How we know

Relative pace compares each recorded section with the same finish's whole-race average pace. Profiles are clipped at the 0.1st and 99.9th percentiles, weighted by section length and grouped by k-means into six landmarks, then named by rules: least late slowing is the Metronome, the next is the Gentle fader, and within the two stronger pairs the one that slows earlier is the Early drifter or Fast-start fader. Left out of these shape statistics: 15 editions whose first 5 km appears to include start delay. Repeat pairs use screened identity candidates with one finish in each of two years at most three years apart; they are not verified people.

• In the six-runner race, each runner follows one type’s average profile scaled to the same finish time, so any time a profile gains early is given back later.
• “Extra time after 20 km” is the median of each finish’s time after 20 km.
• Only runners who raced again appear in the repeat pairs.
• Gender is as recorded: “Other / not recorded” finishes count in totals but not in the gender comparison. The gender shares are standardized to the same mix of 30-minute finish bands, so they hold finish time constant but not course, age or conditions.
• Course shares weight each race edition equally and describe each course’s finishes, not a ranking of courses.
• Type definitions and the F, E and S letters are under Pacing types.

Read the storyarchetypes.json

Counts use eligible elapsed finish times in one-minute bins. The expected count in the minute before a mark comes from a Poisson log-cubic fitted to the ±25 minutes around it, leaving out five minutes before and ten after the mark and short windows around other 10- and 15-minute marks. Ratio intervals resample whole race editions (200 draws). The cliff index compares the step from the minute before a mark to the minute after with the neighbouring minute-to-minute steps. The 40 km view projects each finish from its 40 km time at its own 35–40 km pace; its comparison uses finishes projected away from round marks with the same decile of 35–40 km slowing. None of these are goals, predictions or causal estimates.

• Quarter-hour marks such as 2:45 and 3:45 sit inside neighbouring marks’ windows, so their smooth curves are less reliable.
• The 40 km comparison is a reference, not a forecast or a measure of intent.
• Goals, pacers, pace bands and watches are not recorded. Bunching is consistent with round-number targets but cannot prove why anyone sped up.
• The smooth curve and the late-race comparison are fitted from the data, not outcomes any runner would have had.
• Elapsed times are as published by the timing sources; wave and start offsets are not recorded.
• Across hours and half-hours from 2:30 to 6:00, about 110,000 finishes sit in the five minutes before a mark beyond the curve. That total is sensitive to the fitting window.

Read the storyfinish-times.json

Each replay is a stratified sample of up to 1,500 eligible finishes from one race edition, chosen at evenly spaced finish-time ranks. Everyone starts together at elapsed time zero because wave and start offsets are not recorded, and each runner moves at a constant speed within each recorded section. Colours compare the current section with that runner's own 5–20 km pace.

• The replayed fields hold 23,220 to 58,339 eligible finishes. Real fields start over many minutes.
• Changes inside a 5 km section cannot be seen.
• Only eligible finishes are drawn: runners who did not finish, or whose records were incomplete, are absent.
• Route maps are the current supplied course files; historical routes may differ.
• In the snapshots, finished runners count at the finish line. “Ahead” and “moves past” are clock ranks among eligible finishes, not passes on the road.
• The even-pace ghost’s typical 20 km time is the median share of finish time reached at 20 km by real finishes within 2.5 minutes of its target, not anyone’s plan.
• A clock pack is the finishes that passed 10 km within the same 30 seconds of race time. They share elapsed times, not necessarily the same stretch of road.
• The stretch compares how far the slowest tenth trails the fastest tenth from 20 to 40 km with the first 20 km (1.00× means no change). It uses the 199 race editions with at least 1,000 eligible finishes and a usable first 5 km, leaving out the editions whose first 5 km looks inflated by start delays. These are percentiles at each block, not the same runners.
• The field-spread fan pools every eligible finish from every race.

Read the storyreplay.json

Within each edition every eligible finish is ranked by elapsed time at each checkpoint, ties sharing the average rank. Points gained are 100 × (rank at 30 km − rank at the finish) / (field − 1). Coin-flip rates sample up to 12 partners per finish among finishes behind it within 15 minutes at the checkpoint, weighted so every such pair counts equally, and ask how often the trailing finish crossed the line first. Section reshuffles sample random same-edition pairs. Editions with fewer than 100 eligible finishes and the editions whose first 5 km appears to include start delay are left out.

• 3,970,364 finishes in 200 editions are ranked. Pair rates come from sampled pairs with a fixed seed.
• Section reshuffles compare per kilometer, so the short final section compares fairly. A swap means the pair’s clock order at the end of the section differs from the start.
• Clock ranks compare elapsed times within an edition. Because of wave starts, a clock pass is not necessarily an overtake on the road.
• Only eligible finishes are ranked: runners who did not finish or whose records were incomplete are absent.
• Slowing and places gained both use the late kilometers, so part of their link is built into the measures. It describes; it does not predict.
• The median runner after 30 km is the finishes ranked between the 49th and 51st percentile at 30 km. Its passes compare elapsed times at the 30 km and finish mats, so swaps between mats are invisible. Each median is taken separately, so the rows do not divide into each other.
• The women-and-men comparison uses the 200 editions with at least 100 recorded women and men, and measures mean percentile points gained on the mixed-gender clock. On a shared clock one group’s gains are the other’s losses. Gender is as published, and the comparison says nothing about why.

Read the storypositions.json

Each section is compared with the runner's own 5–20 km pace, the baseline of the sustained-slowdown definition: at least 25% slower for at least 5 km after 20 km. The kick compares the final 2.195 km with 35–40 km. Same-race comparisons use strata of one edition and one 5 s/km band of 5–20 km pace with at least 100 finishes, and leave each finish out of its own stratum's mean. Left out: 15 start-offset editions and 2 editions whose timing grid looks shifted.

• A faster final section means faster than the runner’s own 35–40 km, not faster than the early pace.
• The final section is only 2.195 km, so finish-mat placement and course shape matter more there than in a 5 km section: 50 m shifts its pace by about 2%.
• Recorded 5 km sections cannot show walking breaks, stops or recoveries inside a section, and runners who did not finish are absent.
• 17 editions are left out of the section statistics: Valencia 2021 (median final 2.195 km 15.4% faster than 35–40 km), Warsaw 2022 (median final 2.195 km 23.4% faster than 35–40 km) and the 15 start-offset editions. These are inferences from the timing data.
• Recovery means a later section back within 10% of the runner’s 5–20 km pace. The recovery shares for recorded women and men are not matched on pace, age or course.
• The 5–10, 10–15 and 15–20 km sections sit inside the baseline itself, so their narrow spread is partly by construction.
• Pace bands use the 5–20 km pace expressed as a marathon time, set before the second half is run.
• The warning light is a conditional frequency and partly mechanical: a section already 20% slower needs little more to reach 25%.
• Same-race comparisons hold the race and the early pace fixed, not fitness, goals, experience or conditions. The 5–20 km pace is measured after the opening, so it can already reflect it.
• In the twin runners, part of the gap is built into the definition, since a section 25% slower adds minutes by itself. Each runner is drawn with a straight line to its median finish, and the within-race gap compares strata of one race, recorded gender and 5 s/km band.
• Repeat pairs link screened candidate profiles, not verified people, one to three years apart, compared with what was typical in the later race at the same 5–20 km pace. Only runners who raced again appear.
• Gender is as recorded, and the groups differ in age, course and field.

Read the storykick.json

Women and men are compared within the same race edition and the same whole finish minute between 2:30 and 6:00; each edition-minute cell contributes the smaller of its two group sizes, so neither group dominates. The 20–40 km block change compares the 20–40 km block with the 0–20 km block, not measured half-marathon splits. Exact ages come only from editions whose age field holds at least 30 distinct values with fewer than 40% multiples of five; sources that record age-group floors are left out. Intervals resample whole editions.

• Matching on finish time conditions on the result: a man in a given band is more often someone who started faster and slowed into it, so the same finish does not mean the same fitness.
• Gender is as recorded by the source, and “Other / not recorded” finishes are left out of two-group comparisons.
• In the ghost race each runner follows the matched average for one group, so it is not a real pair, and “ahead” means on the clock, not on the road.
• Exact ages cover 957,255 finishes from Auckland, Austin, New York, Paris and Washington.
• The age ladder compares each group’s sustained-slowdown share with the same race’s same-minute field (both genders), in percentage points. Age comparisons are cross-sectional: different people at different ages, not the same runners getting older.
• 8 editions with incomplete gender or result coverage are left out of the composition charts.

Read the storydemographics.json

Relative pace is each section's pace against the finish's own whole-race average pace. A course curve is the mean of its editions' median curves; the typical curve is the median across courses. The fade multiplier fits each edition curve as a multiple of the typical curve (distance-weighted), and the signature is what remains. Weather is the supplied modelled hour at the scheduled start; within-course slopes compare editions of the same course with each edition weighted equally, and intervals resample whole courses. Editions need at least 100 finishes. Left out of the shape statistics: 15 editions whose first 5 km looks inflated.

• Late fade fits each edition curve as a distance-weighted multiple of the typical curve; 1.00× is typical, and the signature is what remains.
• Name that course matches each edition’s fade-removed signature to the nearest course average, built without that edition. That variant was chosen among four, all reported in the data file: fade removed, weighted distance 59%; fade removed, correlation 56%; raw deviation, weighted distance 45%; raw deviation, correlation 49%.
• Course means the supplied city label, not a verified historical route. Route maps and elevation are the current supplied course files, and the elevation model misses bridge decks; in New York it reads them as water level.
• Relative pace mixes terrain, field ability, weather and race behavior.
• The same-early-pace comparison weights each edition equally and draws only courses with at least 3 editions and 1,000 finishes in the band. Matching on early pace balances speed, not ability, goals or weather.
• The editions whose first 5 km looks inflated (Madrid, New York, Prague and Warsaw) are left out of every shape statistic.
• Field size counts eligible finishes with all nine splits, not official finishers, and missing years include coverage gaps as well as cancellations.
• The before-and-after comparison weights each edition equally within 2015–2019 and 2022–2026 and uses the 14 courses whose editions in each period average at least 95% of raw records eligible. Only the editions the source policy keeps count, so a course can rest on one or two race days in a period. Course mix, entry rules and eligibility shares changed too, so it describes the fields, not why they changed.
• Recorded-gender gaps and changing eligibility shares affect some courses and years.
• The course pages use the same fingerprint and race-morning charts.

Read the storycourses.json

Relative pace is each section's pace against the finish's own whole-race average pace. A course curve is the mean of its editions' median curves; the typical curve is the median across courses. The fade multiplier fits each edition curve as a multiple of the typical curve (distance-weighted), and the signature is what remains. Weather is the supplied modelled hour at the scheduled start; within-course slopes compare editions of the same course with each edition weighted equally, and intervals resample whole courses. Editions need at least 100 finishes. Left out of the shape statistics: 15 editions whose first 5 km looks inflated.

• Weather is never personal or wave exposure, and other things also changed between years.
• Same-course pairs need starts at least 5 °C apart, and pairs share editions, so they are not independent tests.
• Within-course slopes are simple descriptive slopes. The prespecified weather screen remains the adjusted analysis.
• In the untangled view each course’s dots are moved so its own averages sit at the center; the pooled view mixes field, terrain and climate.
• The heat signature is an average within-course association, drawn as a straight line from the typical curve at the average start. Its finish and slowdown readouts use a separate curved fit, compared with a 10 °C start on the same course. Hottest and coolest mornings use courses with three or more editions.
• Race-morning temperatures are modeled for the hour near the scheduled start at one point in the city (see Weather and elevation sources).

Read the storycourses.json

Select finishes in the displayed 15-minute finish-time band, centered on the nearest 15-minute target preset. Show each section’s median pace and middle 50% of observed paces. These are achieved times, not declared goals, and the profile is not an optimal pacing plan.

Open the analysis

Use runners with a recorded best in the two strictly earlier calendar years. Classify the first 10 km as more than 2% faster than that benchmark’s marathon pace, within 2%, or more than 2% slower. For each group, count finishes strictly below the selected time. Groups are observational and pool available editions; a prior-time filter narrows ability but does not eliminate confounding.

Compare finishes in the five minutes strictly below the selected target with finishes from the target up to, but excluding, five minutes above it. Subtract the target’s even-pace time budget from each mean section duration. Means reconcile with mean finish-time differences; selection on the outcome makes this descriptive.

Open the analysis

Compare median percentage pace change from 0–20 to 20–40 km across exact-age bands. When previous performance is supplied, compare the same 15-minute prior-time band. Otherwise compare the same displayed achieved-time band. Gender and course are held to the displayed selection. This is a comparison of different people, not an individual aging trajectory.

Open the analysis

Align the supplied unique city-level course segments with the pacing profile. Historical route validity is unknown. Net elevation change can hide both climbing and descending; bridge decks, tunnels, smoothing and route changes can affect the profile. This does not estimate a historical hill penalty or grade-adjusted effort.

