Survival analysis on every Olympic athletics world record
Fitted only on what was known at the end of 2010, a survival model of every Olympic athletics world record scored a Brier of 0.235 on what fell from 2011 to 2025, against 0.329 for the historical base rate (lower is better), though it ran overconfident. Applied today, it puts Duplantis's 6.31 pole vault first in line to fall before the LA 2028 Games close (48%) and Florence Griffith Joyner's 10.49 last (5%). The records come from World Athletics' own lists for all 46 events.
From tables/backtest_summary.csv, tables/reliability_rolling.csv and tables/watchlist.csv.
Record reigns in 46 events, from World Athletics' ratified lists, checked against its records page on 28 October 2026.
Brier score on the 46 records standing at the end of 2010, scored on 2011 to 2025. The base rate scores 0.329; lower is better.
In the rolling backtest, records the model gave 40 to 60% fell 24% of the time. It overestimates how fast records go.
Chance Duplantis's 6.31 falls before 30 July 2028, the highest of any record. 90% interval 19 to 57%.
The watchlist
Each row is a world record standing on 28 October 2026. The two probabilities are the model's chance that the record falls before the LA 2028 Olympics close on 30 July 2028 (21 months away) and within five years. The bar shows the LA 2028 figure with its 90% interval, which comes from refitting the model on 200 resamples of the 46 events. Click a column heading to sort.
From tables/watchlist.csv. Age is how long the record has stood. Two
events have a faster mark pending ratification, set in Zürich on 27 August 2026; if it is
ratified, that record has already fallen.
The July 2026 version of this page said Kelvin Kiptum's 2:00:35 had a 60% chance of falling within five years. It had already fallen: Sabastian Sawe ran 1:59:30 in London on 26 April 2026, and World Athletics ratified it in July. The old pipeline had a hard-coded reference list and never looked at the current records. This version checks all 46 against World Athletics' records page, and its run fails if any differs.
The backtest
A forecast has to be tested on time it has not seen. I cut the data at 31 December 2010: records set later were removed, and reigns still running were censored at that date. The model was fitted on that, then asked for each of the 46 records standing at the cut: what is the chance this falls in the next 15 years? Then I checked what happened. 26 of the 46 fell.
The base rate is the share of standing records that fell within 15 years at every earlier year end whose outcome was known by 2010 (85%). The pooled rate is one constant hazard for every record. The Kaplan-Meier method is the July 2026 approach. A Brier score is the mean squared gap between the probability and what happened, so lower is better; an AUC of 0.5 is a coin toss.
| Forecast | 2010 cut: Brier | AUC | 2000 cut: Brier | Rolling: Brier | AUC |
|---|
The model's Brier gain over the base rate at the 2010 cut is 0.093, with a 90% interval of 0.030 to 0.157 when the 46 events are resampled, so the gain is unlikely to be luck. Its edge over the Kaplan-Meier method is small (0.235 against 0.250) and well inside the noise for 46 records. At the 2000 cut every method did badly: the model and the base rate both expected most records to fall, and the 2000s turned out to be a slow decade for records.
The rolling test refits at every year end from 1985 to 2020 and predicts the next five years, which gives 1,535 predictions, 508 of them records that fell. The chart puts the predicted chance against the share that fell, in bins. A forecast on the dotted line is calibrated. Below the line means records fell less often than the model said.
tables/reliability_rolling.csv and
tables/reliability_2010.csv. Red is the frailty model; grey is the
Kaplan-Meier method on the same cutoffs. Focus a point, or hover it, to read the bin.Both lines sit below the diagonal from about 40% up. When the model said a coin flip, the record fell about a quarter of the time. The reason is the one the model was built to handle and only partly does: records have been getting harder to break, and a model trained on the past keeps expecting the faster pace of the past. Read the watchlist with that in mind. Its ranking is more useful than its levels.
