Finish Line Forecast

Before the gun, a predicted finish time and placing for every registered runner in a field of thousands, from their public race history, published in advance, with the error published once the results are in. For a race director that is pacing, corral and medical staffing planned from expected finish times rather than guesses; for a runner it is a goal time with an honest interval instead of a hunch.

Status: building. Eighteen years of Newfoundland road results are read, 23,830 runners resolved out of them, the three baselines are measured on every race since 2024, every course's difficulty is measured from the results, and the hierarchical model now beats the strongest of those baselines by 27% for the runners with four or more prior results, 5.4 minutes of mean absolute error against carry-forward's 7.4, with no bias left to speak of. A LightGBM challenger given the same information is slightly more accurate still for every runner with a history, and that is published beside it rather than hidden. No prediction has been made yet. The first live race is the Cape to Cabot 20 km in St. John's on 2026-10-18, with a second on a frozen model on 2026-11-11; predictions are committed, tagged and hashed in this repository before each race and scored against the official results after it. Build plan: PLAN.md.

The website: finishline.peterparker.ca, every live race and its predictions, with the whole method explained in pictures.

Nothing was fetched until the two organisations whose pages this reads had been told, and finishline crawl refuses until they have. That rail is in code rather than in a document: docs/emails.md carries what was sent.

Result

What the archive holds. Measured by finishline catalogue on 2026-09-12 from the year indexes, which carry event names and dates and no runners.

Individual road races, 2008 to 2026	286
Distinct courses	52
Deepest course history	17 editions (Mews Memorial 8 km, Mundy Pond 5 km), 16 (Cape to Cabot 20 km)
Years with no racing	2020

The skip list is the coverage claim, so it distinguishes a duplicate from a hole. Most skips are team standings, awards pages, relays, cross-country and school races, which are other views of races already read or other disciplines; a handful are real holes, mostly races published as a PDF. One is the same race published twice: the 2014 CHCM 10 km is on the index as both .htm and .php with the same 162 finishers in different letter cases, and counting it twice would have given 162 people a second result on a day they raced once.

⚠️ Eighty of those courses were really fifty-two. The 2008 to 2015 index titles a race with its ordinal and whichever sponsor held the naming rights, so Burton's Pond read as six courses of one edition each and CHCM as seven, and every fragment then fell under the thirty-finish floor and vanished from the table below. The merge overshot once before it settled: some races have no name but their sponsor, and stripping it collapsed six unrelated half marathons into a single course called "unknown". docs/data-terms.md has the full table and every page that would not parse.

What reading all of it produced. Written by finishline report; not edited by hand.

Races read	286
Finishes parsed	75,061
Runners resolved	23,830
Runners this refuses to tell apart, and will not publish	381
Runners with one finish	12,737
Runners with two or three	6,280
Runners with four or more	4,786
Pages that would not parse	5

How hard each course is, measured rather than surveyed. Runners cross between courses, so a course's difficulty is identifiable from finishes alone: somebody slower on Cape to Cabot than their own equal-VDOT expectation every year, and faster on Mews Memorial every year, is saying what the hills cost. Effects are centred on the average course somebody actually runs, so zero is ordinary rather than flat, and the interval resamples runners rather than finishes, because two races by one person are not independent evidence about a hill. The last column is a check and not an input: it asks what average grade would explain the measured factor, given the climb the race publishes.