Open the analysis

Identify a net-downhill opening from the first two supplied 5 km terrain segments. Among the earlier-benchmark opening groups, compare median pace change from the 5–20 km baseline to the final 12.195 km. The supplied route is a proxy, and these associations do not establish that an early descent caused later slowing. Where no downhill opening is documented, show the general opening comparison and say so.

Match course, available age/gender groups, checkpoint and a two-minute elapsed-time band at 20, 30 or 35 km. Optional recent-5-km pace is classified relative to elapsed average pace using ±2%. The previous-marathon filter is not used because this comparison conditions on current-race progress. Show historical finish percentiles and the fraction below the visitor’s threshold, not a calibrated personal probability. Only complete eligible finishers are represented.

Open the analysis

For the same displayed age, gender and prior-time cohort across courses, show the 10th, 50th and 90th percentiles of finish-time change relative to the recent recorded best. Require at least 100 observations and three editions per course. Use identical cohort criteria across the displayed courses. These different runners and editions do not isolate a course effect or provide equivalent-time predictions.

Use the supplied modeled start-hour temperature and runners with earlier benchmarks. Compute edition-level median finish-time change, requiring 20 finishes per edition/group. Average edition medians equally within temperature bands, requiring three editions and 100 finishes. Weather exposure is a start-hour proxy, not personal exposure; humidity, sun, wind and route changes are not isolated.

Evaluate the selected time against the historical finish-time distribution in the displayed cohort. Optional previous performance selects a 15-minute band of recent recorded bests. Each whole-minute threshold from 1:30 through 12:00 is calculated exactly using a strict less-than comparison. Without previous performance this describes the selected field, not individual readiness.

Open the analysis

Use consecutive linked appearances in different years, at most three years apart, on the same city course, requiring one eligible recorded race in both endpoint years. Apply the profile to the later appearance and compare both observed section profiles, each normalized to its own finish. Report paired mean finish-time change. Route changes, fitness and selection remain possible explanations; first recorded is not first ever.

Select finishes in the displayed achieved-time band that beat the runner’s fastest recorded finish from strictly earlier years. Compare durations in the opening 10 km, middle 20 km and final 12.195 km with that earlier result. Mean block gains sum to mean finish improvement. This is an earlier recorded best, not a lifetime personal best or a prescription for a future improvement.

Open the analysis

Pace-profile answers describe achieved finishes in the displayed band. The visitor’s time is a comparison threshold, and historical intentions are unknown.

Opening pace is compared with each runner’s earlier recorded best, not a declared goal, and each opening group has its own denominator.

In the downhill question, a positive late pace change means slower over the final 12.195 km than the 5–20 km baseline. This does not establish a downhill effect.

Nearby-finish section charts compare finishes in the five minutes below the target with those from the target to under five minutes above it. Higher values mean more time than the target’s even-pace section budget.

Age answers compare pace retention: positive means the second 20 km was slower than the first, and the final 2.195 km is separate. Each age group shows its own sample, and the comparison does not follow individuals as they age.

In course answers, below zero means faster than the earlier benchmark, and the 10th–90th percentile range describes outcome variation. Course, field selection, fitness changes and conditions stay mixed together.

Weather answers compare finishes with earlier recorded bests, not the visitor’s target, and every edition has equal weight. The weather bands need at least three editions and 100 finishes each, and they do not isolate a temperature penalty.

Without an earlier time, the target answer is a field comparison; with one, its runners had earlier recorded bests in the displayed band. Either way the share is descriptive, not a calibrated prediction.

Returning-runner answers normalize both appearances to their own marathon average. These paired changes can reflect fitness, conditions and route changes as well as familiarity, and they are not selected by the visitor’s target.

Improvement answers split the mean total improvement into three blocks that add up to it.

Terrain answers show the available city route profile beside pooled historical finishes. They do not establish the terrain used in each edition or explain an individual slowdown.

Hills and pacing compares each 5 km section with the same section on every course combined, near the same finish time: each section’s median pace against its own course’s distance-weighted average, minus the same for every course, in percentage points. Late slowing happens on every course, so a section’s pace against its own course alone would credit a hill with it. The page opens on the course whose climbing section (30 m or more of ascent, after the first 5 km, whose pace also reflects the start’s crowding) ran slowest against elsewhere. A route profile that sits at or below 0 m for 0.5 km or more away from the polders records the water under a bridge rather than its deck; those courses are not drawn or offered there.

The earlier-result comparisons and the questions on improvement gains and returning to a course need an earlier recorded result.

The outcome is each edition's median individual percentage pace change from 0–20 km to 20–40 km: 100 × ((time at 40 km − time at 20 km) / time at 20 km − 1). Positive is slowing. These are equal-distance blocks, not measured half-marathon splits; the last 2.195 km is excluded.

Each city/year must identify exactly one race and one weather row with a valid same-year race date. Keep at least 100 complete eligible finishes per edition and weight each retained edition equally. The number of finishes is not the number of independent weather observations.

Use the producer's observed_at archive hour only when it is within 30 minutes of the supplied scheduled local start; either nearest hour is accepted at an exact half-hour tie. One reviewed exception is accepted by name: Chicago 2025 uses the 07:00 archive hour for its 07:35 scheduled start, 35 minutes earlier. Temperature, dew point, humidity and wind must agree with that hour. Warming is temperature exactly four hourly readings later minus temperature at that start hour; no runner's finish time determines the window. Supplied start times may be approximate or refer to elite rather than mass starts.

Fit one prespecified joint linear model of edition median pace change, with course fixed effects, a centered linear calendar-year term, centered start temperature and its square, start dew point, first-four-hours temperature change and start wind speed. Course denotes the supplied city, not a verified historical route. Course fixed effects are removed by within-course centering.

Report each weather coefficient multiplied by that exposure's observed edition interquartile range (75th minus 25th percentile). This is a conditional regression contrast, not a prediction that sets dew point beyond the physically supported temperature range. The other modeled variables are held constant.

Uncertainty resamples whole courses with replacement, keeping every edition within the drawn course and recomputing the model. The fixed seed is 20260911; 6,000 draws produce percentile intervals at 98.333333% confidence, an approximate Bonferroni adjustment for the three screened exposures. At least 95% of draws must have full rank. Repeated runners across courses are not separately clustered.

Before seeing outcomes, require 30 editions, 10 courses, within-course exposure ranges at least half the overall IQR in at least 10 courses, and variance inflation factor no more than 10. A supported association additionally needs an interval excluding zero, an absolute IQR estimate of at least 1 percentage point and unchanged direction in every leave-one-course-out refit. A precise null requires the entire interval and every leave-one-course-out estimate strictly inside ±1 percentage point. This 1-point threshold is a prespecified practical publication threshold, not a physiological boundary or a claim that smaller changes never matter to runners. All three screen results remain in the audit, including withheld candidates.

This is an exploratory, adjusted edition-level association among eligible finishers. It does not establish a weather penalty, an individual forecast or a causal effect. Field composition, fitness, opening congestion, race organization, route changes and unmeasured conditions remain possible explanations. The opening 0–20 km baseline can itself be affected by weather and congestion. Runners who did not finish or lack valid splits are not represented.

Relative humidity changes with temperature. Dew point is the modeled moisture exposure; the raw humidity percentage is retained for context. Wind speed gives no runner-specific headwind or tailwind without a valid historical route, direction and personal exposure timeline. Scheduled-start weather is not personal wave-start weather.

Use complete, strictly increasing elapsed checkpoints at 5, 10, 15, 20, 25, 30, 35, 40 and 42.195 km. Clock strings must parse as H:MM:SS or M:SS. No missing splits are interpolated.

Remove exact duplicate race records, ignoring database IDs, ingestion timestamps and source URLs. Retain finishes from 90 minutes to 12 hours with every section between 2 and 20 minutes per km. These quality filters can exclude genuine unusual performances; the analysis describes this eligible cohort, not every entrant.

Apply the reviewed source-quality edition exclusions for this exact export after the timing checks. Known invalid split grids (including grids that mix timing bases, a checkpoint that stopped recording partway through the field and a virtual edition), incomplete ingestion, unreconciled HOLD editions and selected fields (a top-finisher list, or finishers cut off when a race was stopped) do not contribute to the analyses or prior benchmarks. Report source exclusions separately; an already invalid timing row is not counted twice. Missing age or recorded gender alone does not exclude an otherwise eligible finish from the overall cohort. Other sparse editions are not declared incomplete merely from their size.

For each runner, divide section pace by their full-marathon average pace and subtract one. Plot the median of these individual percentages at each checkpoint; zero is that runner’s own marathon pace. A median curve is not itself one runner’s race and need not integrate to zero.

Compare elapsed time over 0–20 km with 20–40 km. Faster: more than 2% faster; similar: within 2%; moderate slowing: more than 2% through 10%; pronounced slowing: more than 10%. The final 2.195 km appears in the profile but not this equal-distance classification. These are descriptive categories, not slowdown episodes or measured half-marathon splits.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 4,108,749 eligible finishes.

Keep in mind: Calculated directly from 4,108,749 complete, plausible split records in the public export.

Reading the answer: The full-course profile also shows how pace changes within those broad blocks.

Method and data coverage (JSON)The typical pace profile (CSV)How the second 20 km compares with the first (CSV)

Define a substantially improved performance as a finish more than 2% faster than the recent recorded best. Compare each section’s pace with the full-marathon pace of that earlier benchmark, then plot the median by outcome group.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes.

Keep in mind: Compare where those improvements appeared across the race. This benchmark uses earlier results only.

Method and data coverage (JSON)Pace relative to the earlier benchmark (CSV)

Match faster, similar and slower openings within city, year, race, recorded gender and 15-minute bands of recent recorded best. Keep strata with at least 20 in all three groups. Weight every group by the smallest group count in that stratum.

Show the weighted mean percentage change from the prior benchmark. The lower and upper limits are percentile confidence limits from 500 bootstrap draws of whole race editions, seed 20260908. This captures edition clustering but not dependence when the same runner appears in different editions. No multiple-comparison significance claims are made.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 439,681 eligible finishes.

Keep in mind: The chart uses a benchmark from earlier years. It does not establish what would happen if an individual changed strategy.

Reading the answer: Improved fitness could produce both a faster start and a faster finish, so this is not a recommendation to start aggressively.

Method and data coverage (JSON)Performance after matching prior ability and race edition (CSV)

A slow start is 0–5 km pace more than 5% slower than the earlier benchmark’s average. Group the change from 0–5 to 5–10 km as more than 5% acceleration, 2–5% acceleration, or less acceleration/slowing. Plot the 10th, 50th and 90th percentiles of final performance change. These groups are not matched on course or fitness.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 60,634 eligible finishes.

Keep in mind: A slow opening could reflect restraint, terrain or crowding. The available splits cannot identify the cause.

Reading the answer: These groups were not otherwise matched.

Method and data coverage (JSON)Finish outcomes after a slow start (CSV)

Compare the 15–20 km pace with 10–15 km. Accelerating is more than 2% faster, slowing more than 2% slower, and steady within 2%. Match within city, year, race and the same floored minute of elapsed time at 20 km.

Only strata with at least 20 finishes in each of the three groups qualify. Use the smallest group count as each stratum’s common weight for all three groups, then average the within-stratum finish-time differences versus steady. Sample counts are actual observations, not matching weights.

Runners arrive within the same 60-second band, not at an identical instant. Prior fitness and runner intention are unknown. This is a conditional association, not proof that accelerating causes a better finish or an out-of-sample prediction.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 2,123,626 eligible finishes.

Keep in mind: The comparison holds the race edition and the one-minute 20 km time band constant.

Method and data coverage (JSON)Finish-time difference versus steady runners (CSV)

At each checkpoint, start with elapsed time × 42.195 / distance. The elapsed-only model multiplies this by the training median actual/projected ratio in 30-second-per-km elapsed-pace bands. The trend model adds the most recent section’s change from the previous section (faster than −2%, within ±2%, or slower than +2%). At 5 km no trend exists.

Hold out the latest three observed calendar years (2024–2026). Fit every factor and the 10th–90th percentile ratio interval on earlier years only. Training cells need at least 100 records; missing trend cells fall back to the pace-band model, then the pooled training model. No future splits, finishing-time groups, identities or supplied ability fields enter a prediction.

Report median absolute error for all three methods, plus observed coverage and median width of the nominal 80% prediction interval and the 90th percentile absolute error. Model selection is fixed before examining these results. A runner may occur in training and test in different years; identities are not used. Results apply to complete eligible finishers and do not predict withdrawals.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 1,027,178 eligible finishes (validated forecast).

Keep in mind: All model fitting uses editions before 2024; all checkpoints are evaluated on the same 1,027,178 complete finishes in the later years.

Method and data coverage (JSON)Median finish-time error in later race years (CSV)Observed coverage of the 80% prediction interval (CSV)

Scan 20–25, 25–30 and 30–35 km in order. A patch is the first section more than 10% slower than both the immediately preceding section and the 5–20 km baseline pace. Recovery means the next full 5 km is no more than 5% slower than that baseline.

Every runner contributes at most one patch. The outcome is next-section recovery, not a diagnosis or necessarily a return maintained to the finish. Course sections and different runner mixes can account for differences across distances. Five-kilometer timing cannot distinguish stops, walking, fatigue or terrain.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 1,264,439 eligible finishes.

