TAG as a Half Marathon Predictor

The Terrain-Adjusted Grade (TAG) was originally developed as a universal 5km performance metric — one that accounts simultaneously for course difficulty, age, and gender. In our earlier paper Is your parkrun grade the whole picture? we validated TAG across 11.4 million UK parkrun results, showing that it eliminates the majority of grade variation attributable to course terrain. TAG has since been incorporated into the Skamper platform as a cross-distance race predictor: given a parkrun time at any course, it generates predicted times at any target distance and course type.

The analysis in this paper stands on the shoulders of that earlier work in a way worth making explicit. Comparing a runner's performance at one parkrun course with another — or predicting from parkrun to half marathon — is only meaningful if you can put those performances on a common scale. A runner finishing a 5km in 25 minutes at Great Yarmouth North Beach, where competitors run over 25% slower than on the fastest UK courses, is doing something categorically different from a runner finishing 25 minutes at Belfast Victoria on a flat, fast course. Without reliable course factors, this entire analysis collapses into noise.

Those course factors — derived by two entirely independent methods across all 860 UK parkrun courses — are what make the half marathon prediction meaningful. They allow us to strip out the terrain effect from a parkrun PB and express the underlying running ability as a single number: the TAG. That number is then used to project forward to a half marathon.

This paper asks a more pointed question: when a runner enters their 2025 parkrun seasonal PB into the TAG planner and asks for a half marathon prediction, how close is that prediction to their actual race time?

This is a harder test than validating the 5km metric itself. Predicting a 21.1km race from a 5km effort requires assumptions about how well 5km running economy translates to longer distances — assumptions that may break down for recreational runners who approach the two distances at very different relative intensities.

We use data from two UK half marathons held on the same day — 22 March 2026 — to provide independent replication across two different course profiles, two different runner populations, and two different parts of the country.

TAG uses gender-specific ALJ base times throughout — a critical distinction from earlier implementations. For female runners, the female ALJ base time at the runner's recovered age is used; for male runners, the male ALJ base time. This ensures the TAG% is calibrated on the same scale for both genders, and that the Skamper prediction formula — which uses gender-specific age indices — receives a consistently scaled input.

Given a TAG%, the predicted half marathon time is derived using the Skamper prediction formula applied to the race course:

What TAG% means in plain English

A runner's TAG% expresses their 5km parkrun performance as a fraction of what a world-class runner of the same age and gender would achieve on the same course. A TAG of 75% means the runner is performing at 75% of age-and-gender-adjusted world-class level. The prediction then asks: if this runner maintains 75% of that standard over the half marathon, what time does that imply? The key assumption is that TAG — the fraction of world-class — is conserved across distances for a given runner.

Step 1 — Identify club runners in each race

We restrict the analysis to runners affiliated with a running club. Club membership gives us a matching key that reduces false name matches, and club runners are more likely to be racing to their potential. For Wilmslow, the results CSV contains a Club field. For Brentwood, the equivalent is the Team field in the Excel results. Runners are deduplicated on first name + last name + club.

Step 2 — Match to 2025 parkrun PB

For each club runner, we search the 11.4 million 2025 UK parkrun results for a matching full name (normalised to lowercase). A critical filter requires at least 3 parkruns in 2025 — runners with fewer results may have a PB representing an unusually good day rather than a genuine seasonal level. The runner's best (highest TAG) 2025 parkrun run is used.

Matching is performed on full name only — club names frequently differ between race registration and parkrun. Where the same name appears in parkrun under multiple clubs, the entry whose club most closely matches the race registration is preferred as a tiebreaker.

Step 3 — Age recovery

Exact ages are reverse-engineered from the parkrun age grade and finish time: implied_OC = grade × finish_sec / 100, matched against the gender-specific ALJ pr_base table. Female runners use the female pr_base table; male runners use the male table. Where the implied OC matches multiple candidate ages (both sides of the age curve, or the flat zone where multiple ages share the same base time), the Age Group band from the parkrun record is used to select the correct age. Juniors whose OC falls outside the table receive the Age Group band midpoint. The result: 100% of recovered ages fall within their declared Age Group band.

Step 4 — TAG computation

TAG is computed using gender-specific ALJ base times: tag_base_female_s for women, tag_base_male_s for men. The gender-specific course factor (factor_women or factor_men) is applied. This ensures TAG is on a consistent scale for both genders, compatible with the Skamper prediction formula's gender-specific age indices.

