PaperlayerView exact revision on arXiv

Source-linked AI summary

NRCD: An Open Database of Collegiate Running with Unified Performance Standardization

Jonathan A. Karr, Ryan M. Fryer, Ben Darden, Nicholas Pell, Kayla Ambrose, Evan Hall, Ramzi K. Bualuan, Nitesh V. Chawla

arXiv:2608.14776v1cs.LGcs.CYcs.IR

TL;DR

Large-scale collegiate running research lacks bulk-accessible, comparable data. NRCD provides an open dataset and unified standardization framework, reducing median within-athlete cross-meet variability by 51.0% for women and 34.4% for men on cross country.

  • Problem

    Existing collegiate running platforms lack bulk relational export, while course and venue differences limit comparability and prior open analyses use small, male-skewed samples.

  • Method

    NRCD constructs a community-governed open dataset and applies sport- and gender-specific distance, elevation, altitude, and heat adjustments in a unified framework.

  • Results

    51.0% for women and 34.4% for men: full standardization reduces median within-athlete cross-meet variability on cross country versus raw times.

  • Takeaways & Limitations

    NRCD removes the bulk-export barrier and supports longitudinal, environmental-confounder, and gender-equity research in collegiate sport.

  • Takeaways & Limitations

    Validation of variance and improvement uses only cross country and supports environmentally aware standards rather than individual training plans.

Abstract

from arXiv · show

Collegiate running in the United States generates thousands of race results annually in cross country and track and field, yet no large-scale dataset has been publicly available for research. Existing websites such as Athletic.net, MileSplit, and TFRRS host results but do not support bulk download, restricting prior analyses to ~500 performances, often skewing studies toward male athletes. We introduce the National Running Club Database (NRCD), the first openly available collegiate running dataset at scale: 128,963 approved performances from 28,913 athletes across 1,336 meets in four sports (cross country (XC), indoor and outdoor track, and road races), 36.3% women, spanning 2004 through 2026. Within that single export, meets from August 2023 onward carry comprehensive course distance, elevation gain and loss, weather at race time, and track venue metadata (97.7% of XC rows with weather fields); earlier seasons back to 2004 are included with sparser metadata. NRCD is community-governed through open submission and expert approval and is maintained as a live database whose meet volume has grown yearly. We release a unified performance standardization framework that operationalizes established distance, elevation, and heat adjustments in one pipeline. Furthermore, we recommend gender-stratified modeling. On XC, full standardization lowers median within-athlete cross-meet variability by 51.0% (women) and 34.4% (men) versus raw times. We release the dataset and pipeline with a python package `nrcd' under FAIR principles, supporting longitudinal athlete modeling, environmental-confounder studies, and gender-equity research in collegiate sport.

1 Introduction

NRCD addresses the lack of bulk-accessible, comparable collegiate running data by providing an open, community-governed dataset and unified standardization framework. The resource supports longitudinal, environmental, fairness, and reproducible sports-informatics research with gender-stratified validation.

  • Dataset and scope: 128,963 results (36.3% women) span four sports in NRCD, broadening coverage beyond a single running subfield or ability tier.The resource is designed for the full band of runners represented across collegiate, club, and road-race contexts.
  • Motivation: Thousands of annual results exist, but major platforms lack bulk relational export, forcing prior open analyses to use samples of a few hundred records with documented male skew.Cross-country comparability is also limited by differences in measured distance, terrain, and weather.
  • Resource contribution: NRCD combines a public dataset, validated conversion formula, pip-installable ‘nrcd’ Python package, and live moderated contributed website.The live website supports future sustainability and community participation.
  • Standardization and validation: NRCD provides environmentally aware standardization with converted-only and standardized modes, a derived heat law (𝑘=0.0016), and separate empirical analyses for men and women.The framework is empirically validated on cross country, where each fall meet is the same aerobic race for a given gender.
  • Resource contribution: Community-governed curation adds submitted results that domain experts approve, supporting a maintained open resource.This governance model is identified as a core NRCD contribution.

2 Related Work

NRCD addresses gaps in open collegiate running data by providing large-scale meet results with course and weather metadata. It also unifies established performance-adjustment methods while distinguishing governed meet data from fitness-app training records.

  • Open running data: Prior open collegiate running work used fewer than 1,000 hand-curated results, whereas no prior resource offered close to 100,000+ results with course and weather metadata.NRCD is positioned as a substantially larger open resource with environmental and course information.
  • Standardization: NRCD integrates Riegel’s law, Maurer elevation factors, and weather–performance research in one pipeline using a quadratic heat surrogate fit to Hadley’s coaching bands.The framework excludes AQI adjustment.
  • Community sports data: Unlike fitness apps that record training, NRCD provides governed meet results with unique ids for user tracking.This distinction concerns the type of activity data captured and how records can be tracked.

3 The NRCD Dataset

The NRCD dataset spans four collegiate running sports and 128,963 approved results, divided nearly evenly between comprehensive post-August 2023 records and historical records from 2004 through July 2023. It also supports longitudinal analysis, with substantial subsets of athletes competing across at least two or three school years.

  • Dataset scope: 128,963 approved results cover Cross Country, Indoor Track, Outdoor Track, and Road Race.The export includes 64,209 comprehensive results and 64,754 historical results.
  • Temporal coverage: 64,209 comprehensive results date from August 2023 onward, while 64,754 historical results span 2004 through July 2023.The comprehensive subset contains richer metadata and is defined by meet dates from August 2023 onward.
  • Longitudinal coverage: 10,335 athletes (35.7%) have results in at least two school years, and 4,687 (16.2%) have results in at least three.School years are defined as August 1 through July 31.