Course	Finishes	Editions	Slower than flat	95% CI	Grade that would explain it
usr-42195	32	1	+11.2%	[+7.5, +14.8]
huffin-puffin-42195	296	7	+9.6%	[+8.3, +10.8]
provincial-championship-42195	68	2	+9.4%	[+6.2, +12.1]
cape-to-cabot-20000	5,311	15	+9.2%	[+9.0, +9.5]	10.3% average, over the published 550 m of climb
run-from-away-42195	79	1	+7.8%	[+6.1, +9.6]
trapline-5000	138	6	+6.3%	[-1.9, +15.1]
bell-island-blast-16093	181	3	+4.6%	[+3.8, +5.5]
trapline-42195	55	7	+4.5%	[+0.5, +8.2]
... 34 more
five-and-dime-5000	1,052	10	-3.0%	[-3.6, -2.4]
quidi-vidi-5000	334	4	-3.2%	[-3.9, -2.4]
ane-mile-1609	577	12	-3.7%	[-5.0, -2.3]
provincial-championship-5000	907	8	-4.0%	[-4.5, -3.5]
turkey-tea-10000	2,296	14	-5.1%	[-5.4, -4.9]	the published climb cannot explain it
mews-memorial-8000	4,893	16	-5.3%	[-5.5, -5.1]
pearlgate-5000	88	1	-5.7%	[-6.8, -4.7]
oceanview-5000	108	1	-6.3%	[-7.2, -5.5]

The check is worth more than either number alone. Cape to Cabot is the only course here with a published elevation, and the two routes to its difficulty were computed independently: 5,310 finishes say +9.3%, and 550 m of climb against 450 m of drop over 20 km through Minetti's cost-of-running curve needs a 10.3% average grade to produce that. The race's own course page says "grades of more than 10 per cent in some parts". Neither number was tuned to the other.

⚠️ A per-runner career trend is doing more work here than it looks. Fitted with one constant per runner, Cape to Cabot's edition effect climbs almost monotonically from +3.8% in 2013 to +14.4% in 2025, which reads as a course getting harder every year. It is not: a career-long constant has nowhere to put the fact that runners age, so the edition effects absorb it, and all thirteen well-covered courses drift upward at a median of +0.60% a year. With a per-runner trend the median drift is +0.00% and the signs scatter. A model fitted the first way and asked for 2026 would extrapolate ten points of course inflation that does not exist, and the table would look entirely reasonable.

What the morning costs. The course layer says how hard a road is; this says how much of what is left over is the weather. It fits the same edition effects against the observed temperature and wind at St. John's airport, with the temperature coefficient allowed to grow with distance, because a 5 km field meets fifteen minutes of weather and a marathon field meets four hours of it.

Fitted on 231 editions near St. John's airport, 55 of 286 excluded as too far from it or without an observation. Explains 34% of the edition-to-edition variance within a course, leaving sd 2.31%.

Race length	Cost per degree above neutral	What a 20 C morning costs
5 km	-0.05%	-0.5%
10 km	+0.24%	+2.4%
Tely 10	+0.45%	+4.5%
Cape to Cabot 20 km	+0.54%	+5.4%
marathon	+0.86%	+8.6%

Neutral is 10 C and 20 km/h, which is the middle of this archive rather than a laboratory ideal.

Term	Estimate	95% CI
Temperature at 10 km, per degree	+0.245%	[+0.135, +0.372]
Tailwind along the bearing, per km/h	-0.033%	[-0.101, +0.025]

The check. The Tely 10 on its own is the cleanest natural experiment in the archive: eleven editions of two to four thousand finishers, run anywhere from 3.6 to 22.7 C because two COVID years pushed it into October. It gives +0.41% per degree on its own, +0.42 controlling for year, R-squared 0.64. The pooled model above, fitted across every course and never told about the Tely, returns +0.425 at that distance. Both sit in the range the marathon literature reports for mid-pack runners.

⚠️ This took two wrong answers first, and both are in PLAN.md section 13. Fitted unweighted, temperature came out at +0.03% per degree with the interval through zero, which reads as "the weather does not move a race in a climate this cool"; an edition effect from thirty finishers is mostly noise and the small races were shouting down the large ones. Fitted on raw edition effects rather than within-course ones, it partly measured the fact that the hard courses here run in October and the easy ones in June. What caught the second one was the residual coming out at 4.42%, larger than the 2.79% scatter it was supposed to be explaining: a model cannot explain something and leave more behind than it started with.