Step 5 — Anomaly removal

Three filters remove unreliable results:

• Parkrun PB slower than 30:00 — age grading unreliable at very slow paces
• Parkrun PB faster than 13:44 — faster than the parkrun world record, indicating a data error
• TAG% > 105% — extremely rare but possible given the formula construction; above this threshold indicates a data error

A further filter removes runners whose predicted time differs from actual by more than 25 minutes — almost always the result of a name collision where two different people share a name in both datasets. These are saved separately for inspection but excluded from the core analysis.

Wilmslow Running Festival Half Marathon — 22 March 2026

The Wilmslow Running Festival is one of the largest half marathons in the North West of England, attracting a strong field of club runners from across Greater Manchester, Cheshire, and beyond. The course is entirely paved with 100m of total ascent over 21.1km — gently undulating rather than flat. The official CSV results contained 3,077 half marathon finisher records. Of these, 1,138 were affiliated with a running club — 37% of all finishers, reflecting Wilmslow's strong club running character.

Brentwood Half Marathon — 22 March 2026

The Brentwood Half Marathon is an established Essex race with a slightly more undulating course than Wilmslow — 143m of total ascent compared with 100m. Results were provided in Excel format with 2,160 finisher records. The Team field is the club equivalent; it was populated for only 364 of 2,160 runners — 17%, compared with 37% at Wilmslow. This lower club density reflects Brentwood's broader recreational field profile relative to Wilmslow's competitive club running base.

2025 UK parkrun results

The parkrun dataset covers all UK parkrun results from the 2025 calendar year: 11,203,744 rows across 860 courses, processed in Python with Apache Arrow. After filtering for valid times and age grades, 9.95 million rows remained with valid TAG scores. Of athletes with at least 3 parkruns in 2025, 565,508 unique runners were available for matching.

Matching club runners from both races to the parkrun dataset produced the following:

Match rates of 76–82% are strong for name-only matching without a unique runner ID. The unmatched runners are accounted for by: runners who did not do parkrun in 2025; runners whose name differs between race registration and parkrun; and runners who completed fewer than 3 parkruns in 2025.

Wilmslow produced the larger and more competitive sample: 744 runners in the core analysis set, drawn from clubs across the North West and beyond. The TAG prediction uses the Skamper open time of 3,904 seconds for Wilmslow (21.1km, 100m ascent, 100% paved, band 0).

The distribution is left-skewed: more runners finish faster than predicted than slower, with 61% of the core set running faster than their TAG prediction. This is the opposite of a naive predictor — TAG is slightly conservative on average, particularly for runners racing hard. The median runner beats their prediction by 2.0 minutes.

The negative bias is concentrated in the faster cohorts: among the slowest third of runners (finish times over 1:55), the distribution is much more symmetric, with roughly equal numbers above and below prediction. This is consistent with TAG being calibrated for best effort: competitive runners race hard and beat their prediction; recreational runners run more cautiously and cluster around it.

Gender breakdown — Wilmslow

Both genders show a negative median variance — runners finishing faster than predicted. The male median is 2.0 minutes more negative than female. We examine whether this reflects anything systematic in Section 9.

Closest predictions

Among the 744 runners, the 10 most accurately predicted included Jack Astbury (predicted and actual both 1:34:32 — a perfect prediction), Matthew Allen (2 seconds), Katie Burt (3 seconds), and Martin Sands (3 seconds). Individual prediction accuracy of this calibre reflects the robustness of the TAG formula when inputs are genuinely representative of the runner's fitness.

Brentwood produced a smaller but independently valid sample: 249 runners in the core analysis set. The TAG prediction uses the Skamper open time of 4,008 seconds for Brentwood (21.1km, 143m ascent, 100% paved, band 1 — moderate hills).

The Brentwood distribution is even more left-skewed than Wilmslow: 62% of runners finished faster than their TAG prediction. The median variance of −2.9 minutes is slightly more negative than Wilmslow's −2.0 minutes, consistent with Brentwood's more competitive club field relative to its total field size — only 17% of finishers were club-affiliated, meaning the club runners who do match are disproportionately competitive.

Gender breakdown — Brentwood

The overall median variance figures already show a slight negative bias — runners beating their predictions on average. This effect is much stronger when we isolate the fastest 10% of runners by actual finish time: those most likely to be racing to their potential.