Keep in mind: Recovery is assessed after the first qualifying patch; the patch is not selected because recovery later occurred.

Reading the answer: A qualifying patch was more than 10% slower than both the preceding section and the early baseline.

Method and data coverage (JSON)Returned near baseline in the next 5 km (CSV)

Within each complete-race pattern, divide finishes more than 2% faster than the earlier benchmark by all linked finishes with that pattern. This conditions on a pattern known after the finish; it is a retrospective association, not a pre-race strategy trial.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes.

Keep in mind: We compare equal 20 km blocks because no measured halfway checkpoint is available.

Reading the answer: These are observed associations, not proof that a particular split pattern causes success.

Method and data coverage (JSON)Improved finishes within each race pattern (CSV)

Restrict to finishes more than 2% faster than the recent recorded best. Divide the number in each equal-distance pattern by all such improved finishes. Suppress any pattern with fewer than 100 improved finishes; if a pattern is suppressed, visible shares need not sum to 100.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 238,209 eligible finishes.

Keep in mind: A common successful pattern can simply be a common pattern overall; compare this with the success rate within each pattern in the previous question.

Reading the answer: The popularity of a successful pattern is different from its success rate.

Method and data coverage (JSON)Race patterns among improved finishes (CSV)

Within each prior-time band and opening group, calculate the 10th, 50th and 90th percentiles of finish-time percentage change. Also calculate the share more than 5% slower than the earlier benchmark. These distributions are unadjusted for course and edition, and describe observed finishers only.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes.

Keep in mind: The spread is grouped by prior performance, while the second chart records finishes more than 5% slower than that prior benchmark.

Reading the answer: These are outcome ranges, not confidence intervals or causal strategy effects.

Method and data coverage (JSON)Outcome spread by prior performance (CSV)Finishes more than 5% slower than the earlier benchmark (CSV)

Compute individual section pace relative to each runner’s full-marathon average, then take the median by city and section. Every checkpoint within a city uses the same complete-record cohort.

Cities pool available race editions. Terrain, weather, field composition and route changes are not separated. Course-profile validity years are absent from CORE, so historical elevation has not been assigned to these runners.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 4,108,749 eligible finishes.

Keep in mind: Individual split profiles are available for 27 cities with at least 100 eligible finishes each.

Method and data coverage (JSON)Pace through the selected course (CSV)

Join a unique supplied course segment by city and exact checkpoint distance. Classify net grade above +0.15% as uphill, below −0.15% as downhill and the remainder near level. At each distance, plot median individual section pace relative to that runner’s full-marathon average.

The course profiles use GPX geometry and digital elevation models, sometimes smoothed over 800 m. Net grade conceals mixed climbs and descents; bridge decks, tunnels and route changes may be wrong. Historical races are compared with the available city profile as a proxy only. Different terrain groups contain different courses and fields; no causal hill penalty or physiological effort is inferred.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 4,108,749 eligible finishes (route proxy).

Keep in mind: The route files mostly describe recent editions. Their historical validity is unverified, so this cannot measure the effect of the hills actually run in every year.

Reading the answer: This is an exploratory alignment with a route proxy.

Method and data coverage (JSON)Observed pace alongside supplied terrain classes (CSV)

Use consecutive linked races in different calendar years, at most three years apart, with one recorded eligible race in each endpoint year. For each course pair, calculate the mean destination-minus-origin finish time separately for runners taking each race order, then average those two means equally.

Require at least 20 pairs in each order and at least 100 overall. Balancing order reduces simple order imbalance but cannot remove fitness, weather, aging, motivation or entry-selection effects. The same runner can supply more than one pair. Positive means a slower finish at the destination.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 108,588 linked race pairs (partial comparison).

Keep in mind: These are observed time differences across consecutive recorded races, not a personalized equivalent-time calculator.

Reading the answer: Fitness, weather and selection still differ; this is not a personal equivalent-time forecast.

Method and data coverage (JSON)Observed time difference at the destination (CSV)

Within city and prior-time band, report the 10th, 50th and 90th percentiles of finish-time change versus the earlier benchmark. Pool available editions and require at least 100 observations per city and band. A wide percentile range describes individual variation, not uncertainty in the median.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes.

Keep in mind: Prior-time bands improve comparability, but this is still a comparison of different people and race days.

Reading the answer: The charts show the much wider spread within each course; these are different fields, not a course ranking.

Method and data coverage (JSON)Performance and spread on each course (CSV)

Familiar means at least one eligible linked appearance in the same city in an earlier calendar year. Match familiar and first-recorded groups within edition, recorded gender and 15-minute prior-time bands; each group needs 20 finishes per stratum. Use the smaller stratum count as a common weight.

The outcome is mean percentage change from 0–20 to 20–40 km. Both groups have some prior recorded marathon history, but familiarity itself is not randomly assigned. Route changes and visits absent from the dataset are unknown.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 328,524 eligible finishes.

Keep in mind: A first appearance in this export may not be a runner’s first visit to the course, and returning runners are a selected group.

Reading the answer: That association does not isolate the benefit of knowing the course.

Method and data coverage (JSON)Pace retention after matching prior ability (CSV)

Use the supplied Open-Meteo archive hour nearest the scheduled local start. Join the single city-year weather row whose race date parses and matches the record year. Group temperature below 10°C, 10–14.9°C, 15–19.9°C and at least 20°C.

First calculate each edition’s median outcome among linked runners with a recent benchmark, requiring 100 finishes. Then average edition medians equally within temperature bands, requiring five editions. The profile is normalized by each runner’s own marathon average; performance is relative to the earlier benchmark.

The modeled weather is a start-hour proxy, not each runner’s exposure. Start offsets are absent, temperatures change during the race and humidity, wind, sunshine, terrain and fitness remain potential confounders. Edition counts in the source table are the number of weather exposures; finish counts are not independent weather observations.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes (weather proxy).

Keep in mind: Each race edition receives equal weight. Temperature bands still contain different courses and fields, so these differences are not a temperature penalty.

Reading the answer: The course and field mix differs, so this is not a measured temperature penalty.

Method and data coverage (JSON)Open-Meteo historical weather methodsPace through the race by start-hour temperature (CSV)Performance relative to earlier races (CSV)

For every city and race year, calculate the median and 10th–90th percentiles of percentage finish change versus each linked runner’s recent recorded best. Require 100 such finishes per edition. Compare an individual’s percentage change with that edition median to describe their position relative to the field.

This contemporaneous edition reference is retrospective, includes the runner when eligible, and may shift with selection and fitness changes. It is not a weather correction or a prediction available before race day.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes (partial comparison).

Keep in mind: The chart supplies a race-day benchmark. It cannot divide a runner’s lost time into weather, pacing, illness or other causes.

Reading the answer: A race-wide shortfall and an individual shortfall answer different questions.

Method and data coverage (JSON)How the field performed against its earlier benchmarks (CSV)

Chip elapsed times do not establish physical proximity when runners start in different waves. Absolute checkpoint times and individual start offsets are needed to compare group continuity and race outcomes.

Why it has no answer: The current release cannot establish who was physically running together.

Measure: Pace consistency, surges, group continuity, and finish outcomes.

Compare: Runners who stay with a group, join or leave one, or run without a sustained group, at comparable ability and course positions.

Still needed: Absolute checkpoint timestamps and start offsets. Similar chip elapsed times do not establish physical proximity.

Current source release

A slow first section alone does not establish congestion. Start waves, corrals, clock times and comparable earlier performance are needed.

Why it has no answer: The current release cannot separate a crowded start from deliberate restraint.

Measure: Start delay, first-section pace, catch-up behavior, and the remainder of the race.

Compare: Runners of similar prior ability across waves and corrals. Use crowd and timing evidence to distinguish congestion from deliberate restraint where possible.

Still needed: Individual chip and gun times, wave or corral assignments, checkpoint clock times, and dated start arrangements.

Current source release

Normalize section pace by each runner’s own full-marathon average, then take the median by city and performance group. Performance groups use more than 2% faster than, within 2% of, or more than 2% slower than the prior benchmark. Require 100 per city and group.

The outcome group is known after the race. Courses pool editions; neither terrain versions nor race-day conditions are held constant. This is a descriptive course-response profile, not evidence of optimal effort allocation on hills.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes (partial comparison).

Keep in mind: A stronger finish does not prove better terrain adaptation. Historical route validity and direct effort measurements are still missing.

Reading the answer: This identifies how their pace distributions differ on the same named course.

Method and data coverage (JSON)Course pacing profile by performance group (CSV)

On pace means elapsed time within ±1% of the target’s even-pace time at that checkpoint. A goal is achieved only when the finish is strictly under the target. Divide successes by all eligible on-pace finishes for that goal and checkpoint.

This window includes runners just ahead of and just behind the target; it is narrower than the older core table’s time-budget definition. Results pool editions and are descriptive historical frequencies, not a validated personal forecast. Goals are inferred benchmarks, not runners’ declared intentions.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 4,108,749 eligible finishes.

Keep in mind: Each checkpoint selects a different group of runners.

Method and data coverage (JSON)Target achieved among runners close to target pace (CSV)

Rank the same complete-record finishers at 30 km and at the finish within each city, year and race. Ties receive their average rank. Positive rank change means gaining places; all signed changes sum to zero within each edition.

Convert rank changes to percentile points using field size minus one. A five-point move corresponds to about 500 positions in a 10,000-person eligible field. Missing-split finishers and non-finishers are absent. Different start waves mean elapsed-time ranks cannot count physical passes.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 4,108,749 eligible finishes.

Keep in mind: These are changes in elapsed-time rankings, not a count of physical overtakes.

Method and data coverage (JSON)How much elapsed-time position changes after 30 km (CSV)

Project finish time at 40 km by multiplying elapsed time by 42.195/40. Keep projections within five minutes of a round target and divide the margin into four bands. The kick measure is 100 × (40–42.195 km pace / 35–40 km pace − 1); negative means a faster final section.

Plot the median kick and the share finishing strictly below the target. Different courses, ability and fatigue can produce the same projected margin. These are unadjusted associations, not evidence that a milestone caused a sprint.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 1,219,011 eligible finishes.

Keep in mind: Groups are defined at 40 km, before the finishing section. Round times are possible incentives, not declared goals.

Reading the answer: A finishing acceleration does not by itself establish a goal-driven sprint.

Method and data coverage (JSON)Final-section pace change (CSV)Target achieved from that position (CSV)

The B.A.A. history lists 18–34 standards of 3:05 for men and 3:35 for women for 2013–2019 Boston races; the 2020 standards became 3:00 and 3:30, announced in September 2018. Compare 2016–2017 performances with 2019, excluding the transition year. Exact ages 18–31 avoid crossing the 35-year boundary within the next two years.

Around each fixed old/new benchmark, select finishes within ±2 minutes. Within each city and period, calculate the share at or below the benchmark. Keep cities with at least 20 close finishes in both periods and weight both periods by the smaller city-period count. Report at least 100 observed finishes per displayed period.

Acceptance cutoffs, declared intentions, age on a future Boston race day, certification and qualifying windows are not assigned to individuals. Round-number appeal, historical field changes and anticipatory behavior can explain bunching. This is not a difference-in-differences causal estimate, and the comparison does not cover every rule change.

Official sources: https://www.baa.org/races/boston-marathon/qualify/ ; https://www.baa.org/news/2020-boston-marathon-qualifier-acceptances-announced/ . Historical values verified 2026-09-08.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 854 eligible finishes (partial comparison).

Keep in mind: This initial analysis focuses on ages 18–31 in 2016–2017 and 2019; it does not assign individual Boston eligibility or acceptance.

Reading the answer: Only 1 city met both-period sample requirements near the faster standard. This narrow comparison cannot establish that the qualifying-rule change altered pacing.

Method and data coverage (JSON)B.A.A. historical qualifying standardsB.A.A. 2020 standards and acceptance announcementAt or below the benchmark among nearby finishes (CSV)

At 20 km keep elapsed time from 99% through 100% of the target’s even-pace budget. Find the first 25, 30, 35 or 40 km checkpoint where elapsed time exceeds that budget. Divide eventual strict sub-target finishes by all such first-crossing observations.

This uses inferred round-time benchmarks and elapsed chip times. It cannot establish when a runner mentally abandoned a goal. First-crossing groups differ, and results include only complete eligible finishes.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 306,308 eligible finishes.

Keep in mind: Each runner was on or slightly ahead of the target at 20 km, and then first crossed behind it at a later checkpoint.

Reading the answer: These checkpoints select different runners.

Method and data coverage (JSON)Still finished under the target (CSV)

A finishing acceleration is the percentage change in pace from 35–40 to 40–42.195 km. Group it as more than 5% faster, up to 5% faster, or similar/slower. Use consecutive cross-year linked pairs at most three years apart.

Measure next-finish percentage change relative to the first finish, report its 10th/50th/90th percentiles within the first race’s pacing pattern, and the share improving by more than 2%. This does not measure effort reserves, account for absent follow-up, or establish what would have happened with a harder earlier effort.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 681,266 linked race pairs (partial comparison).

