How long does a track & field world record stand — and which of today's records is the survival model betting against?
🌐 Overview: https://lyhjeremy.github.io/world-record-half-lives/
Athletics world records don't fall randomly. Some are broken within months; others — Kratochvílová's 800 m from 1983 — have outlasted the careers of everyone who has chased them. This project treats each record as a survival-time problem: every record "lives" until the next one breaks it, and the records still standing today are right-censored. From ~250 record reigns across nine events I fit a Kaplan-Meier survival curve, a Cox proportional-hazards model of what makes a record fall sooner, and turn the model on the current records to estimate each one's probability of falling in the next five years.
- A record's hazard is front-loaded. The Kaplan-Meier half-life of a record is only about 1.5 years — most records are beaten quickly. But that median hides a heavy tail: the records still standing have all survived the dangerous early window.
- Records last far longer than they used to. Split by era, a record set in 1980 or later survives markedly longer than a pre-1980 record (see the curves), and the Cox model agrees: being a modern record roughly halves the hazard (HR ≈ 0.52).
- Big leaps last. The larger the improvement that set a record, the longer it survives (HR ≈ 0.65 per unit) — Beamon-style jumps buy durability.
- Vulnerability today. Given how long each current record has already stood, the model's five-year fall probabilities run from ~60% (the 2023 marathon record, in the super-shoe era) down to ~0% (the 42-year-old women's 800 m).
Progressions come from Wikipedia's "… world record progression" pages
(pandas.read_html). These pages mix the chronological progression with all-time
top-lists and nav-boxes, so a naïve scrape silently truncates or contaminates an
event's history. The fetcher therefore keeps an event only if its reconstructed
running-best equals the current ratified world record (and its history reaches
back far enough) — a hard completeness check. Nine events pass cleanly: 100 m,
800 m, 1500 m, 5000 m, 10,000 m, marathon and shot put (men), plus the women's
800 m and high jump. Events whose modern records live in tables the scraper can't
reliably isolate are dropped rather than analysed half-complete.
pip install -r requirements.txt
python fetch_data.py # scrape + verify progressions -> data/raw/ (git-ignored)
python analyze.py # survival models + curve fits -> figures/ + tables/| Path | What it is |
|---|---|
fetch_data.py |
Scrape Wikipedia progressions; keep only events verified complete vs the real WR |
analyze.py |
Reconstruct reigns → Kaplan-Meier + Cox + asymptotic fits → figures & tables |
figures/ |
Survival curve, progression fits, vulnerability ranking |
tables/ |
record_reigns.csv, cox_summary.csv, current_record_vulnerability.csv |
Reigns are reconstructed from ratified progressions, so the model sees when records fell but not the live "pressure" on them (how many athletes are within 1% right now) — a natural next feature. Estimated event "limits" are omitted from the headline because for still-improving events the asymptote isn't statistically identifiable; the vulnerability score leans on the survival model instead, which needs no such extrapolation.
MIT © 2026 Jeremy Lee · record data from Wikipedia