Start: the question
Good advice, or advice nobody follows?
Training plans say to hold back early and finish faster, and the physiology behind it holds up. I had never seen a count of how many ordinary runners actually manage it.
Boston publishes a half-marathon split for every finisher, so the count is easy to make. Each race splits into the time to the halfway mat and the time from that mat to the line. If the second number is smaller, the runner ran a negative split. The analysis is that subtraction, done 31,912 times.
13.1 mi: the answer
Almost nobody speeds up
The rust bars at the left of the chart are the negative splits: 803 runners. Another 3.9% finished between even and 2% slower. Everyone else, about 94% of the field, gave time back, and the most common fade was 12 to 15%. Pick a finish band to see how the shape moves from the front of the field to the back.
All 31,912 finishers: 2.5% ran the second half faster
Hover the chart, or focus it and use the arrow keys, to read a bar.
tables/web_splits.json, written by export_web.py
with the same definitions as analyze.py.
20 mi: who fades
Even pacing goes with being fast
Split the field by finish time and the median slowdown grows at every step, from 4.5% for the 2:00 to 2:30 group to 26.8% for the 5:00 to 6:00 group. Negative splits thin out the same way: 8.2% of the 2:00 to 2:30 group ran one, and 0.5% of the 5:00 to 6:00 group.
| Finish band | 1st half | 2nd half | Fade | |
|---|---|---|---|---|
| 2:00–2:30 | 1:10:14 | 1:13:32 | +4.5% | |
| 2:30–3:00 | 1:23:25 | 1:28:35 | +5.2% | |
| 3:00–3:30 | 1:33:45 | 1:43:03 | +8.6% | |
| 3:30–4:00 | 1:45:07 | 1:59:05 | +12.2% | |
| 4:00–4:30 | 1:56:41 | 2:17:30 | +17.3% | |
| 4:30–5:00 | 2:07:52 | 2:36:24 | +22.0% | |
| 5:00–6:00 | 2:22:20 | 3:01:42 | +26.8% | |
| All 31,912 | 1:47:50 | 2:03:46 | +14.0% |
Fade is the median of each runner's own second-half/first-half ratio, so it is not the ratio of the two median columns. Bands as in figure 3 of analyze.py.
The cause runs both ways. Even pacing probably helps, but fitter runners are also further from their limit at halfway. The data shows the association and can't separate the two.
23 mi: a smaller gap
Women paced slightly more evenly
The median slowdown was 13.4% for women and 14.8% for men. Women were a little less likely to run a negative split (2.3% against 2.7%), so the difference is in how far each group faded. Both gaps are small next to the one between the front and the back of the field.
25 mi: your race
Where your split lands
Enter your halfway split and your finish time from any marathon. The page works out your fade and compares it with the 2014 Boston runners who finished in the same band.
h:mm:ss or h:mm. Finish between 2:00 and 6:00 to compare with a band.
October 2026 upgrade · every Boston year with splits
2014 was one of the worst years to try it
The counts above are one race. The same rule (second half against first, from the halfway mat) was run on every Boston year with public splits: 2014 to 2017 and 2021 to 2024. The share of negative splits ran from 2.4% in 2024 to 10.8% in 2023, and the median fade from 5.8% to 14.0%. 2014 sits at the hard end of both: the second-lowest share and the biggest median fade of the eight. The years go together, too: when the field fades less, more runners manage a negative split.
| Year | Finishers | Negative split | Median fade | ||
|---|---|---|---|---|---|
| 2024 | 25,757 | 2.4% | +13.7% | ||
| 2023 | 26,674 | 10.8% | +5.8% | ||
| 2022 | 24,745 | 6.9% | +8.3% | ||
| 2021 | 15,291 | 4.7% | +11.7% | ||
| 2017 | 26,393 | 3.1% | +12.8% | ||
| 2016 | 26,613 | 2.6% | +12.1% | ||
| 2015 | 26,570 | 7.7% | +7.2% | ||
| 2014 | 31,859 | 2.5% | +14.0% | ||
| 2013 (halted) | 16,096 | 6.0% | +7.7% |
From tables/years_splits.json, written by multiyear.py from the multi-year table that the
companion repo builds. This table drops wheelchair and handcycle entrants by division; the 2014 headline above
(31,912) kept 53 wheelchair finishers, one of them a negative split, so the 2014 row has 31,859 runners and the same
2.5% and 14.0%. 2013 was stopped at about 4:09 on the race clock, so its finishers are the faster part of the field;
it is shown for completeness and left out of the comparison. 2018 and 2019 are missing because the public files for
those years carry no splits.
The companion study now has a live split tool: enter your splits as you pass each mat and it predicts your finish band, your chance of hitting the wall and your pacing type, using a model trained on these years and tested on 2024.
Limitations: what this doesn't show
One race, one day, one course
The headline numbers are one edition of one marathon, and the per-year table shows it was a hard one. Boston's course also works against even pacing. It drops for the first few miles and puts the Newton hills between miles 16 and 21. Some of the 14% belongs to the course, so a split from a flat race will look better against this field than it would at home.
Boston is also a qualifying race, with a faster and more experienced field than a typical big-city marathon. That should push toward more even pacing, so a general-entry field would probably fade more.
The halfway split is the only checkpoint used. A runner who surged and faded within one half looks the same as one who ran it steadily. Weather on the day is not modeled. The spread between years (2.4% to 10.8% negative splits) says the day matters a great deal; how much of that is weather, this data can't say.
26 mi: what to take from it
Read it backwards
The title invites the reading that negative splitting is a myth and not worth trying. I take the opposite lesson: the runners closest to even pace were the fastest ones, and the field's pacing errors run almost entirely in one direction.
Of 31,912 finishers, almost none started too slowly. Whatever pace feels right at the start line, this field says it is too fast.
The finding, as adviceFinish: how it was built
Where the numbers come from
fetch_data.py downloads the results and analyze.py
writes the figures and tables/summary.csv.
export_web.py reads the same data through
analyze.load() and writes the band histograms and percentiles
this page draws from. The raw data is git-ignored and downloaded on demand.
- Data
- Official 2014 Boston Marathon results, from the public llimllib/bostonmarathon mirror
- Sample
- 31,912 finishers with a valid half-marathon split
- Definition
- First half is the half-marathon mat time; second half is finish minus half. Ratio is second over first. Negative split means a ratio below 1; even means 1.00 to 1.02
- More years
- 2015 to 2017 from a Kaggle mirror, 2021 to 2024 from williamsorg4/BostonMarathonTimeModel, assembled by
build_dataset.pyin the companion repo; per-year counts in this repo'smultiyear.py - Stack
- Python, pandas, NumPy, matplotlib; the charts here are plain SVG
- Reproduce
pip install -r requirements.txt, thenpython fetch_data.py && python analyze.py && python export_web.py- Companion study
- Hitting the Wall follows the same field 5 km at a time to find where the fade starts