Keep in mind: The comparison conditions on returning for another recorded race, and the next race may have a different course and conditions.

Reading the answer: This does not prove unused capacity in the earlier race.

Method and data coverage (JSON)Performance in the next race (CSV)Next finish improved by more than 2% (CSV)

Define the first section after 20 km whose average pace is more than 10% slower than 5–20 km. Look only through 40 km so all tested sections are 5 km long. Group runners using their observed 5–20 km pace.

Among runners with such a section, plot the distribution of its end distance. For each distance and early-pace group, show median elapsed arrival at the start and end of that section. The medians bound a typical observation interval; they are not confidence limits. This broader 10% section definition is separate from the sustained slowdown definition.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 2,775,783 eligible finishes.

Keep in mind: The timing mats locate the first slow section, not the exact instant the pace changed. This descriptive comparison cannot identify a distance-driven or time-driven cause.

Reading the answer: Distance and elapsed time describe different aspects; these observations do not identify the cause of slowing.

Method and data coverage (JSON)First slow section by early-race pace (CSV)Elapsed time at the two timing mats (CSV)

Use consecutive linked finishes in different years, no more than three years apart, with exactly one recorded eligible finish in each endpoint year. Calculate Pearson correlation between the two 0–20 versus 20–40 km pace changes, grouped by calendar-year gap.

For each earlier race pattern, divide the number of next races in each pattern by all eligible pairs with that earlier pattern. These conditional transition percentages sum to 100 within the earlier pattern. Correlation is descriptive; repeat observations of a runner are not independent.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 681,266 linked race pairs.

Keep in mind: Persistence may reflect the runner, repeated course choices, or shared conditions. It is not an immutable pacing personality.

Method and data coverage (JSON)Similarity in pace retention between races (CSV)Pattern in the next recorded race (CSV)

Use exact integer ages from 18 through 89, grouping 18–29 then ten-year bands. Do not infer exact ages from age-group labels. Report median 0–20 km pace and median percentage pace change from 0–20 to 20–40 km, separately by recorded gender.

This is cross-sectional: different people, races and performance levels are being compared. Selection into marathons, missing ages, course and prior ability can explain part of the pattern. It is not an estimate of an individual’s aging trajectory.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 1,283,674 eligible finishes.

Keep in mind: This comparison uses 1,283,674 finishes with an exact age and a recorded women’s or men’s category. Age-group-only records are excluded.

Reading the answer: These records can describe those differences, but they cannot tell us which changes first as an individual ages.

Method and data coverage (JSON)Opening 20 km pace by age (CSV)

Compare average percentage change from the first 20 km to the second 20 km. The pooled chart uses all eligible women’s and men’s records. The matched chart uses only race-edition and one-minute 20 km strata containing at least 20 finishes in each category.

Use the smaller category count as a common stratum weight. Thus both matched estimates have the same race and opening-time distribution. Opening time is an observed race performance, not an independent measure of prior fitness. Age, experience and other factors remain uncontrolled.

The export field is sex; labels follow its recorded categories. Other or missing categories remain in the overall pacing analyses but are not included in this two-category comparison.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 3,528,456 eligible finishes.

Keep in mind: Matching the opening performance makes the comparison more specific, but does not establish why the difference exists.

Method and data coverage (JSON)Pooled pace change (CSV)Pace change after matching the race and opening time (CSV)

Select linked finishes within two minutes of each round target. Under is strictly below the target; an exact target time belongs to the at-or-over group. Require two subsequent calendar years with at least 100 eligible finishes in the index city, and exclude the latest two observed years as index years.

Return means any eligible linked finish anywhere in the export during the next two calendar years. Same-year returns do not count. Divide returns by every eligible index finish in each group, including those with no observed return. Coverage checks reduce administrative censoring but do not establish complete race ingestion. Repeat index observations, false/missed links, changing coverage and unrecorded goals can affect the association.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 266,787 eligible finishes (partial comparison).

Keep in mind: A missing later record means no eligible linked appearance was observed here. It does not mean the runner stopped racing.

Reading the answer: This measures recorded return, not whether somebody continued running.

Method and data coverage (JSON)Observed again in the following two years (CSV)

Count eligible linked finishes in strictly earlier calendar years. Show median opening pace relative to the recent benchmark and median change between the two 20 km blocks by prior count.

Separately, use consecutive cross-year pairs to calculate the median next-minus-previous pace-retention change, grouped by the previous pattern. Selecting an unusually good or bad first race creates regression to the mean, so improvement after pronounced slowing is not proof of learning. Continued participation and changing fitness also affect these comparisons.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes.

Keep in mind: Experience here means appearances in this dataset. The first observed race is not necessarily a marathon debut.

Reading the answer: Age, course choice and who keeps racing can affect this comparison.

Method and data coverage (JSON)Pacing by recorded experience (CSV)How retention changed in the next race (CSV)

Use only linked identity groups for which every eligible record has a unique supplied race date. Order by that date, pair adjacent races and retain intervals of 1–1095 days. Unlike the year-based analyses, this includes same-year pairs. Dates come from the supplied calendar overlay and have not all been independently reverified.

Calculate 100 × (next finish / previous finish − 1) and its 10th/50th/90th percentiles in the displayed day bands. The second view requires the earlier race to beat every recorded finish in earlier calendar years; same-year bests are not used for that label. Fitness, course, motivation and selection into short or long intervals remain confounders.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 714,482 dated race pairs.

Keep in mind: These are observed scheduling choices, not recommendations for a recovery interval.

Reading the answer: The data do not identify an optimal recovery interval.

Method and data coverage (JSON)Next performance by recorded race interval (CSV)

For each city and year, calculate median first-20-km pace and median percentage pace change from the first 20 km to the second. Show city-years with at least 100 eligible finishes, and cities with at least three available years.

The same cities, runners and course versions are not represented every year. Choose a city to avoid pooling changing city coverage into one global trend. Even within a city, field composition, route changes and conditions remain uncontrolled. Gaps are years without publishable data, not interpolated observations.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 3,952,130 eligible finishes.

Keep in mind: The charts show observed yearly medians; they do not isolate changes in training, shoes or fueling.

Method and data coverage (JSON)Opening speed over time (CSV)

For each eligible linked finish faster than every eligible recorded finish in strictly earlier calendar years, select the fastest earlier-year record as comparator. Break tied best times by earlier year and then stable raw record ID. Both races must have all nine valid checkpoints.

Subtract current from earlier elapsed time over 0–10, 10–30 and 30–42.195 km. Positive means time gained. Compute means so that the three block means sum exactly to the mean finish improvement; separately divide each block by its own distance to show seconds gained per kilometer. Assert that block gains sum to total gain for every pair and in the published means. Different courses and conditions can contribute to gains.

Shared rules it follows: 1, 2, 4, 5, 8, 9.

Eligible sample: 363,511 earlier-best comparisons (partial comparison).

Keep in mind: This is a best within the recorded earlier-year history. Races absent from the export and other races in the same year prevent a lifetime-PB claim.

Method and data coverage (JSON)Where the finish-time improvement appeared (CSV)

Average positive section slowing relative to the 5–20 km pace, weighted by section distance. Faster sections contribute zero. This differs from a sustained episode and is not a physiological diagnosis.

Eligible sample: 4,108,749 eligible finishes.

Reading the answer: The severity distribution describes average positive slowing across 20–42.195 km.

Current study results, methods and source checksums (JSON)

Use the supplied Open-Meteo archive hour nearest the scheduled local start. Join the single city-year weather row whose race date parses and matches the record year. Group temperature below 10°C, 10–14.9°C, 15–19.9°C and at least 20°C.

First calculate each edition’s median outcome among linked runners with a recent benchmark, requiring 100 finishes. Then average edition medians equally within temperature bands, requiring five editions. The profile is normalized by each runner’s own marathon average; performance is relative to the earlier benchmark.

The modeled weather is a start-hour proxy, not each runner’s exposure. Start offsets are absent, temperatures change during the race and humidity, wind, sunshine, terrain and fitness remain potential confounders. Edition counts in the source table are the number of weather exposures; finish counts are not independent weather observations.

Shared rules it follows: 1, 2, 4, 5, 6, 7, 8, 9.

Eligible sample: 648,914 eligible finishes (weather proxy).

Keep in mind: Each race edition receives equal weight. Temperature bands still contain different courses and fields, so these differences are not a temperature penalty.

Reading the answer: The course and field mix differs, so this is not a measured temperature penalty.

Method and data coverage (JSON)Open-Meteo historical weather methodsPace through the race by start-hour temperature (CSV)Performance relative to earlier races (CSV)

Half-open one-minute windows immediately before and after 3:00, 3:30 and 4:00. These counts use eligible complete finishes. They do not reveal declared goals or establish a psychological mechanism.

Eligible sample: 4,108,749 eligible finishes.

Reading the answer: Compare the recorded finishes in the minute before and the minute after each milestone.

Current study results, methods and source checksums (JSON)

For each runner, divide section pace by their full-marathon average pace and subtract one. Plot the median of these individual percentages at each checkpoint; zero is that runner’s own marathon pace. A median curve is not itself one runner’s race and need not integrate to zero.

Compare elapsed time over 0–20 km with 20–40 km. Faster: more than 2% faster; similar: within 2%; moderate slowing: more than 2% through 10%; pronounced slowing: more than 10%. The final 2.195 km appears in the profile but not this equal-distance classification. These are descriptive categories, not slowdown episodes or measured half-marathon splits.

Shared rules it follows: 1, 2, 3, 4.

Eligible sample: 4,108,749 eligible finishes.

Keep in mind: This overview describes the current eligible field. Open the pacing-pattern analysis to choose comparable finishers. It does not prescribe an optimal strategy.

Reading the answer: The full-course profile also shows how pace changes within those broad blocks.

Method and data coverage (JSON)The typical pace profile (CSV)How the second 20 km compares with the first (CSV)

At each checkpoint, start with elapsed time × 42.195 / distance. The elapsed-only model multiplies this by the training median actual/projected ratio in 30-second-per-km elapsed-pace bands. The trend model adds the most recent section’s change from the previous section (faster than −2%, within ±2%, or slower than +2%). At 5 km no trend exists.

Hold out the latest three observed calendar years (2024–2026). Fit every factor and the 10th–90th percentile ratio interval on earlier years only. Training cells need at least 100 records; missing trend cells fall back to the pace-band model, then the pooled training model. No future splits, finishing-time groups, identities or supplied ability fields enter a prediction.

Report median absolute error for all three methods, plus observed coverage and median width of the nominal 80% prediction interval and the 90th percentile absolute error. Model selection is fixed before examining these results. A runner may occur in training and test in different years; identities are not used. Results apply to complete eligible finishers and do not predict withdrawals.

Shared rules it follows: 1, 2, 4, 8, 9.

Eligible sample: 1,027,178 eligible finishes (validated forecast).

Keep in mind: This page shows the current checkpoint model’s held-out validation. Use the interactive checkpoint comparison for a selected course and elapsed time; 20 km is before halfway.

Method and data coverage (JSON)Median finish-time error in later race years (CSV)Observed coverage of the 80% prediction interval (CSV)

Recorded or modeled weather from past editions must not be presented as a forecast for an upcoming race.

Why it has no answer: This database contains historical race weather, not live race-day forecasts.

Current source release

Historical course validity ranges are missing. The supplied city route can support an explicitly labeled terrain proxy, but it cannot establish a runner’s physiological effort or how much time hills caused.

Why it has no answer: A validated course-adjusted pacing estimate is not available from the current release.

Current source release

The course shape. On a course whose pace band publishes your goal, your plan and its wristband strips take the course’s shape by default. The group is the pace band’s for that course, all recorded genders (on the pace band, the recorded gender you choose): every screened finish from G − 5:00 to G − 0:01, where G is the whole minute at or below your goal, each with all nine official timing points (the 5–40 km mats and the finish). The group’s median elapsed time at each timing point is multiplied by your goal divided by its median finish, so the band ends exactly on your goal, seconds kept, and keeps the course’s proportions. Between timing points the band holds each segment’s pace evenly, so its mile and kilometer rows and halfway are interpolated within a segment, not observed: the observed facts are the nine timing points, and the planner’s card shows only those (5K, 20K, 30K and the finish), with the pace through 20K and from 30K to the finish. A shape is published wherever the group has 100 or more finishes from at least 3 editions, the same minimum as the course results below: one or two race days describe their own weather and fields more than the course, so a course where fewer editions passed the screens (Singapore one; Auckland, Barcelona and Paris two) has no course shape, and neither has a goal at the edge of a course’s range whose group pools fewer editions than that. Otherwise, and on all courses together (there is no course to shape it), the band is your goal spread evenly, the even pace. Both pages then say in one line, where the choice would be, why there is no shape: the pace band has no group for the course (or, on the pace band, for the recorded gender you choose), too few editions passed the screens for the course or at your goal, the goal is outside the published goals (2:30 to 6:30), or the course has too few finishes at your goal. Where a shape exists, the Splits choice switches between Course shape and Even pace; course shape is the default and stays out of the link, and even pace is kept in it (pace=even) from one course to the next, though not through All courses. The wristband strips are the same as the pace band’s.