⚠️ A wind speed is not a wind, and the tailwind term is not yet significant. The prevailing wind in Tely season is westerly and the Tely runs east-north-east, so the usual wind pushes that field along; Cape to Cabot runs north-west, so the same westerly is a headwind, and it runs into one in 13 of its 16 editions. Fitted as a single speed for the whole province those cancel, which is exactly what the first attempt showed. Wind now enters as a speed, which a loop pays whichever way it blows, and as a signed tailwind along a course bearing, which only a point-to-point course has. Only two courses carry a bearing so far, so the tailwind coefficient has the right sign and an interval that still includes zero. It is reported that way rather than kept quiet because the sign is pleasing.

How well the obvious approaches do. Every race from 2024 on, each predicted only from results dated strictly before it. Coverage sits beside error in every row, because a model that answers for the easy half of a field is not better than one that answers for all of it.

Prior results	Runners	Model	Answered	MAE, minutes (95% CI)	Mean % error	Skill vs carry-forward
0	5711	carry-forward	0%	-	-	baseline
		best-equal-vdot	0%	-	-	-
		category-median	95%	18.4 (17.9 to 18.9)	18%	-
		hierarchical	100%	18.3 (17.9 to 18.8)	18%	-
		lightgbm	100%	18.6 (18.1 to 19.1)	17%	-
1	2713	carry-forward	100%	9.6 (9.2 to 10.0)	10%	baseline
		best-equal-vdot	64%	7.5 (7.1 to 8.0)	8%	21%
		category-median	96%	16.1 (15.5 to 16.7)	16%	-68%
		hierarchical	100%	9.4 (9.0 to 9.9)	10%	2%
		lightgbm	100%	8.8 (8.4 to 9.2)	9%	8%
2 to 3	2910	carry-forward	100%	9.3 (8.9 to 9.7)	9%	baseline
		best-equal-vdot	73%	7.8 (7.4 to 8.2)	8%	16%
		category-median	95%	15.8 (15.2 to 16.4)	16%	-70%
		hierarchical	100%	8.5 (8.2 to 8.9)	9%	8%
		lightgbm	100%	7.7 (7.4 to 8.1)	8%	17%
4 or more	7490	carry-forward	100%	7.4 (7.2 to 7.6)	8%	baseline
		best-equal-vdot	87%	7.2 (7.0 to 7.4)	7%	2%
		category-median	95%	14.0 (13.7 to 14.4)	18%	-90%
		hierarchical	100%	5.4 (5.2 to 5.6)	6%	27%
		lightgbm	100%	5.2 (5.0 to 5.3)	6%	30%

lightgbm against hierarchical on the same runners. The difference is in mean absolute error as a percent of each runner's own finish time; negative favours lightgbm, and the 95% CI resamples races.

Prior results	Runners	Races	MAE, minutes, lightgbm vs hierarchical	Difference, points of a finish time (95% CI)
0	5,703	53	18.6 vs 18.3	+0.67 (-0.85 to +1.60)
1	2,707	52	8.8 vs 9.4	-0.91 (-1.65 to -0.57)
2 to 3	2,904	52	7.7 vs 8.5	-1.00 (-1.39 to -0.79)
4 or more	7,480	51	5.2 vs 5.4	-0.39 (-0.59 to -0.21)

⚠️ The LightGBM challenger is more accurate than the hierarchical model for every runner with a history. Gradient-boosted quantile trees on hand-built features (models/gbm.py: form on the same Daniels scale, history depth and age, course difficulty, the raw weather), refitted per quarter on the same history, with hyperparameters fixed before the first run. On the same runners, the paired table above: 0.39 to 1.00 points of a finish time better for runners with one or more prior results, intervals clear of zero; level for first-timers. It also orders a field better, 4.2 places closer than carry-forward against the hierarchical model's 1.8. Its own quantile ranges under-cover at every depth (68 to 74% at 80%) and the conformal layer repairs them, which is the coverage table below. Its features and parameters were searched on 2022 and 2023 only, never on these races, and that search bought about one percent (PLAN.md section 13 item 34, which also has the six ideas that made it worse). The published predictions are still the hierarchical model's, because the placing simulation needs joint draws of a whole field on one morning, which quantile trees do not give; which model or blend publishes Cape to Cabot is decided before its model lock on 2026-10-11. PLAN.md section 13 item 33 has the detail.

