Does TAG predict
marathon performance
from parkrun — and from a half marathon?

In our earlier paper Does TAG predict half marathon performance from a seasonal parkrun PB? we showed that TAG predicts half marathon times to within ±5 minutes for 47% of club runners and within ±10 minutes for 80%. The formula is slightly conservative — competitive runners consistently beat their predictions — and shows no systematic gender bias. The key assumption underlying the prediction is that a runner's TAG% is conserved across distances: a runner performing at 80% of age-adjusted world class over 5km will also perform at approximately 80% of world class over 21.1km.

This paper tests whether that assumption holds at 42.2km. The marathon is a categorically harder prediction problem than the half marathon. Glycogen depletion, pacing strategy, marathon-specific training, and aerobic base all play a role that is largely invisible in a 5km parkrun effort. The question is not whether TAG predicts marathons perfectly — it clearly cannot — but whether the formula remains correctly calibrated, and how its accuracy scales with distance.

We also introduce a novel test: using an actual half marathon race time as the prediction input, rather than a parkrun PB. This eliminates the "casual parkrun" noise and tests the formula with two genuine race-day efforts. By matching runners who completed both Wilmslow Half Marathon (March 2026) and Chester Marathon (October 2025), we can compare predictions made from parkrun against predictions made from a half marathon — and see whether the input quality changes the output accuracy.

Before presenting the results, it is worth addressing how accuracy should be compared across distances. A ±5 minute error at half marathon distance represents about 4% of the typical finish time. The same absolute error at marathon distance represents only 2% — it is a proportionally smaller mistake. Comparing marathon prediction accuracy using the same absolute bands as half marathon prediction is therefore misleading.

Since a marathon is approximately twice the distance of a half marathon, a fair comparison doubles the tolerance bands. A runner predicted within ±5 minutes at half marathon is equivalent to a runner predicted within ±10 minutes at marathon. This paper reports both absolute and scaled figures, but the scaled comparison is the meaningful one.

The 2025 MBNA Chester Marathon took place on 5 October 2025 — a well-established race with a mildly undulating course through the city and surrounding countryside. Course details: 42.2km, 200m total ascent, 100% paved. At 4.74 m/km, the course falls into Skamper's band 0 (below 5 m/km), the same band as Wilmslow Half Marathon, with power coefficient 0.1 and an open-class time of 2:19:30.

The matching and filtering methodology is identical to the half marathon study: name-only matching with club as a tiebreaker, minimum 3 parkruns in 2025, anomaly filters for implausible parkrun times, and an extreme variance filter removing predictions more than 25 minutes from actual. The proportion of collisions removed was higher than at half marathon (343 of 903 predictions removed, 38%) — consistent with the longer gap between parkrun PBs and the race date, and a broader field of runners less likely to have raced parkrun recently.

The full 2025 parkrun dataset was used, including results after the October race date. This is appropriate for a seasonal PB analysis — the best 2025 parkrun performance reflects the runner's capability across the year, regardless of when it was set.

The overall median variance of +1.4 minutes means TAG is essentially unbiased as a marathon predictor — the median runner performs almost exactly at their parkrun-implied level. This is a stronger result than at half marathon distance, where the median was −2.0 minutes (runners beating predictions). The distribution is more symmetric at marathon distance, with 267 runners faster than predicted and 293 slower.

The absolute accuracy figures are lower than at half marathon: only 26.2% within ±5 minutes (vs 47.8% at HM) and 54.3% within ±10 minutes (vs 80.5% at HM). However, as discussed in Section 2, these are not the meaningful comparison. At scaled bands — doubling tolerances to match the doubled distance — the story changes significantly.

The distribution has longer tails than at half marathon distance, particularly on the slow side. The 26.6% of runners more than 10 minutes slower than predicted reflects the classic marathon challenge: runners who cannot sustain their 5km pace equivalent over the full distance due to fuelling, pacing, or insufficient marathon-specific preparation. This is a behavioural and physiological phenomenon, not a formula error.

Calibration check

An important diagnostic: adjusting the open-class time to minimise median variance shifts it by only 54 seconds (from 2:19:30 to 2:20:24). This adjustment changes the median variance to zero but has almost no effect on the spread — ±10 minute accuracy moves from 53.9% to 53.9%. The formula is correctly calibrated; the spread is inherent to the distance, not an artefact of a misspecified open-class time.

Gender breakdown

Male runners are essentially unbiased at median (−0.9 minutes), while female runners show a +5.4 minute positive median — running slower than their parkrun-implied time. The female ±10-minute accuracy is actually slightly better than male (58.0% vs 54.8%), suggesting the female median shift reflects a consistent moderate effect rather than extreme outliers. This gender difference is not seen at half marathon distance and merits further investigation with larger samples.

Of the 3,077 Wilmslow Half Marathon finishers (March 2026) and 4,913 Chester Marathon finishers (October 2025), 149 runners appear in both results by name — 124 male and 25 female. After age assignment from the parkrun dataset (or race category midpoint as fallback), computing TAG from the actual Wilmslow chip time, and applying the 25-minute extreme variance filter, 89 runners form the core analysis set.

This is a fundamentally different test from the parkrun analysis. Both inputs are race-day efforts — the Wilmslow time reflects genuine maximum effort on a measured course, not a training parkrun. If the TAG conservation assumption holds, these predictions should be the most accurate of any we produce.