The shape describes finishes that made the goal: the group is chosen by its outcome, so it says how those finishes were spread over the course, not how a race will go. It is not a prediction or a recommended plan, running these splits is not shown to produce that finish time, and it is a median of finishes, not anyone’s actual splits. Weather and the supplied route elevation do not enter it, and it is not an adjustment for hills.

On a course, the results are the course chooser’s row for that course (below). Rows are published for goals from 2:30 to 6:30 every 5 minutes, and the planner shows the nearest one, saying so when it differs from your goal. On all courses together there is no course row, so the planner shows the pace band’s group of finishes in the five minutes up to the goal.

With the course shape, the mat table shows each mat’s course-shape time beside its even-pace time and the number of finishes the shape comes from; the course shape is the one figure that combines your goal with Pace Notes data, by the scaling above. With even pace, the mat table and the “H” line on the strips are the pace band’s groups for the course (or all courses), all recorded genders, in the same window, split by whether they had a sustained slowdown; each group shows its median elapsed time at the official mats, your even-pace times sit in their own column, and no number subtracts one kind from the other.

When a course has no row at a goal, the reason is one of these: no edition passed the data screens; fewer than 3 editions had 20 or more finishes at that pace; or those editions held fewer than 100 such finishes in all. Race months and start temperatures come from one supplied, modeled observation per edition near the scheduled start, not anyone’s exposure. The climb is the sum of the 5 km sections of the supplied current route, whose historical validity is unknown.

Pace is time divided by distance, and speed is distance divided by time. A marathon is 42.195 km (26.2 miles) and a half 21.0975 km. Every split row is the elapsed time at an even pace; with a second-half difference, the elapsed time at an even pace within each half. A custom distance can be from 0.1 to 1,000 km. Split tables stop at 500 rows, so very long distances offer only the longer split intervals. The same-pace table is pace × distance, not a predicted race time.

Certified road courses are measured along the shortest route a runner could take (World Athletics Rule 55.3). Measurers add 0.1%, the short course prevention factor, so a later re-measurement does not find the course short: each kilometer is laid out as 1,001 m. GPS watches usually record a slightly longer distance, so their average pace reads faster than course pace. How long a watch reads depends on tangents, buildings and tunnels, so the watch table shows several readings and assumes no typical error.

Source: World Athletics and AIMS, The Measurement of Road Race Courses (revised edition, 2023), Rule 55.3 and the short course prevention factor, checked October 7, 2026.

Every number on the pace charts is an even pace. The marathon chart lists the official timing mats every 5 km from 5 to 40 km, halfway at 21.0975 km and, in miles, every fifth mile. None of them are recorded splits, and the charts use no race data, weather, elevation or course.

Each goal page (3:00, 3:30, 4:00, 4:30 and 5:00) sets the even-pace splits for the goal beside the pace band’s All-courses group for it: every screened finish from 5:00 to 0:01 under the goal, all recorded genders, each with all nine official 5 km checkpoint times. These are achieved finishes, not stated goals. The page’s sentences are built only from that goal’s pace-band values.

A section’s median pace is the median of every finish’s own pace over that section, so it need not equal the gap between two mat medians. The middle half is observed variation between finishes, not uncertainty and not a plan.

The bunching panel uses the finish-time data behind the Round numbers story. A smooth curve is fitted to the minutes around the goal, and the ratio’s 95% interval comes from resampling whole race editions. The panel appears only when that whole interval is above 1.

Riegel (1981): T₂ = T₁ × (D₂ ÷ D₁)^b with b = 1.06, the value most calculators use. Riegel fitted the power law to records over efforts of about 3.5 to 230 minutes. Vickers and Vertosick (2,303 recreational runners) found it well calibrated for races up to the half marathon. For the marathon it was a median 10:09 too fast, and at least 10 minutes too fast for half of the runners; about three in four ran slower than predicted.

Daniels–Gilbert equations (the basis of VDOT): oxygen cost VO₂ = −4.60 + 0.182258v + 0.000104v² (v in meters per minute), and sustainable fraction 0.8 + 0.1894393e^(−0.012778t) + 0.2989558e^(−0.1932605t) (t in minutes). Their ratio scores the race, and the target time with the same score is found by bisection. They are not applied to a race slower than 10:00 per km, about walking speed. For the marathon they imply an exponent of about 1.05 to 1.06, so they are about as optimistic as Riegel.

Typical recreational range: 4,402 RunningAHEAD training logs had a half and a full within a year. Their half-to-full exponent was 1.09 most often, 1.13 at the median and 1.15 on average (SD 0.084). From a half or longer, the estimates are T × (42.195 ÷ D)^b for those three exponents. From a shorter race, Riegel 1.06 first gives a half-marathon equivalent, where Vickers and Vertosick found it well calibrated, and the exponents apply from there. That range is less certain, so its ends are drawn faded, and no published figure says how much wider real outcomes spread. This composition of published estimates is Pace Notes’ own, and nothing in it is fitted to Pace Notes data. Error grows with the distance ratio: Keogh et al. (2019) reviewed 114 marathon equations and found no single best one.

Personal exponent: b = ln(T₂ ÷ T₁) ÷ ln(D₂ ÷ D₁) from two races, applied from the longer one. The page warns when b falls outside 1.00 to 1.30 or when the races may be more than six months apart, and shows no number outside 0.90 to 1.50. In Vickers and Vertosick’s study a two-race regression model was the most accurate (about 14 minutes typical error).

Tanda (2011): marathon pace (s/km) = 17.1 + 140 × e^(−0.0053K) + 0.55P, where K is the average weekly kilometers and P the average training pace (s/km) over the eight weeks before the race. It was built on 22 runners and 46 marathons of 2:47 to 3:36, with a standard error of about four minutes, so outside that range no number is shown.

Not modeled: adding weekly mileage cut Vickers and Vertosick’s typical error from about 19.5 to about 15 minutes, but their coefficients are not reproduced. No course, weather, age or sex adjustment is applied, and every estimate assumes marathon-specific training, a flat course and cool weather.

The Pace Notes data panel takes the even pace of the estimate you pick and finds the race-day projector’s All-courses group at 20 km. That group holds complete finishes whose 20 km time put them in the same 2-minute band of even-pace finish (20 km time × 42.195 ÷ 20). It shows their finish percentiles (stored from the 5th to the 95th in steps of 5), the share under the estimate (interpolated between percentiles, so approximate) and the share with a sustained slowdown. Pace Notes holds marathons only, so it cannot test a half-to-full prediction.

Why marathon predictions run fast. The classic formulas assume your endurance carries over: Riegel’s 1.06 and the Daniels–Gilbert equations both describe runners trained for the distance, which is why the page labels Riegel the classic formula and sets the typical range beside it. On the page, Riegel 1.06 is the classic formula, Daniels–Gilbert the fitness equations and the half-to-full exponents (from the RunningAHEAD training logs) the typical range; the accuracy of the estimate from your two races is from Vickers and Vertosick’s study of recreational runners, and the estimate from training volume is Tanda (2011).

Vickers & Vertosick (2016), BMC Sports Science, Medicine and Rehabilitation 8:26RunningAHEAD half-to-full exponents, 4,402 training logsRiegel (1981), American Scientist 69:285–290Daniels–Gilbert equations (Oxygen Power, 1979), transcribedTanda (2011), Journal of Human Sport and Exercise 6(3)Keogh et al. (2019), IJSPP 14:1159

On All courses the band is an even pace for your goal, with mile rows and halfway at 21.0975 km. On a course whose window (below) for your goal and recorded gender holds 100 or more finishes, the band is the course shape from that course and recorded gender’s group, as on the race planner, unless you choose even pace. The watch option divides the goal pace by 1 plus the overrun you choose; in course shape, whose pace changes along the course, that figure is the watch’s average over the whole race and is labeled as an average. That overrun is your own assumption about how long your watch reads, not a typical error, and the band’s times still refer to the course markers. Most watches read a certified course a little long, so their pace runs quick (see the pace calculator). The link to the race-day projector keeps your course only when the projector publishes it, and the projector’s recorded-gender groups cover All courses only, so a course link drops the gender.

Few marathoners run evenly: in a large published analysis of US marathons, men ran the second half 15.6% slower than the first on average and women 11.7% (Deaner and colleagues, 2015). That is why the band sits beside what real finishes ran.

For a whole-minute goal G, the observed columns hold every screened finish from G − 5:00 to G − 0:01 on the chosen course (or all courses) and recorded gender, each with all nine official 5 km checkpoint times. A goal with seconds uses the whole minute at or below it (3:30:30 uses the 3:30 window), so every finish in the window still beat your goal. Each group shows the median elapsed time at the 5–40 km mats and the finish, the pace in each section and, where published, the 25th and 75th percentiles. The chart draws each group’s median elapsed time against even pace at every mat, even pace as its zero line. The window’s finishes split into those with a sustained slowdown and those with none.

A course that other Pace Notes data covers but the pace band does not publish has no five-minute window with 100 screened finishes. The optional elevation strip for the back of a printed band is the supplied current route, drawn on the same 200 m (656 ft) scale on every course so flat courses print flat. Its historical validity is unknown, and it may miss bridge decks.

Source: Deaner RO, Carter RE, Joyner MJ, Hunter SK (2015), Men are more likely than women to slow in the marathon, Medicine & Science in Sports & Exercise 47(3): 607–616.

For a goal G, even pace is G ÷ 42.195 km. On each course, a finish counts when its 5–20 km pace is within ±2% of that even pace. An edition counts when it has at least 20 such finishes. A course row is published only with at least 3 counted editions and 100 finishes in all. The 5–20 km pace is measured during the race, so a row describes finishes that happened to run that pace, chosen after the fact. It is not an estimate of what anyone can run.

Under goal is the share of the row’s finishes that came in below G, with all counted editions pooled, so large editions weigh more. The sustained-slowdown share and the time after 20 km are worked out per edition and then averaged, every edition counted equally. Finish-time percentiles (10th, median, 90th) pool all of the row’s finishes.

This is not a ranking. The default order is alphabetical, and no course is called fast, slow, easy or hard. Courses differ in field, qualifying rules (Boston has qualifying times), weather, era and route. Averaging editions equally stops one large or unusual year from dominating a row, but it does not remove those differences. No course factor or equivalent time is calculated. Finishes after a course closes are not recorded, so near a time limit a row describes only the finishes inside it.

Race months come from the dates of editions with a weather row. Start temperature is the range, across editions, of one supplied modeled observation per edition for the hour near the scheduled start at one point in the city. The route profile is the supplied current route: climb, descent and the end-to-end net change exactly as supplied (net is not climb minus descent), and historical routes may differ. A downhill opening means the first 5 km of that profile drops more than 25 m (82 ft).

Never published, with fewer than 3 editions after the screens: Auckland, Barcelona, Paris and Singapore, plus Prague, which has none.

The course comparison analysis uses wider pace bands. The pace band’s group (finishes within five minutes under the goal) differs from a row here, which is chosen on 5–20 km pace. So a course can have a row at a goal and still have fewer than 100 finishes just under it; those rows link to the all-course band.

Your start temperature is rounded to the nearest whole °C. The window is that center ±2 °C (±3.6 °F), or ±3 °C (±5.4 °F) if you widen it. A past edition matches when its start temperature lies in the window and it has at least 20 finishes whose 5–20 km pace falls in your 15 s/km pace band. A goal time is turned into its even pace and used as the 5–20 km pace, which is an assumption. On a warmer morning runners may already have started slower, so the same band can hold different runners on warm and cool mornings. A window and band are published only with at least 3 editions and 100 finishes.

The figures are the share of finishes with a sustained slowdown, the median time after 20 km, the median pace from 30 km to the finish against the 5–20 km pace, and each 5 km section’s median pace against the 5–20 km pace. Each is worked out per edition and averaged with every edition counted equally, because each edition is one weather observation. Finish-time percentiles pool all the finishes and only describe the field.

The reference column is the same pace band in editions that started at 6–10 °C (42.8–50 °F), inside the start-temperature band with the fewest sustained slowdowns in the race-day weather analysis. The two columns are never subtracted: the editions differ in course, field, year and route as well as weather, so the difference is not a heat effect, and no heat-adjusted time is calculated from Pace Notes data. Each edition has one supplied, modeled observation for the hour near its scheduled start at one point in the city. It is not anyone’s exposure: wave starts, sun, shade and wind on the course are unknown. Dew point is context only. The Pace Notes humidity question was withheld as inconclusive, which does not show that humidity has no effect. Historical weather is not a forecast, and the page uses only the forecast you type.

Warm data is sparse: of 198 editions with weather, 6 started at 20 °C (68 °F) or warmer: Stockholm 2024 (25.6 °C), Singapore 2025 (23.5 °C), Helsinki 2017 (22.1 °C), Stockholm 2011 (21.6 °C), Chicago 2021 (20.7 °C) and Austin 2017 (20.0 °C). Windows centered above about 21 °C rest largely on these few editions, and no published window is centered above 25 °C (77 °F).