⚠️ The first run of this model lost to carry-forward. This is the second run, and what changed is in PLAN.md section 13 items 28 and 29. Two things were wrong at once: each runner's improvement trend was carried in a straight line to race day, which predicts years of improvement nobody has, and the race-edition effects were quietly absorbing a calendar drift, so that 2023 and 2024 races came out 6 to 8% slower than their own course averages. Form is now a random walk over the years a runner actually races, and a shared year effect walks the whole province from one calendar year to the next. The bias is gone: on the 12,714 runners both can answer for, predictions are 0.7% too fast on average (95% CI 1.8% too fast to 0.6% too slow), against carry-forward's 0.4% too fast (2.6% too fast to 1.9% too slow), the intervals resampling races rather than runners. With each race's typical error removed the model is ahead rather than level, by 0.59 percentage points of absolute log error (95% CI 0.36 to 0.93), so it has both the level and the order of a field better than the baseline it lost to before.

⚠️ It still samples badly, and the tables above are what that badly-sampled model predicts. Across the eight quarterly fits the worst R-hat runs from 1.41 to 2.15 and the smallest bulk ESS is about 5, with no divergences. The cause is identification, not tuning: years since a runner's first race and the calendar year move together, so the group drift, the year effect and the group means trade off along a ridge that the sampler wanders. Checked on the worst fit, the quantity a prediction for a returning runner reads converges (R-hat 1.006 median, 1.08 worst, over 600 runners), so those predictions are not sampler noise; a first-timer's starting level does not (R-hat up to 1.27), so theirs carry some, and the conformal layer calibrating first-timers separately is what keeps their intervals honest. The obvious fix, constraining the year effect, made every chain disagree and was taken out. No individual coefficient from this fit should be read on its own. PLAN.md section 13 items 29 and 31 have the detail.

⚠️ Weather is the felt heat above 12 C, with the sun estimated, and it is worth about a percent of a field's level, not of a runner's error. The model reads each morning at St. John's airport over the hours the field was actually out, from each race's published start time, and charges nothing below 12 C. Above it, each degree of felt heat costs a fraction that grows with distance, and full sun adds degrees to the felt temperature, how many being a parameter the data estimates. Fitted on the whole archive, a degree above the knee costs +0.05% at 5 km (95% CI +0.00 to +0.13), +0.31% at 10 km (+0.23 to +0.39), +0.50% on the Tely (+0.37 to +0.61), +0.58% on Cape to Cabot (+0.43 to +0.71) and +0.87% at a marathon (+0.64 to +1.06); a km/h of wind +0.027% (+0.011 to +0.045). A full sun adds 2.1 C of felt temperature (95% CI 0.1 to 9.2), which the data narrowed to half its prior's spread without pinning down, because only six of 281 mornings in the archive had strong sun. The tailwind term stays the null it has always been, because only two courses carry a bearing.

Paired on the 18,278 predictions both runs made before the 2026 USR results were added (the ablation is rerun overnight, and its rows are missing from the tables until then), weather lowers absolute log error by 0.0010 (95% CI -0.0019 to +0.0002), better at every depth and clear of zero only for runners with no history. That is the size it should be: weather moves the level of a whole field by one to five percent, individual error is around ten, and a correct level shift is close to invisible per runner. It is not invisible to a race director planning a finish-line clock, and the check that matters is the race level. There, the model without weather leaves bias that climbs with the heat, +0.42 (+/- 0.30) of each point of heat cost left in the errors, and the model with weather leaves -0.15 (+/- 0.30), which is zero. The previous weather model, linear in temperature from 10 C, leaves +0.11 (+/- 0.31) on the same measure, so on three years of backtest races the two weather models cannot be told apart; the case for the felt-heat shape is eighteen years of editions, in docs/rejected.md. The 2025 USR, the hottest mornings in the backtest at 22 C, went from 1.0 to 1.4% too fast to 0.6 to 1.2% too slow.