The median variance of +5.8 minutes means the typical runner performed worse at Chester than their Wilmslow time implied. This positive bias — larger than in the parkrun analysis (+1.4 minutes) — reflects several factors: the seven-month gap between the two races means Chester fitness was not necessarily equal to Wilmslow fitness; runners may have targeted one race more seriously than the other; and the inherent difficulty of sustaining half marathon pace proportion over a full marathon.

However, the ±20-minute accuracy matches the parkrun analysis exactly: 91.0% in both cases. Using a race-quality half marathon time as input rather than a parkrun PB does not improve the broad accuracy of the prediction, though it does narrow the tails slightly.

Fastest individual predictions

Among the 89 runners, the closest predictions include Emily Farrimond (predicted 3:29:29, actual 3:29:34 — a 5-second error), Thierry Ngoga (16 seconds), John Williams (41 seconds — also one of the closest predictions at Wilmslow itself), and Ciaran Wright (1:04). At the elite end, Joshua Griffiths — winner of the Wilmslow half marathon in 1:05:04 — was predicted 2:19:28 at Chester and ran 2:17:16, just 2 minutes faster than the Skamper prediction. The formula correctly identified him as a 2:19 calibre runner.

The key question is whether TAG accuracy degrades at marathon distance, or whether the apparent degradation is an artefact of using the same absolute tolerance bands at twice the distance. The scaled comparison answers this directly.

The secondary band (±10 mins at HM, ±20 mins at marathon) captures 80.5%, 91.0%, and 91.0% of runners respectively. The marathon predictions are actually slightly more accurate at the secondary band than the half marathon, likely because the larger absolute tolerance absorbs more of the individual variation. The primary band (±5 mins at HM, ±10 mins at marathon) shows more variation: 47.8% at HM vs 51.7–53.9% at marathon. Again, the marathon performs comparably or slightly better.

This is a strong result. It means TAG is not a formula that works at 5km and half marathon but breaks down at marathon. Rather, the uncertainty in marathon prediction is proportional to distance — the same fraction of runners fall within proportionally scaled bands. The formula correctly captures the physics of pace scaling across distances; what it cannot capture is the human variation in marathon performance relative to shorter-distance performance.

Among the top 10% of male finishers at Chester (n=40), the median variance is −13.5 minutes — these runners beat their parkrun-implied time by over 13 minutes on average. All but 3 of the 40 finished faster than predicted. This is an even stronger overperformance relative to prediction than at half marathon distance (where the male top 10% beat predictions by a median of 6.7 minutes).

This is not a formula failure — it reflects genuine elite performance. Competitive marathon runners often run at a higher fraction of their aerobic ceiling than their parkrun efforts suggest, because marathon training builds the specific endurance that short-distance running does not test. A runner who jogs a parkrun at 80% effort but races a marathon at 95% will consistently beat their TAG prediction at marathon distance.

For the female top 10% (n=15), the median variance is −6.2 minutes — also strongly negative, with 13 of 15 finishing faster than predicted. This is consistent with the half marathon finding: among competitive runners genuinely racing to their potential, TAG is conservative rather than optimistic.

1. TAG is correctly calibrated at marathon distance. The open-class time requires only a 54-second adjustment to achieve a zero median — well within the margin of a single dataset. The formula is not systematically wrong for marathons.
2. At scaled tolerance bands, marathon accuracy matches half marathon accuracy. The ±20-minute band captures 91% of marathon runners — comparable to the 80.5% captured by the ±10-minute band at half marathon. The degradation in absolute accuracy at marathon distance is proportional to distance, not a model failure.
3. Using a half marathon time as input does not significantly improve broad accuracy. Both the parkrun and half marathon inputs produce 91% within ±20 minutes. The input quality matters less than the inherent variability of marathon performance itself.
4. For elite runners, the formula is remarkably precise. Runners who genuinely race to their ceiling at both distances — where TAG% is truly conserved — are predicted within 1–3 minutes even at marathon distance. This validates the conservation assumption for consistent performers.
5. The systematic positive bias for slower runners reflects marathon-specific factors. Fuelling, pacing, marathon-specific endurance, and the training gap between parkrun and marathon all contribute to slower-than-predicted performances. These are characteristics of the runners, not errors in the formula.

Sample size

The HM→marathon matched set (n=89) is small, with only 19 female runners. Results at this scale should be treated as indicative rather than definitive. The finding that both methods produce 91.0% within ±20 minutes is striking but based on a coincidence of numbers that warrants replication with larger samples.

Single race and single year

Chester 2025 is one race on one day. Results may differ at hillier courses, in adverse weather, or with fields of different competitive character. The gap between Wilmslow (March 2026) and Chester (October 2025) means the matched runners ran their marathon before their half marathon — an unusual sequence that may introduce selection effects.

Collision rate

The high collision removal rate (38% of matched predictions) at marathon distance suggests name matching is less reliable for marathon events, possibly because marathon fields are more diverse and less concentrated in club running. An athlete ID system would substantially improve the quality of this analysis.

Female marathon variance

The +5.4 minute female median in the parkrun→marathon analysis (versus −0.9 minutes for males) is a notable asymmetry not seen at half marathon distance. With only 157 female marathon runners in the core set, this cannot be conclusively attributed to a formula effect, a training effect, or a selection effect. Further work with larger female marathon samples is needed.