Anything about brands we never synced. We hold deep Garmin data and no equivalent Coros, Polar, Suunto, Whoop or Oura data. When we write about those, we are reading their documentation and published research, and we say so.

We publish corrections in the article itself, not in a footnote nobody reads. Our own race-predictor article used to repeat the correction factors that circulate on running forums. When our data contradicted them, we said so on the page and replaced them.

If you find something wrong here, tell us and we will fix it in public: hello@gneta.app.

The analysis scripts and their aggregate outputs live in our public repository. Every number in a study article traces back to a JSON file produced by a script you can read. We would rather be checked than trusted.