⚠️ One hot race the heat does not explain: the 2026 Tely 10 is still 3.0% too fast. 18 C, light sun, calm, with a tailwind. The model charges it 3.1% and it ran about 6% slower than a neutral morning. The whole-archive fit is the source of the charged figure, not the backtest's own quarterly fit, so this is approximate; either way the gap is not heat, since no setting of the knee or the sun closes it without breaking the other hot races. The conditions table above is the older linear check on edition effects alone, kept as the independent cross-check it was built to be; PLAN.md section 13 item 30 has the felt-heat model, the alternatives it was measured against, and why the knee is fixed at 12 C rather than fitted.

Getting the order right, which is the number a race director actually plans from.

Model	Races	Mean absolute place error	Spearman, predicted vs actual
carry-forward	53	24.6	0.836
best-equal-vdot	52	17.6	0.838
category-median	46	95.4	0.349
hierarchical	53	50.9	0.706
lightgbm	53	49.6	0.733

The same runners: each model against carry-forward, both ranked among the runners both answered for in each race (races with at least 10 of them). Negative place error and positive Spearman differences favour the model; the 95% CI resamples races.

Model	Races	Runners	Place error, model vs carry-forward	Difference (95% CI)	Spearman, model vs carry-forward	Difference (95% CI)
best-equal-vdot	51	10,377	17.9 vs 19.0	-1.1 (-2.4 to -0.1)	0.849 vs 0.847	+0.002 (-0.006 to +0.010)
category-median	44	12,427	68.9 vs 27.9	+41.1 (+20.3 to +66.9)	0.368 vs 0.852	-0.484 (-0.536 to -0.439)
hierarchical	51	13,081	23.7 vs 25.5	-1.8 (-3.1 to -0.7)	0.857 vs 0.849	+0.008 (-0.005 to +0.023)
lightgbm	51	13,081	21.3 vs 25.5	-4.2 (-6.7 to -2.0)	0.875 vs 0.849	+0.027 (+0.016 to +0.037)

⚠️ Read the second table, not the first. The first ranks each model among the runners it answered for, so carry-forward is ranked over the 13,113 runners who have a prior result and the model over the whole field, the 5,711 entrants with no history included, whose order is close to a guess. That is where its 50.9 comes from. The second ranks both over the same runners, race by race: there the model is 1.8 places better than carry-forward (95% CI 0.7 to 3.1), 23.7 against 25.5, with Spearman 0.857 against 0.849. It orders a field better than the baseline, by less than it times one. Both tables stay, because scoring each model only on the subset that suits it is how a table stops being checkable.

How often the intervals hold. A predicted time with an interval is two claims, and the second one is checked here: the model's own 80% and 90% intervals, and the same intervals after conformal adjustment on the races before each one, by how much history a runner has.

hierarchical, every race from 2024 on. Each race's intervals are adjusted using only races dated before it, separately for each history depth. Coverage is the share of runners whose finish fell inside; the 95% CI resamples races, not runners, because runners in one race share its morning.

Prior results	Level	Runners checked	Races	Model's own interval	After conformal	Median width, minutes (own to conformal)
0	80%	5,649	51	77% (75 to 79)	78% (73 to 82)	54.4 to 55.2
1	80%	2,654	47	73% (70 to 75)	77% (73 to 79)	20.9 to 23.3
2 to 3	80%	2,841	47	71% (68 to 74)	78% (76 to 82)	18.4 to 22.8
4 or more	80%	7,390	50	77% (74 to 79)	77% (75 to 79)	13.4 to 13.6
0	90%	5,649	51	87% (86 to 89)	89% (86 to 91)	70.8 to 74.2
1	90%	2,654	47	84% (82 to 85)	89% (87 to 90)	28.8 to 34.4
2 to 3	90%	2,841	47	82% (80 to 84)	88% (86 to 89)	24.9 to 29.8
4 or more	90%	7,390	50	87% (85 to 88)	88% (87 to 90)	18.2 to 19.0

lightgbm, every race from 2024 on. Each race's intervals are adjusted using only races dated before it, separately for each history depth. Coverage is the share of runners whose finish fell inside; the 95% CI resamples races, not runners, because runners in one race share its morning.