Survival by event group
Across all 1,251 reigns the Kaplan-Meier median is 1.1 years. Groups differ a lot. Half of all throws and jumps records were gone inside a year, many of them broken in the same season by the same athlete. Sprint records have a median of 3.0 years. The dashed line is the model's curve for a record set today in a group with no other recent records.
web/data.json, which analyze.py writes. Time on a square-root
scale. Black dots are the records in the group standing today.The model
The main model is a Weibull proportional-hazards model with a shared gamma frailty per event, fitted by maximum likelihood in SciPy. The Weibull part says how risk changes with a record's age: the shape is 0.69, so the yearly risk drops as a record gets older. The frailty is a multiplier for each event, estimated from that event's own history and pulled toward the average when the history is short. Every reign is split at each 1 January so two covariates can change while a record stands: the calendar period, and how many records the event has seen in the last three years.
| Covariate | Hazard ratio | 95% interval | Reading |
|---|
A hazard ratio above 1 means records fall sooner. Three effects are strong. Since 2000, records at risk fall at about half the rate of 1980 to 1999 (0.53). Each percentage point by which a record beat its predecessor cuts the hazard by a fifth (0.79). And an event that has just seen records keeps seeing them: a record with three others set in its event in the last three years carries about a quarter more risk than one that is the only recent record (1.28 per unit of log count). Throws and jumps fall about twice as fast as sprints.
Kaplan-Meier and a Cox model are the baseline. The Cox model's proportional-hazards test fails for the era covariates (p < 0.001 for records set in 1980 to 1999), which is why the main model uses the calendar period a record is at risk in rather than the year it was set. By AIC the final model beats a plain Weibull by 66 points and log-normal and log-logistic versions by more than 150.
| Model | Parameters | Log-likelihood | Δ AIC |
|---|
One caveat about the order of work. I specified the first frailty model before running any
backtest. It used the era a record was set in, and it scored worse at the 2010 cut (0.273).
I then added the calendar period and the recent-record count. Both choices were made after
seeing those scores, so the 2010 and rolling numbers for the final model are not a clean
holdout of the model choice. All three versions are in tables/backtest_summary.csv.
The data
The progressions come from the "Progression of Official World Records" section of World Athletics' Oregon22 statistics handbook, which lists every ratified record to 31 May 2022, equalling marks included. The 30 records ratified since then come from World Athletics' progression pages, 11 of them Duplantis vaults. Wikipedia is only a cross-check: 1,456 of the 1,464 listed records appear there with the same mark and year.
Hand and automatic timing are two chains. World Athletics accepted only automatic times for events up to 400 m from 1 January 1977, so the hand-timed record standing that day is censored then rather than counted as broken, and automatic records older than 1975 enter the risk set late. Javelin chains restart with the new implements of 1986 and 1999. Road records count only from 2004, when World Athletics began ratifying them. Every rule, and every record dropped, is in docs/data_decisions.md.
A validation script flags every reign over 20 years and every unusually large jump (95 flags) and looks each one up. The flag that started this rebuild was the women's 800 m. The July version had three reigns, one of them a 1:54.0 from 1925 that stood 55 years. The handbook lists 29 women's 800 m records from 1922, the fastest of the 1920s being 2:16.8. The 1925 row was a misread.
Limitations
It knows when records fell and how busy each event has been. It does not know who is close now. Duplantis has set the record 15 times since February 2020, usually by a centimeter, so 48% is probably too low for the pole vault. The women's pole vault (28%) shows the opposite problem: the event's busy 1990s and 2000s still lift its frailty, while nobody has come near 5.06 for years.
The backtest is small. 46 records at the 2010 cut, and the rolling test reuses the same records many times, so its 1,535 predictions are far from independent. The calibration problem is real; the exact size of the model's edge is not well pinned down.
Holder age was on the list of covariates to test. The handbook has no dates of birth, and World Athletics' progression pages, which do, only load in a browser; I did not collect them for all 1,464 records, so age is not in the model. The watchlist does not adjust for doping-era records either, which the model treats like any other.
October 2026 upgrade
The July 2026 analysis used 248 reigns from 9 events that passed a completeness gate on a Wikipedia scrape. This version covers all 46 Olympic events from World Athletics' own lists, adds a parametric model with per-event frailty, tests it with a time-split backtest, and forecasts to LA 2028.
Superseded on 28 October 2026. The July headline numbers were a median reign of 1.5 years, a Cox hazard ratio of 0.52 for records set in 1980 or later, 0.65 per point of improvement, and five-year fall chances from 60% (the 2023 marathon) to 0% (the 1983 women's 800 m).
On the full data the median reign is 1.1 years. The 1980 split in median reign is gone (1.09 years before, 1.03 after). The slowdown shows up as a lower hazard in the period since 2000, not in the year a record was set. The margin effect holds at 0.79 per point. The 0% for the women's 800 m came from a data error and a flat stretch of the Kaplan-Meier curve; the model now puts that record at 9% before LA 2028.
How it was built
web/data.json