The male top 10% shows a larger median gap than the female top 10% at both races — 6.7 vs 4.9 minutes at Wilmslow, 8.2 vs 4.5 at Brentwood. This likely reflects the harder racing effort male club runners apply at the sharp end of these events, rather than any formula effect. Female ±5-minute accuracy (52–57%) is actually higher than male (24–30%), consistent with women in the competitive cohort racing at a more predictable fraction of their potential.

The consistent gap between prediction and actual in the top 10% reflects parkrun PBs being set under conditions of lower relative effort than a race-day half marathon. Competitive runners often treat parkrun as a quality training run rather than an absolute maximum effort, meaning their seasonal PB understates their true 5km ceiling. A TAG built on a genuinely maximal parkrun effort would produce tighter predictions. This is not a formula failure — it is the expected result when inputs are training performances rather than race performances.

Across both races, male runners show a 2.0–2.1 minute larger negative median variance than female runners — meaning men beat their predictions by more than women on average. This is the opposite of what a formula bias against women would produce. We examine potential explanations.

The formula uses gender-specific base times throughout

TAG is computed with gender-specific ALJ base times (tag_base_female_s for women, tag_base_male_s for men), and the Skamper prediction uses gender-specific age indices (female rebased index for women, male index for men). Both steps are calibrated separately and consistently. A female runner with TAG 98% is expressing 98% of world-class female standard — directly comparable with a male runner at TAG 98%.

The gender gap is not a formula bias

If the formula were biased against women, we would expect women to finish slower than predicted relative to men — i.e. a larger positive variance for women. Instead, both genders finish faster than predicted, and women finish closer to prediction. The gap is in the direction and magnitude of conservative underestimation, not in any systematic over-prediction for one gender.

Most likely explanation — relative racing effort

The most plausible explanation is that the gender difference reflects variation in how runners distribute effort across different formats — parkrun, training runs, and target races. This varies considerably between individuals regardless of gender, and the data do not support drawing directional conclusions about how any group approaches racing. What the data do show clearly is that the formula itself is not the source of the gap.

Across 993 club runners at two independent half marathons, the TAG planner produced the following results:

1. TAG is a valid half marathon predictor, slightly conservative for competitive runners. Among the top 10% of finishers, TAG predicts within ±5 minutes for 24–57% of runners depending on gender and race, with a median error of 4.5–8.2 minutes in the conservative direction. The formula performs better as a floor than a ceiling: competitive runners exceed it rather than falling short.
2. The negative bias in the overall dataset is consistent and real. The median club runner at both races finished 2–3 minutes faster than their TAG prediction. This is not a formula error — it reflects TAG being calibrated for best effort, while many parkrun PBs are set as training runs rather than maximum efforts. Runners who use parkrun casually will consistently beat their TAG-implied half marathon time.
3. TAG is not biased against women. Both genders beat their predictions in the full dataset and in the top 10%. Female ±5-minute accuracy in the competitive cohort (52–57%) is higher than male (24–30%). There is no evidence of systematic miscalibration for women.
4. The Skamper course model correctly differentiates course difficulty. The Brentwood prediction used a different grade band in the Skamper model (band 1, power 0.105) than Wilmslow (band 0, power 0.1), producing an open time 64 seconds higher. The Brentwood results are consistent with Wilmslow after this adjustment, validating the course model.

Name matching without unique IDs

The principal limitation of this analysis is the absence of a unique runner identifier linking parkrun and race results. Name matching inevitably introduces both false positives (two runners sharing a name) and false negatives (the same runner listed differently). Our 25-minute extreme variance filter catches most collisions, but moderate collisions in the 5–15 minute range cannot be entirely ruled out. An athlete ID system shared across events would be transformative.

Sample size and selection bias

The Brentwood female top 10% group contains only 7 runners — too small for robust statistical inference. The Wilmslow female top 10% (n=23) is more useful but still limited. Both samples cover club runners only — a self-selected group more competitive than the average half marathon field. Generalising to all half marathon participants requires caution.

Single race day and conditions

Both races took place on 22 March 2026 — a dry, cool day generally favourable for half marathon running. Results may differ in adverse weather. A multi-race, multi-season analysis would provide more robust estimates of TAG accuracy across conditions and training cycles.

Parkrun PB timing relative to the race

We use each runner's best 2025 parkrun TAG — which may have been set months before or after the race. A runner peaking for March may have their highest TAG in February; a runner who peaked in October may have a lower TAG in March. This introduces noise that a seasonal or proximity-weighted PB selection might reduce.