Prior results	Level	Runners checked	Races	Model's own interval	After conformal	Median width, minutes (own to conformal)
0	80%	5,649	51	74% (70 to 77)	80% (76 to 84)	52.1 to 59.2
1	80%	2,654	47	71% (67 to 77)	76% (72 to 81)	20.6 to 22.6
2 to 3	80%	2,841	47	68% (65 to 73)	77% (74 to 82)	17.3 to 20.6
4 or more	80%	7,390	50	70% (66 to 75)	77% (72 to 82)	11.2 to 12.9
0	90%	5,649	51	85% (83 to 88)	89% (86 to 92)	68.3 to 74.8
1	90%	2,654	47	83% (80 to 86)	87% (85 to 90)	27.9 to 30.8
2 to 3	90%	2,841	47	81% (78 to 86)	87% (85 to 91)	24.2 to 28.7
4 or more	90%	7,390	50	82% (79 to 85)	88% (85 to 91)	15.8 to 18.3

The assumption. Conformal coverage is guaranteed on average over races within a history-depth group, provided a new race's errors look like the earlier races' errors (exchangeability). It is not a promise about any one runner or any one race, and it fails when a race meets conditions or a field the earlier races did not: a gale on Signal Hill is exactly that. The first races of the backtest have too few earlier errors to calibrate on (under 50 per group) and are left out of this table rather than given an interval nobody could trust.

Live: predicted before the gun, scored after. Cape to Cabot 20 km on 2026-10-18, then Run to Remember 11 km on 2026-11-11 on the same frozen model. finishline score reads each prediction from its tag, refuses one tagged less than 24 hours before the gun, and writes a row here and a race page under docs/predictions/ with every finisher's prediction beside their result.

No prediction has been scored yet. finishline score <race> fills a row here once a tagged prediction's official results are posted.

Who is entered for the first live race. Cape to Cabot, from the club's published start list of 2026-09-12, matched against the archive.

	2016 on, no 2026 Tely	2008 on, with it
Entrants on the list	453	453
Resolve to a runner in the archive	386 (85%)	419 (92%)
Have a 2026 result, so current-season form	166 (37%)	332 (73%)
Have four or more prior results	252	289
No history at all	67	34

That second row is why one race mattered more than the other 282. The 2026 Tely 10 is not on the association's own site: it timed the race on Race Roster in June and its Tely page links out, where every edition from 2018 to 2025 is published in place. It is 4,147 finishers, the largest field in the province, and adding it doubles the share of this field whose current-season form the prediction can see. The basis for reading it, which is the association's ownership of the race and its knowledge of this project rather than the platform's terms, is set out in docs/data-terms.md.

Course and conditions. Week 2 fills this; the normalisation is not built yet.

Course	Estimated course factor (95% CI)	Physics prior	MAE without normalisation	MAE with normalisation
not yet

What the baselines already say

A third of the runners in any race have never raced here before. 5,594 of the 18,308 runners predicted across those 49 races had no prior result at all, even with eighteen years of archive behind them, and for those runners the only answer any of these models can give is the middle of their age and sex category, which is out by 18.5 minutes on average. That number, not the model comparison, is the size of the real problem. Adding eight more years of history moved it by less than three hundred runners, which is worth knowing: the cold start is not a gap in the archive, it is people who have genuinely never raced here.

The race calculator beats the last result, but only where it will answer. Reading the best recent form off Daniels' curve is 19 percent better than carrying the last race forward for runners with one prior result, and it gains four percent on runners with four or more. What moves is coverage: it answers for 63 percent of the thin histories and 90 percent of the deep ones, because a runner with more results is likelier to have one inside the range the curve is fitted for. A calculator that quietly extrapolated instead would have posted a better-looking average over a worse-defined group.