Published research panel. Relative humidity uses the Magnus form (Alduchov and Eskridge 1996). Wet-bulb uses Stull (2011), valid for 5–99% humidity and −20 to 50 °C. Shade WBGT ≈ 0.7 × wet-bulb + 0.3 × air temperature. It is a shade estimate: direct sun can add several degrees, and a psychrometric wet-bulb only approximates the natural wet-bulb a WBGT meter reads. The simplified “ACSM-87” WBGT is not used because it is not an accurate approximation (Brimicombe et al. 2023).

The flags are the race-risk categories of the 1996 ACSM position stand on distance running (Armstrong et al.): green below 18 °C WBGT, yellow 18–23, red 23–28, black above 28. They describe heat-illness risk for the field, not a pace. The current ACSM consensus (Roberts et al. 2023) uses region-specific thresholds instead. Ely et al. (2007) is shown as the extra slowing over its coolest WBGT band (5–10 °C). Mantzios et al. (2022) is shown as 0.2% (elite marathon finalists) to 0.4% (all endurance events) per °C WBGT above 15 °C. Hadley’s table adds temperature and dew point in whole °F, and the RunnersConnect bands use the unrounded dew point. Hadley and RunnersConnect are rules of thumb, not studies. Ely and Mantzios describe elite and top finishers. Knechtle et al. (2019) is one race’s association (Boston 1972–2018), shown as reported and not applied to your goal. On the page, Ely et al. is labeled “Top finishers, seven marathons”, Mantzios et al. “Elite marathon finalists”, Knechtle et al. “Boston finishers, 1972–2018”, Hadley’s table “Temperature + dew point table” and the RunnersConnect bands “Dew-point bands”. Inside Mantzios et al.’s 7.5–15 °C WBGT best range no warm-side slope applies; the study also reports slower times in the cold, which the panel does not quantify. Each method stays a separate range, and minutes are the published percentage multiplied by your goal. None of these figures is fitted to Pace Notes data, combined with it or passed to any other tool.

Ely, Cheuvront, Roberts & Montain (2007), Med Sci Sports Exerc 39:487–493Mantzios et al. (2022), Med Sci Sports Exerc 54:153–161Knechtle et al. (2019), PLoS One 14:e0212797Stull (2011), J Appl Meteorol Climatol 50:2267–2269Armstrong et al. (1996), ACSM position stand on heat and cold illnesses during distance runningRoberts et al. (2023), ACSM expert consensus statement on exertional heat illnessBrimicombe et al. (2023), GeoHealthMark Hadley’s temperature + dew point table, as reproduced by Triple Threat LifeRunnersConnect dew-point bands

An elapsed time E at a 5 km mat implies an even-pace finish P = E × 42.195 ÷ mat km. Finishes are grouped in 2-minute bands of P, so “on 3:58–4:00 pace” means the same thing at every mat; at 5 km a band spans only about 14 seconds of elapsed time. Each group is one course (or All courses), one mat and one band, optionally narrowed by trend or by recorded gender. Gender and trend are never combined, and gender groups exist for All courses only.

Trend compares the latest 5 km pace with the average pace so far: (E at this mat − E at the previous mat) ÷ 5 against E ÷ mat km. More than 2% quicker is “faster”, within ±2% is “similar” and more than 2% slower is “slower”; there is no trend at the 5 km mat.

Each group stores finish percentiles from the 5th to the 95th in steps of 5, and the 10th, 50th and 90th percentile elapsed times at every later mat. It also stores the median pace over the rest of the race and the share of finishes with a sustained slowdown, split by whether it was already recorded by that mat. Shares under a target are interpolated between neighbouring percentiles, so they are approximate. Arrival windows run from the 10th to the 90th percentile, widened to whole minutes, so about one in five finishes arrived outside them. They are observed arrivals of similar finishes, not live tracking. Nothing is interpolated between mats. The even-pace times are elapsed time × distance ÷ mat km, and the needed pace is the time left to the target divided by the distance left.

How well it has held up. The same groups were rebuilt from 2005–2021 races only and scored on every 2022–2026 finish on All courses (about 1.48 million finishes). The 10th–90th range held the finish 78–80% of the time at every mat, against the 80% it should hold. The group median missed the real finish by a median 10:24 at 5 km, 7:25 at 20 km, 4:28 at 30 km and 0:52 at 40 km. Even pace missed by 12:18, 10:29, 6:32 and 0:44. With the last-5 km trend, the median miss fell to 6:22 at 20 km and 3:35 at 30 km. Course groups and gender groups were not tested separately. The checkpoint analysis elsewhere on the site uses its own elapsed-time groups and editions, so its figures differ slightly.

The projector’s groups use the same editions and screens as the other tools (see What the tool data includes).

Your sections come from your own times. Each section’s pace is its time divided by its length: eight 5 km sections between the official timing mats, then the final 2.195 km. Your reference is your own 5–20 km pace. Every percentage is a section’s pace divided by that reference, minus one, so +27% means 27% slower. Your opening is placed in one of the six opening groups. Only the mats you type are used: no halfway or mile splits are derived, and a missing mat is never filled in.

The sustained-slowdown reading is the same rule Pace Notes applies to every finish in its data. It describes the shape of a race without explaining it. Your pacing type compares each section’s pace with your own average pace for the whole race, not with the 5–20 km pace, and is the nearest of the six pacing types, with the same classifier as the pacing-types story. The share shown with it is that type’s share of the story’s finishes. Its shares come from that story’s cohort, which keeps Valencia 2021 and Warsaw 2022; the comparison leaves those out for a shifted mat grid.

The headline names the slowest 5 km section after 20 km and how much slower it was than your 5–20 km pace; a sustained slowdown needs 25% or more over at least 5 km. The 25% line in the section chart marks that threshold. The final 2.195 km can extend a run of slow sections but cannot count on its own, so the step list marks it as the final stretch.

The comparison window, for a finish F, runs from G − 5:00 to G − 0:01 with G = ⌊F ÷ 60⌋ + 3 minutes, so it holds your time with at least two minutes on each side. It contains every screened complete finish in that window on the chosen course (or all courses) and recorded gender, split into those with a sustained slowdown and those with none. When neither group reaches 100 finishes, the median of all finishes in the window is shown instead. All courses pools every race in Pace Notes, so if your race is not one of them, that group comes from other races.

Eligibility follows the rules used across Pace Notes: times must increase, every section must work out between 2 and 20 minutes per km, and the finish must be between 1:30:00 and 12:00:00. Everything is calculated in your browser. The data files are requested by course and recorded gender alone, without the page address.

Every standard, age rule, qualifying window and application date was transcribed from the race’s own pages and checked on October 8, 2026, the date each card shows. Standards change every year, so check the official page before you apply; Pace Notes reviews them before every registration season. Meeting a standard is not entry: each race then applies its own cut-off, cap, review or lottery.

Age rules differ by race. Boston, New York, Chicago and Sydney use your age on their race day. London Good For Age uses your age on the day you ran the qualifying time. Berlin uses the age you reach during the race year, so it goes by birth year. A February 29 birthday counts from March 1 in other years.

The margin is the standard minus the counted time, so a positive margin is under the standard. New York (“at least as fast as”) and Chicago (“within the standards”) accept a time equal to the standard. Boston, Berlin and Sydney do not say, so the checker counts an equal time as meeting them. London requires a time strictly under (“sub”). A result counts only if the race date falls inside the race’s qualifying window and no later than the application deadline, because proof goes in with the application. Boston 2028’s window runs through 2027 registration week, which the B.A.A. has not dated, so the checker treats a time run after September 2027 as outside it and flags one run during September 2027. The checker uses chip (net) time on a certified course; Sydney also accepts gun time. A race page marks a time rule the race does not state as “assumed; not stated by the race”. Each race page and checker card links the race’s official qualifying page; the standards are as that race publishes them.

Boston’s downhill index: a course’s net drop (start elevation minus finish elevation) adds 5:00 for 1,500–2,999 ft and 10:00 for 3,000–5,999 ft, and a drop of 6,000 ft or more is not accepted. A drop typed in meters uses the B.A.A.’s own metric thresholds (457.2 m, 914.2 m and 1,828.6 m), which are not exact conversions. The B.A.A. does not list affected courses, and its statements differ on how long the index will last.

Past cut-offs are not a forecast. The B.A.A. accepts qualifiers fastest relative to their standard until the field is full and announces the cut-off afterwards; neither the B.A.A. nor Pace Notes predicts cut-offs. For 2027 about 1,000 qualifiers who missed the cut-off were drawn at random. The “about 11%” is Pace Notes’ own calculation from B.A.A. counts: 1,000 ÷ (8,019 turned away + 1,000 drawn). New York’s capped-pool cut-offs are as reported by NYRR. Application badges compare your device’s clock with the official opening and closing times where the race states them, and with the listed dates otherwise. Tokyo is not included because its standards could not be verified from an official page.

Times to aim for are each standard minus your buffer (and Boston’s index for your drop), as an even pace. They are not advice about what buffer is enough. The even-pace table on each race page divides each standard by the marathon distance; it is calculated, not recorded, and real races are rarely run at an even pace.

B.A.A. qualifying standards and historyB.A.A. Boston Marathon rules and policies (non-binary entries)B.A.A. 2027 registration updateNYRR time qualifiersNYRR 2026 drawing resultsNYRR 2025 drawing resultsLondon Marathon Events: Good For AgeChicago Marathon: applyBerlin Marathon registrationSydney Marathon High Performance ProgramB.A.A.: record field of qualifiers accepted into the 131st Boston MarathonB.A.A.: 2026 and 2027 registration updates (downhill index)London Marathon Events: Championship entry

Open the qualifying checker

Every eligible finish (4,108,749 finishes) by its whole finish minute from 2:40 to 3:10, pooled over all editions; the minute before 3:00 is lit. Finishes pooled across editions.

Women's and men's finishes from 2:30 to 6:00 in start-screened editions without gender-recording gaps, matched on edition and finish minute (each cell weighted to its smaller group; matched weight 1,101,398 per group); the share of each with a sustained slowdown. Men's share is higher in 24 of the 24 races with a matched weight of at least 2,000. Pairs of linked records, one pair per consecutive race.

Every finish in start-screened editions (3,970,364 finishes): the time from 20K to 40K against the time to 20K (the last 2.2 km left out). Finishes pooled across editions.

Within each start-screened edition of at least 1,000 finishes, finishes sorted by recorded finish second; each neighbouring pair with the same second counts once (2,477,045 pairs), and the gap is the difference between their 20K splits (net running times, so not who ran beside whom). Pairs of linked records, one pair per consecutive race.

Records linked by name with four or more finishes in start-screened editions (75,766 records): their first four in race-date order, the fourth finish’s sustained-slowdown share by how many of the first three had one. Linked records are not verified people. Finishes pooled across editions.

The busiest finish minute (the clock minute holding the most finishes) of each edition with 3,000 or more finishes in Valencia, Boston and Berlin (36 editions, 1,142,315 finishes), grouped as the two minutes before 3:00, 3:30 or 4:00, or another minute; each edition counts once. Each edition counted equally.

Every New York finish with an exact recorded age (746,156 finishes, 16 editions), counted by age; a round age is compared with the mean of the two ages either side. Ages 40, 50 and 60 each stand more than 5% above their neighbours in 16 of the 16 editions. Finishes pooled across editions.

Every sustained slowdown in start-screened editions (3,970,364 finishes, 1,217,483 slowdowns), by the 5K mat where it began. Finishes pooled across editions.

Records linked by name (candidates, not verified people) returning to the same race 1–2 calendar years later, both finishes in editions that pass the start-offset screen: 534,804 pairs of consecutive finishes. Each line is the median change in each 5K section's pace (and the last 2.2 km) against the same record's previous visit, by the change in the edition's start temperature: warmer by 6 °C or more (54,712 pairs, 28 edition pairs, 13 races), within 1 °C (81,612 pairs, 40 edition pairs, 16 races) or cooler by 6 °C or more (61,014 pairs, 27 edition pairs, 10 races); the end notes are the median change in finish time. On much warmer returns the 35–40K change exceeded the 0–5K change in 25 of the 26 edition pairs with 100 or more pairs, and leaving out any one edition pair keeps the 35–40K change between +4.7% and +6.0%. Weather is edition context at the scheduled start, never a personal exposure. Pairs of linked records, one pair per consecutive race.

The twelve cities with the most eligible finishes (3,653,333 finishes), pooled over their editions: the share of each city's finishes slower than 4:00:00. It describes the fields, not the courses. Finishes pooled across editions.

Every pair of finishes within 2 seconds of each other at the 40K mat whose 30K splits were at least 30 seconds apart, in the 199 start-screened editions with at least 1,000 finishes (30,692,116 pairs; equal finish times dropped): the share where the one behind at 30K finished ahead, edition by edition (editions with at least 1,000 such pairs, or 100 within a gap band), as the median over editions; pooled over pairs it is 76%. Order is clock order, not road position; a pack of k level finishes gives k(k−1)/2 pairs. Each edition counted equally.

Women's share of men's and women's finishes at each exact recorded age, 18 to 80, in the granular-age editions (1,067,027 finishes; every age holds at least 107 finishes). Finishes pooled across editions.

