Study S1 · August 2026
What 22,422 daily race predictions reveal
Key finding
Garmin's own numbers say your 5K does not predict your marathon the way the textbook formula claims. All 85 athletes we tracked showed a steeper slowdown curve than Riegel's 1.06 — median 1.13.
Gneta, August 2026 · 85 athletes · 22,422 prediction-days
85
athletes
22,422
prediction-days
2016–2026
data window
What we did
Our analytics service recorded Garmin's four race predictions every morning for consenting users. That produced 22,422 daily snapshots across 85 athletes, a median of 38 tracked days each, with some series running over a year.
We asked three things of that data: what slowdown curve Garmin's own four numbers imply, how those predictions compare with what athletes actually ran, and how much the predictions move over time.
What we found
Finding 1
Garmin builds in more slowdown than the textbook formula
The Riegel formula — the standard used by nearly every race calculator online — assumes your pace decays with an exponent of 1.06 as distance grows. We derived the exponent implied by Garmin's own four predictions for each athlete.
Median implied exponent from 5K to marathon: 1.131 (IQR 1.114–1.143). All 85 athletes came out above 1.06. The half-marathon-to-marathon step was steepest at a median of 1.164.
For the median athlete in our cohort, that is the difference between a textbook projection of about 3:47 and Garmin's own 4:21 — roughly 34 minutes of extra fade priced in.
Finding 2
Against training runs, predictions look wildly optimistic
For each athlete we took their fastest genuine effort at each distance during their tracking window and compared it with the prediction standing that day.
At 5K the best training effort was a median 22.7% slower than predicted (48 athletes). At 10K, 20.4% slower (39 athletes). Only about one athlete in four ever came within 10%. The marathon cell is suppressed — only one athlete qualified, below our five-athlete floor.
Finding 3
In actual races, it was nearly right
Eleven efforts in our data were flagged as races and had a prediction recorded the day before. Median error: +0.6%. The middle half fell between 4.3% faster than predicted and 1.6% slower.
Eleven efforts from six athletes is a case series, not a study. We report it because the direction is striking and because withholding it would be the dishonest choice — not because it settles anything.
Finding 4
The number moves rarely, then moves a lot
Predictions changed on roughly one day in ten (8.2% of consecutive days at 5K, 10.7% at marathon). But over a median 38-day observation window, the median athlete's marathon prediction still swung 11 minutes, and the top decile swung more than 35 minutes.
Limits you should know before citing this
- The cohort is self-selected. Garmin owners who sought out a third-party analytics tool. Fitter and more data-curious than average.
- The race sample is tiny. Eleven efforts from six athletes. Finding 3 is a case series and nothing more.
- “Best effort” favours the short end. We take the fastest qualifying run in each distance band, which slightly flatters the actual times — meaning Finding 2 understates the gap rather than inflating it.
- Six accounts excluded. Near-identical multi-year histories under different Garmin accounts. See the methodology.
- Observational. No causal claim is made anywhere in this study.
The data
The aggregate output and the script that produced it are public. Every number on this page traces to that file.
View the aggregate data (JSON) →