Nobody has said anything about an interval yet. These three produce a time and no sense of how sure it is, which is the gap the hierarchical model and the conformal layer exist to fill.

What this does not do

• It uses no training data. The original idea included each runner's public Strava activity. Strava's API agreement (effective 2026-06-01) forbids displaying or disclosing other users' data even when public, forbids training models on API data, and its acceptable use policy forbids scraping. The public model here works from public race results only, which is weaker, and this page says so first. An opt-in channel, where a runner sees a prediction from their own data and nobody else does, is deferred.
• It does not know who will start. Where the race publishes an entrant list, as Athletics NorthEAST does for Cape to Cabot, it predicts for that list and publishes the no-show rate afterwards. Where there is no list, it predicts for the runners its history says are likely to run and publishes how many of the actual finishers it covered.
• It will sometimes merge two runners who share a name. Hometown is not allowed to split them, for the reasons above, so two people of the same name and a compatible age become one runner with one muddled history. The number that bounds it is published: a measured 2.7 percent of resolved runners have a printed hometown that changes back and forth rather than once, which is the shape two merged people make. Reading them shows most are one person spelling their own town differently across entry forms (Paradise and Pradise, Conception Bay South and Cbs), so the true number is lower than 416, and 416 is what gets published because it is the one that can be checked.
• It does not predict a runner it cannot tell apart from another runner of the same name. 355 were held back on the current archive, and they are counted rather than guessed at.
• It does not publish anything about a runner beyond what the race results already publish: name and hometown as printed, and the prediction.
• Its intervals are calibrated on past races. Coverage is guaranteed on average within a history-depth group under exchangeability, not for any one runner or one race, and the tables above say where it held.

How it works

See PLAN.md. Public road-race results from the Newfoundland and Labrador Athletics Association are crawled once, parsed and resolved to runners across races. There is no runner identifier anywhere in that archive, and most of the readable races print no hometown either, so the resolver works from the name and from the one piece of evidence the pages give away for free: a runner cannot get younger. Every printed age band on a dated race implies a window of birth years, and one person's windows have to intersect. That is the only thing allowed to split a name into two runners. The hometown breaks a tie and never splits, because runners move: of the 1,656 names appearing under two or more towns, 1,103 show a single clean switch over time. Where a result could belong to either of two runners and the page printed no age band, it is held back and counted. Each finish is put on one scale as a ratio to a Daniels reference time. A Bayesian hierarchical model on that ratio gives every runner a fitness level and a distance fade shrunk toward their age-sex group, and a form that walks from one racing year to the next; every race gets its course's measured difficulty, a shared effect for its calendar year, and the heat and wind observed at St. John's airport that morning. A live prediction walks each runner's form forward to race day and draws the weather from the forecast, corrected by how wrong that forecast was on past race mornings at the same lead. The design history, including the models the data refuted, is PLAN.md section 13; docs/methods.md is the short version. Intervals are conformalised on rolling-origin residuals, stratified by how many results a runner has. Placing is simulated from the whole field's predictive distributions, and at the biggest races the runners with no results here are drawn from how that course's past first-timers finished. From a week before a race, each registered runner is predicted the first morning they appear on the entrant list; the day before, the whole field is predicted again with the latest forecast and given places. Every file is committed, tagged and hashed before the gun and scored after, and the website (finishline site, rebuilt on every push) shows them with a search box for anyone looking for their own name.

Part of a portfolio

One of fifteen projects. It reuses the running arithmetic from Overload, the AI coaching team for runners: Daniels' VDOT, the heat and wind corrections and the age-grading tables, here in a public repository with tests.

How this was built

Design, methodology, evaluation choices and judgement are Peter Parker's, including years of racing and coaching himself on these courses. AI coding assistants (Claude Code) were used for implementation and drafting, the way a senior engineer uses them in 2026. Every number in the results tables is reproducible from this repository with one command, and every live prediction is verifiable from a tag that predates the race it predicts.

This page is the repository's README, rendered as of 2026-09-20. Relative links open on GitHub.