Research
What 371,000 days of training data actually show
We ran a Garmin analytics service for a year. Before shutting it down and deleting the personal data, we computed aggregate statistics from 81,000 activities and 371,000 daily health summaries — some accounts holding nine years of history.
These are the findings. Read the methodology first if you plan to cite any of it — particularly the part about who our athletes are and what this data cannot tell you.
Garmin race predictor: what 22,422 daily predictions show
S1All 85 athletes got a steeper slowdown curve than the textbook formula predicts.
85 athletes · 22,422 prediction-days
Read the study →Does Training Readiness predict how a run goes?
S2Median within-athlete correlation with running efficiency: 0.056. Effectively nothing.
39 athletes · 5,136 runs
Read the findings →How much does Garmin move your VO2 max?
S3It changes about every sixth day — by 0.2 points. A year of training moves it 2.0.
42 athletes · 180+ days each
Read the findings →Body Battery, HRV and sleep score: three metrics or one?
S4They agree for 98% of athletes. Largely one signal wearing three interfaces.
82–93 athletes · 30+ paired days each
Read the findings →Does last night’s sleep score predict today’s run?
S5Median correlation 0.04 across 7,492 runs. The control worked; the signal was not there.
54 athletes · 7,492 runs
Read the findings →Using these numbers
Cite them freely with a link back. If you are writing something that depends on one of these findings, read the caveats in the study first — every one of them has real limits, and we would rather be quoted accurately than widely.
Questions, corrections, or a dataset you want us to look at: hello@gneta.app.