Complete finishes in each race’s 2021 and latest edition as a share of its 2019 finishes, in the races with 2019 and 2021 editions whose eligible share of records moved under 5 points across the three editions (left out: Boston, Eindhoven, Madrid). Each edition counted equally.

Finishes from 3:59:00 to 3:59:59 in editions that pass the start-offset screen (42,160 finishes), set against 4:00 splits (a constant pace for 4:00:00) at every mat; the lines are the quartiles and median of that margin. Finishes pooled across editions.

Finishes in start-screened editions (794,475 finishes) whose sustained slowdown began in the 20–25K, 25–30K or 30–35K section; each later full 5K up to 40K is compared with the same finish's 5–20K pace. Finishes pooled across editions.

Every eligible finish (4,108,749 finishes), pooled, in two-minute bins of whole finish minutes from 2:00 to 7:00: of the 60-minute windows starting on a whole minute from 2:00 to 6:00, the one that holds the most finishes is lit. Finishes pooled across editions.

Finishes in start-screened editions of 23 races with a supplied route profile and 1,000+ such finishes (3,033,995 finishes): the share whose quickest of the eight full 5K sections was 0–5K, against the route's net elevation change over its first 5 km as a gradient (New York and Dubai, whose profiles read the water under a bridge, are left out); the line is a least-squares fit. The highlighted races are those whose first 5 km climbs more than 0.3% net, with the highest share among them noted; the falling-start figure is the median over races whose first 5 km falls more than 0.3% net. Finishes pooled across editions.

Men’s finishes under 3:00 as a share of men’s finishes in eight races, every edition of 1,000+ finishes without a gender-recording gap; each race’s 2022–25 mean share (editions equal) over its 2013–19 mean, then the median over the eight races. Entry rules differ by race, so it describes who finished, not why. Each edition counted equally.

Median finish by exact recorded age in the granular-age editions without gender-recording gaps (1,067,027 finishes); ages 20 to 75 with at least 100 finishes per age and gender are drawn; the three anchor ages pool two years either side. A cross-section of different finishes at each age, not people ageing. Finishes pooled across editions.

Finishes in start-screened editions (449,090 finishes) that reached 20K in 1:50:00–1:54:59, grouped by how their 0–5K pace compared with their own 5–20K pace. Finishes pooled across editions.

Finishes in start-screened editions of at least 1,000 finishes (3,969,514 finishes) that had another finish within 5 seconds of elapsed time at 5, 10, 15, 20, 25, 30, 35, 40K and the finish. Chip times only: nothing says the two ran together. Finishes pooled across editions.

Every eligible finish (4,108,749 finishes) against round finish times, pooled over all editions; the median over the 214 editions of 1,000 or more finishes keeps one quick field from carrying the share under 3:00. Finishes pooled across editions.

Every eligible finish (4,108,749 finishes) grouped by its record linked by name (3,240,150 records); the bars are the 178,201 records with three or more finishes and the finishes they hold. Links differ by race, so repeat counts are approximate. Finishes pooled across editions.

The 10th, 50th and 90th percentile finish of every edition with at least 2,000 finishes, in the 21 races with three or more such editions (3,921,134 finishes); the ratio of the 90th percentile's distance from the median to the 10th's, edition by edition: Boston is the median over its editions, and "elsewhere" the median over the 181 editions of the other races; the chart shows the eight races with the most finishes, ranked by the ratio of their median distances. The closest other edition (Madrid) sits near Boston's least lopsided one. Each edition counted equally.

Men's finishes with an exact recorded age (613,196 finishes at 25–79): the share of each five-year age band from 35 quicker than the median men's 25–34 finish pooled over the same editions. Finishes pooled across editions.

In each start-screened edition of at least 2,000 finishes (184 editions, 3,946,081 finishes), the running time of its quickest 1% of finishes (the 1st-percentile finish) against every finish's split at each mat; the median edition. This is running time, not a moment on the clock: waves start apart. Each edition counted equally.

Every finish in start-screened editions (3,970,364 finishes): which of its eight full 5K sections was quickest and which slowest. Finishes pooled across editions.

Each New York finish from 2016 to 2025, and whether its record linked by name holds a New York finish in the five calendar years before. Finishes pooled across editions.

Finishes with a recorded age group of 70–74 or older (7,552 finishes in the 48 editions that record age groups), by finish time. Finishes pooled across editions.

Every eligible finish (4,108,749 finishes), by how many eligible finishes in the same edition share its whole recorded second; pooled within each band of edition size (eligible finishes). Finishes pooled across editions.

193 editions that pass the start-offset screen with 100 or more finishes and a weather row, in races with three or more such editions: each edition's sustained slowdown share and its warming over the four hours after the scheduled start, both against the race's own average, with the within-race start-temperature slope removed; editions grouped by warming more than 2 °C below the race's usual, 0.5–2 below, within 0.5, 0.5–2 above and 2 or more above, each edition counted once. Weather is edition context, never a personal exposure. Each edition counted equally.

Consecutive finishes of records linked by name in race-date order (868,569 pairs of finishes); each finish’s fifth of its race’s complete finishes by finish time, as a share of the pairs starting in that fifth. Pairs of linked records, one pair per consecutive race.

Every eligible finish outside the 5 gender-recording-gap editions (3,845,060 finishes), pooled: the share of men's finishes, and of all finishes, under 4:00. The pooled median men's finish is 3:59:57 and the median over editions of each edition's men's median (210 editions with 100+ men's finishes) is 3:57:22. Finishes pooled across editions.

Finishes in start-screened editions (3,970,364 finishes) whose every full 5K section stayed within 3% of their own average 5K pace (the last 2.2 km left out), by finish time. Finishes pooled across editions.

The seven races with the most finishes among those with three or more editions of 1,000 or more finishes: each edition's 10th, 25th, 50th, 75th and 90th percentile finish times, then the median over the race's editions (3,054,403 finishes); the middle half is the span from the 25th to the 75th. Each edition counted equally.

Every pair of finishes in the same start-screened edition of at least 1,000 finishes (3,969,514 finishes) within 5 seconds of each other at 5, 10, 15 and 20K, grouped by the finish the earlier 20K split projects (20K time × 42.195/20); for the 2:45–3:15 group, the section in which each pair parted (first more than 5 seconds apart at the mat that ends it) or that it stayed within 5 seconds to the finish. Across groups, the share still within 5 seconds at every later mat rises with the projected finish. A pack of k finishes gives k(k−1)/2 pairs, so the same five groups were also counted per finish (finishes with such a partner that kept one to the finish) and per edition (the median edition of 100+ pairs): the order of the groups is the same. Pairs of linked records, one pair per consecutive race.

Men's finishes under 3:00 with an exact recorded age in the granular-age editions without gender-recording gaps (21,601 finishes), by five-year band of recorded age. Of all men's finishes with an exact recorded age (640,904 finishes), 55% were at 40 or older. Finishes pooled across editions.

Records linked by name with five or more finishes (37,154 records): their first five in race-date order, each finish time against the one before; a shuffled order would give one in 120 for either. Finishes pooled across editions.

Every sustained slowdown in start-screened editions (3,970,364 finishes, 1,217,483 slowdowns), by the fifth of its edition's clock order at the 20K mat and by the 5K mat where the slowdown began, as shares of each fifth's slowdowns (23% of front-fifth finishes and 34% of back-fifth finishes had one); the front fifth's share beginning at 35K was the higher in 181 of the 182 editions with 100 or more slowdowns in both fifths. Finishes pooled across editions.

Every pair of finishes in each start-screened edition of at least 1,000 finishes (3,969,514 finishes), by their gap in elapsed time at 20K or 30K, and whether the finish order was the reverse (ties at either point left out); the median edition (every pair pooled, 5–10 minutes at 20K: 15.6%). Each edition counted equally.

Edition by edition, the median finish of men's finishes recorded 50–54 and 55–59 against the women's recorded 30–34, in the 38 editions (6 races) outside those where recorded gender is incomplete with at least 100 finishes in each group; one dot per edition. Pooled over those editions' men's and women's finishes with a recorded age group (1,201,165 finishes), the men's 50–54 median is 18 minutes ahead of the women's 30–34 median. Each edition counted equally.

Consecutive finishes of one record linked by name (a candidate, not a verified person), in race-date order from the weather evidence: 5,352 doubles 1 to 21 days apart at two different races (mainly Washington then New York, Berlin then Chicago, Berlin then Amsterdam and Boston then London), against 13,964 pairs 301 to 430 days apart at the same two races in the same order, reweighted within each band to the doubles' mix of race pairs (race pairs with at least 10 such pairs in the band). Each pair is banded by the mean of its two finish times rather than the first, so no regression to the mean is built into either group. Doubles are noisier pairs: under 3:30, the two finishes differ by more than 10% in 18% of doubles against 13% of year-on pairs. The share of pairs whose second finish was quicker; ties count as not quicker. Pairs of linked records, one pair per consecutive race.

Finishes in start-screened editions (3,970,364 finishes): which of the eight full 5K sections was slowest, grouped as 0–30K, 30–35K and 35–40K, for three races and for every race pooled; each bar is the share whose slowest was 30–35K. Finishes pooled across editions.

Start-screened finishes (3,970,364 finishes) by their margin on 4:00 splits at 20K, 30K, 35K and 40K (4:00 × km/42.195 minus the split, in 30-second bins centred on whole half-minutes, so "on 4:00 splits" is within 15 seconds); the share of each bin that finished under 4:00. Finishes pooled across editions.

Finishes whose eight full 5 km sections all sit within 3% of their mean 5 km pace, in start-screened editions of 1,000+ screened finishes of eight races; each race’s mean share over its 2013–19 and its 2022–25 editions. Each edition counted equally.

In each of 183 editions with at least 500 recorded women's and 500 men's finishes, at most 2% other or unrecorded gender and at least 80% of their records kept as finishes (the editions with gender-recording gaps left out), the share of women's finishes quicker than that edition's median men's finish; the chart shows the median edition. Each edition counted equally.

Start-screened editions of at least 1,000 finishes (3,969,514 finishes, 199 editions): within each edition, every finish is ranked on each section's split on its own (ties at their average rank) and on its finish clock time, and counted when the two places differ by under 5% of the field; each bar is the median over editions (the cumulative 30K clock time, ranked the same way, reached 71%). The 20–25K or 25–30K split ranked closest in 199 of 199 editions. Each edition counted equally.

Consecutive finishes of records linked by name whose first finish was 2:55:00–3:09:59 (48,191 pairs on different days): the share whose next finish was under 3:00, by the minute of the first. Pairs of linked records, one pair per consecutive race.

Finishes with a 5–20K pace of 5:30–6:00 per km in editions that pass the start-offset screen, in Hamburg, Berlin, Chicago, London, Boston and New York (511,298 finishes in 79 editions with 100 or more each): the 20–25K pace against the 5–20K pace, the median within each edition, then the median over the race's editions. Each edition counted equally.

Finishes at 60–69 and 18–29 (exact recorded ages, start-screened editions, finishes from 2:30 to 6:00) matched on edition, recorded gender and finish minute, each cell weighted to its smaller group (matched weight 33,776); the gap is the difference of their weighted mean clock times at each mat. Pairs of linked records, one pair per consecutive race.

Races with at least six editions of 3,000+ finishes spanning eight or more years (11 races, 143 editions): the gap between the 10th and 90th percentile finish, averaged over the first three and the last three such editions (Amsterdam 2012–2014 against 2022–2025; Berlin 2005–2007 against 2024–2026; Boston 2014–2018 against 2024–2026; Chicago 2006–2009 against 2024–2026; Copenhagen 2014–2016 against 2024–2026; Hamburg 2013–2015 against 2023–2025; Madrid 2013–2015 against 2024–2026; New York 2006–2010 against 2023–2025; Rotterdam 2013–2015 against 2022–2024; Warsaw 2015–2017 against 2024–2026; Washington 2008–2010 against 2016–2018). Each edition counted equally.

Finishes in start-screened editions (3,545,576 finishes) in six 30-minute finish bands from 2:30 to 5:30; each cell is the median of a section's pace against the same finish's 5–20K pace. Finishes pooled across editions.

Consecutive finishes of records linked by name, both in editions that pass the start-offset screen (844,470 pairs on different days): whether the 20–40K took less clock time than the 0–20K, in the first finish and in the next. Pairs of linked records, one pair per consecutive race.

Finishes in start-screened editions (3,970,364 finishes) whose elapsed time at a mat falls in the whole-second window of 2:59 splits (a constant pace for 2:59:00 to 2:59:59, scaled to that mat) against the next window of the same number of seconds (3:00 splits). Finishes pooled across editions.

Each start-screened edition with at least 500 finishes, in races with three or more such editions (193 editions, 23 races, 3,813,745 finishes): the median of each 5K section's pace relative to the finish's own 5–20K pace, 5–10K to 35–40K; the nearest other edition by Euclidean distance over those seven values, against the share a random other edition would give. "The race" covers its course, field and weather together, and is grouped by race name. In 9 of the 23 races, half or more of the editions are nearest to the same race. Each edition counted equally.

Each edition with at least 2,000 finishes (199 editions, 4,084,466 finishes): the distance of its 1st, 10th, 25th, 75th, 90th and 99th percentile finishes from its own median, and the median of each distance over those editions; in 197 of the 199 editions the 99th percentile sat further from the median than the 1st. Course cut-offs bound the far tail. Each edition counted equally.

Women's share of men's and women's finishes under 3:30 in 108 editions of 1,000+ finishes in eight races (Berlin, Chicago, New York, Boston, Amsterdam, Rotterdam, Hamburg, Copenhagen), editions with gender-recording gaps left out. Each edition counted equally.

Finishes in start-screened editions (3,970,364 finishes) whose each 5K section from 20–25K to 35–40K was slower than the one before it; it includes finishes with and without a sustained slowdown. Finishes pooled across editions.

Men's finishes with an exact recorded age in start-screened editions (565,539 finishes), by ten-year age band and 30-minute finish band from 3:00 to 6:00 (three bands drawn); cells under 100 finishes are left out. Finishes pooled across editions.

Every finish in start-screened editions of at least 1,000 finishes (3,969,514 finishes): its clock place within the edition at each mat against its place at the finish, counted as near when the two differ by at most 5% of the field; rows are the five fifths of the finishing order. Finishes pooled across editions.

Finishes in editions that pass the start-offset screen (3,970,364 finishes), by finish minute: the share whose pace over 40K to the finish was quicker than over 35–40K. Finishes pooled across editions.

193 start-screened editions with 500+ finishes and a reviewed start temperature, from races with three or more (3,813,745 finishes): each edition's share of even-paced finishes (every full 5K within 3% of the mean 5K pace) and its start temperature minus its race's mean; editions grouped into thirds of the temperature deviation; each dot is one race with five or more such editions, its warmest-third mean minus its coolest-third mean; a gap under half a point counts as about the same. Weather is edition context, never a personal exposure. Each edition counted equally.

New York finishes with an exact recorded age, edition by edition: the shares recorded under 30 and at 60 or older. Each edition counted equally.

In each of 164 editions with at least 1,000 men's and 1,000 women's finishes (gender-recording gap editions left out), the women's finish at each percentile of the women's field over the men's finish at the same percentile of the men's field; the median of each gap across editions, every edition counting once. Each edition counted equally.

Finishes of 3:59:00–3:59:59 in start-screened editions (42,160 finishes), split by whether a sustained slowdown was recorded; lines are the median pace of each section against the finish's own average pace. Finishes pooled across editions.

Finishes in start-screened editions of at least 1,000 finishes (3,969,514 finishes), by their tenth of the edition's field at 20K (clock-time order): the 20–40K time against the 0–20K time, in half-percent bins, and places gained between 20K and 40K (in hundredths of the field); the first bin where the median places gained is no longer positive (all finishes: 12.5%). The median slowdown of all finishes was 9.4%. Finishes pooled across editions.

Finishes in start-screened editions of the eight races with the most such finishes (3,160,270 finishes, 102 editions): per edition, the share whose pace from 40K to the finish was quicker than their 35–40K pace; each race shows its lowest, median and highest edition. Each edition counted equally.

Consecutive finishes of records linked by name (834,550 pairs of finishes on different days, first finish under 6:00): the share whose second finish was within 2% of the first. Pairs of linked records, one pair per consecutive race.

Editions of 30,000 or more eligible finishes (56 editions, 2,403,986 finishes): finishes from five minutes before to five minutes after each round time, scaled to a field of 40,000; the median over editions, window by window. Each edition counted equally.

Pairs of finishes with exact recorded ages within 5 seconds of each other at all nine timing points (5K to 40K and the finish), in the 32 start-screened editions with exact recorded ages and at least 1,000 finishes (53,113 pairs; 323 pairs identical at all nine points left out), against every pair within 5 seconds at the finish but 30 seconds or more apart at some mat, reweighted to the in-step pairs' edition and 30-minute finish band. Pairs of finishes, not people. Pairs of linked records, one pair per consecutive race.

Finishes in start-screened editions (3,970,364 finishes): pace from 40K to the finish against the same finish's 35–40K pace and its 5–20K pace. Finishes pooled across editions.

Every pair of a recorded woman's and a recorded man's finish in each start-screened edition of at least 1,000 finishes with 100 of each and at most 2% of finishes without a recorded gender, editions with gender-recording gaps left out (194 editions, 3,705,825 finishes): the share of pairs with the woman ahead by elapsed time at each mat (a tie counts as half); the median edition. Places are clock-time orders, not road positions. Each edition counted equally.

193 start-screened editions with 500+ finishes and a reviewed weather row, from 22 races with three or more (3,813,745 finishes): each edition's sustained slowdown share, start wind and start temperature minus its race's mean, grouped into thirds of the wind and of the temperature deviation; each bar is a third's mean, editions equal. The windy third's 95% interval resamples races (2,000 draws). Weather is edition context, never a personal exposure. Each edition counted equally.

Pairs of finishes within 5 seconds of each other at all nine timing points in the start-screened editions (154,113 pairs of different linked records; pairs identical at all nine points left out), keyed by their two linked records (149,353 keys); every later start-screened edition that both records finished (14,625), and whether the two finishes were in step at all nine points again. The comparison column does the same for pairs of linked records within 5 seconds at the finish but 30 seconds or more apart at some mat (9,188,042 keys, 2,448,654 later shared editions, 432 in step). Records linked by name are candidates, not verified people. Pairs of linked records, one pair per consecutive race.

Men's finishes with an exact recorded age in the 28 editions holding at least 10,000 of them (626,833 finishes): the quickest 1%, 10% and 25% and the slowest 10% by finish time within each edition, ages pooled across the editions, against the median age of all of them. Finishes pooled across editions.

Every eligible finish (4,108,749 finishes) in ten-second steps of its whole recorded second around 3:00, pooled over all editions. Finishes pooled across editions.

Every finish in start-screened editions (3,970,364 finishes), by how far its 15–20K pace drifted from its 5–10K pace; the share with a sustained slowdown. Finishes pooled across editions.

Women's share of men's and women's New York finishes in each five-year band of exact recorded age, 2006 against 2025 (92,243 finishes; every band holds at least 189 finishes of each gender). Finishes pooled across editions.

Places are clock ranks within each of the 199 start-screened editions with at least 1,000 finishes (3,969,514 finishes), compared at 20K, 25K, 30K, 35K, 40K and the finish: the share that gained places in all five sections against the share that lost places in all five, as the median over editions. In every section after 20K the typical edition has most finishes gaining places (57–61%) while a minority lose many, so steady gainers can outnumber steady losers. Outside the editions where recorded gender is incomplete, men's finishes 27% against 17%, women's 40% against 7%. Each edition counted equally.

The sustained slowdown share of each edition of 100+ finishes that passes the start-offset screen, for the races with at least eight such editions (11 races, the 8 widest drawn); each row runs from the lowest edition to the highest, the dot at the median edition. Each edition counted equally.

Start temperature at the scheduled start from the reviewed weather rows, one value per edition (editions with at least 100 finishes), in the 10 races with eleven or more such editions: each race's warmest minus its coolest morning, as a multiple of the gap between the coolest and the warmest race's median morning. Races are grouped by race name, so a city that changed race counts as two. These are mostly spring and autumn races in temperate cities. Over the 17 races with five or more editions, differences between years of the same race hold 76% of the variance of start temperature. Each edition counted equally.

Consecutive finishes of the same record linked by name (not a verified person), ordered by race date, 300 to 429 days apart (434,299 pairs); same-day pairs are left out. A check on New York pairs one year apart whose exact ages step by 0 to 2 years gives the same picture. Pairs of linked records, one pair per consecutive race.

Every finish in start-screened editions of at least 1,000 finishes (3,969,514 finishes, 199 editions): the change in its clock place within the edition over each 5K, as a share of the field; the bars are the median over all finishes, shown against the 15–20K one, and each edition also compares its own median change in 30–35K with its median change in 15–20K. Finishes pooled across editions.

Women’s share of recorded men’s and women’s finishes in each race’s earliest and latest edition of 1,000+ finishes, races whose editions span ten years or more; editions with gender-recording gaps, over 2% unrecorded gender or under 80% of records retained are left out. The eight races with the most finishes are drawn; the earliest edition is the first that passes these rules, not the race’s first edition. Each edition counted equally.

Finishes with a 5–20K pace of 5:30–6:00 per km in editions that pass the start-offset screen, in the 24 races with a supplied route and 1,000 or more such finishes (723,108 finishes): each race's sustained slowdown share, its editions (100 or more such finishes each) weighted equally, against the whole-route climb of its supplied course file (r = 0.39, 95% interval -0.02 to 0.68). Each share folds in the weather of its editions; the climb is context, never an adjustment. Each edition counted equally.

Records linked by name (not verified people) with two or more finishes in start-screened editions (497,725 records with a gap over a minute): the gap between each record's quickest and slowest finish, split by section and pooled over every record, against each section's share of the distance (the median record opened 68% of its gap after 20K). Pairs of linked records, one pair per consecutive race.

Every eligible finish (4,108,749 finishes) counted by whole finish minute: across each ten-minute mark from 2:40 to 3:50 and from 4:10 to 5:50 other than the hours and half hours, the finishes in the minute before over the minute after, divided by the same ratio one and two minutes either side (geometric mean). Finishes pooled across editions.

Start-screened finishes with a recorded men's or women's gender, gender-recording gap editions left out (3,704,847 finishes); even pace is every full 5K within 3% of the mean 5K pace; women's and men's finishes are matched on edition and finish minute, each cell weighted by its smaller group (1,161,077 per group). Pairs of linked records, one pair per consecutive race.

Finishes in start-screened editions of the 22 races with three or more such editions of 300+ finishes (3,813,745 finishes, 193 editions): the share whose 25–30K took at most 98% of their 20–25K time (at least 2% quicker). Each race dot is its typical edition (the median over its editions) and the typical race is the median of those dots. New York's 15 editions range from 37% to 46%; pooled in half-hour finish bands from under 3:00 to 6:00+, its lowest band (34%) is above the other races' highest (19%). Counting any quicker 25–30K, the typical New York edition has 61% and the typical race 19%. 27,376 records linked by name with finishes in New York and in another race sped up by 2% or more in 42% of their New York finishes and 10% of their finishes elsewhere. Each edition counted equally.

Records linked by name whose first linked finish was in an edition with exact ages, and each later finish of the same record: the median change in finish time by whole years since the first and age at the first (points with 100+ pairs; only records still finishing). Records first seen at 60+ are not drawn; at eight years they lost more time than the 50–59 group. Pairs of linked records, one pair per consecutive race.

Every finish in start-screened editions (3,970,364 finishes): its slowest 5K section after 20K against its own 5–20K pace, in bins of 2.5 points; the first bin also holds every finish 7.5% quicker or more, and finishes beyond 80% slower are counted in a note. Finishes pooled across editions.

Women's share of recorded men's and women's finishes in each 15-minute finish band, in the 84 editions of at least 2,000 finishes without gender-recording gaps that hold at least 100 such finishes in every band (under 3:00, then 3:00 to 6:15); each band shows the median over those same editions, every edition counting once. Each edition counted equally.

Each start-screened edition of at least 2,000 finishes (184 editions, 3,946,081 finishes): the gap between its 10th and 90th percentile elapsed time at every mat, and the median over editions; the dashed line is eight times the median gap at 5K, the spread the field would reach by 40K if it kept spreading at its first-5K rate. Each edition counted equally.

Men's and women's finishes outside the 6 gender-recording-gap editions (3,843,228 finishes), counted by whole finish minute: at each mark, the finishes in the minute before over the minute after, within each gender. Finishes pooled across editions.

Records linked by name with five or more finishes in start-screened editions (36,147 records): the sustained-slowdown share at each of their first five finishes in race-date order, the same records in every column. Finishes pooled across editions.

Women's share of recorded men's and women's finishes under 4:00 in the latest edition, 2023 or later, of each race with at least five editions of 1,000 such finishes without gender-recording gaps and 10,000 finishes in that latest edition. Finishes pooled across editions.

The 17 races with at least five start-screened editions of 1,000+ finishes and a reviewed weather row: in the coolest and the warmest of them (temperature at the scheduled start, edition context only), the share of all pairs of finishes whose order at 20K is the reverse of their finishing order; a race counts as clearly higher on one morning when the gap is at least half a percentage point. The chart shows every race’s gap. Pairs of linked records, one pair per consecutive race.

Every finish in start-screened editions (3,970,364 finishes): the share whose 5–10K section was quicker than its 0–5K, by finish time (53% of all finishes); "slower" in the title means strictly slower, and equal sections count as neither. Finishes pooled across editions.

Finishes in start-screened editions of the 22 races with three or more screened editions of 100+ finishes (3,813,745 finishes, 193 editions): per edition, the share whose 20–40K took less time than their 0–20K; each race's mean over its editions. Each edition counted equally.

New York finishes (746,234 finishes, 16 editions) grouped by record linked by name; counts are records, which are candidate links rather than verified people. Finishes pooled across editions.