4 Community-Governed Curation

NRCD uses community-contributed meet submissions combined with nationwide administrative review, while domain-expert approval controls which entries enter the database. Approved-meet coverage expanded from 68 to 142, 340, and 281 meets per year across 2022–2025.

  • Community Contributions: Coaches, athletes, and volunteers submit meets, making NRCD open-contributed.Community submissions provide the initial source of meet records.
  • Expert Approval: Every submitted meet requires approval by an administrator with domain expertise against public postings.This review occurs before an entry is added to NRCD.
  • Coverage Expansion: Administrative reviews of public results and registered teams add missing meets beyond user submissions, broadening coverage nationwide.Admins review public meet results and registered teams to identify omitted meets.

5 Unified Standardization Framework

The framework integrates approved results with athlete, team, event, and course metadata, then applies sport- and gender-specific factors to produce comparable times. For cross country, standardization follows a fixed sequence of environmental, course, altitude, and distance adjustments before Riegel conversion.

  • Pipeline overview: Approved results join athlete, team, event, and course metadata before sport- and gender-specific factors produce comparable times within each sport and gender.The pipeline is summarized in Figure 1 and uses factors listed by sport in Table 5.
  • Cross-country standardization: For XC, factors apply in fixed order: weather, grade, venue altitude, measured- versus reported-course length, then Riegel conversion to d_target.The XC sequence uses f_alt for venue altitude and Equation 2 for conversion to the target distance; heat is addressed in Section 6.1.
  • Transformation: Raw time t_raw maps to standardized time through multiplicative factors followed by Riegel conversion.The framework presents this mapping using the cited Riegel method.

1 For XC,

For cross country and road performances, NRCD applies a unified adjustment pipeline combining Riegel distance conversion with weather, elevation, course-distance, and altitude corrections when metadata are available. Analyses are recommended to be fit and reported separately by gender rather than pooling men’s and women’s performances.

  • Gender-stratified analysis: Gender-stratified analysis recommends fitting, evaluating, and reporting every model separately, never pooling raw or adjusted seconds across men and women.The recommendation applies to both raw and standardized performances.
  • Weather adjustment: AQI is measured when available but omitted from the adjustment factor because it is confounded with temperature and dew point.Heat adjustment instead uses heat index H, defined as temperature plus dew point.
  • Distance and elevation: Measured course lengths differing from reported distances add (d_actual/d_reported)^b before conversion to target distances.Target distances are 8000 m for men and 6000 m for women, following NIRCA defaults.
  • Distance and elevation: Elevation adjustment uses f_elev = 1.04^g·0.9633^l for gain and loss grades, with grades derived from measured length.The conversion accounts for both course elevation and measured course length.
  • XC standardization: XC and road standardization combines Riegel distance adjustment, weather, elevation, course-distance accuracy, and altitude correction when metadata exist.The pipeline uses f_weather, f_elev, course-distance accuracy, and f_alt alongside Equation 1.

6 Formula Validation (XC)

XC formula validation uses gender-stratified checks of heuristic fidelity, monotonic weather behavior, plausibility, within-athlete variance, and within-season improvement, without predictive benchmarking or pooling 6 km and 8 km seconds. Full standardization reduces cross-meet variability, while converted-only times inflate improvement estimates and retain environmental effects.

  • 6 Formula Validation (XC): Validation used three formula-focused checks, with all summaries gender-stratified and absolute seconds never pooled across 6 km and 8 km targets.The checks were not predictive benchmarks.
  • 6.1 Heat Surrogate: 0.34 pp RMSE against Hadley band midpoints and 70% of H∈[101, 180] within published ranges support heuristic fidelity for k=0.0016.The validation explicitly does not constitute independent physiological validation.
  • 6.1 Heat Surrogate: For H≤100, f_weather=1; above 100, the factor decreases monotonically, with mean factor 0.990 across 35,909 weather-observed XC rows.Finish-time recalibration of k was unstable and rejected; metrics were stable for the reported k set, though the passage truncates that set.
  • 6.2 Within-Athlete Variance: 51.0% for women and 34.4% for men are the reported reductions in median within-athlete cross-meet variability after full standardization.The analysis included athletes with ≥3 regular-season XC meets per season, excluded nationals, and used gender-specific 6 km or 8 km targets.
  • 6.3 Within-Season Improvement: 12% for women and 21% for men are the reported inflation of median within-season improvement from converted-only versus full standardization.Converted-only retains fall cooling and grade effects.
  • 6.3 Within-Season Improvement: Converted-only finish times produced higher test R^2 than standardized times for the reported gender-stratified random-forest and gradient-boosting models.Models trained on 2023 and tested on 2024 predicted within-season improvement rate from season-level XC features.
  • 6.3 Within-Season Improvement: The validation reflects NIRCA club competition, while pre-August 2023 entries lack complete metadata and adjustments target environmentally aware standards rather than individual training plans.Variance and improvement validation use XC only.

7 Conclusion

NRCD is presented as the first open large-scale collegiate running dataset with documented standardization. Validation and gender-stratified checks support full standardization and gender-stratified modeling for derivative work.

  • NRCD is the first open large-scale collegiate running dataset with documented standardization.
  • 51.0% for women and 34.4% for men: full standardization reduces within-athlete cross-meet variability versus raw times.
  • The derived heat surrogate is validated against Hadley’s piecewise bands rather than finish-time regression.
  • Gender-stratified modeling is recommended for all derivative work because converted-only times can overstate within-season improvement.

Ethics Statement

The study uses publicly accessible NRCD data with permission, removes personally identifiable information, and was classified by Notre Dame’s IRB as not human subject research.

  • NRCD data are publicly accessible, and the researchers had permission to use and publish the dataset.
  • The researchers removed personally identifiable information, including links and people’s names.
  • Notre Dame’s IRB office classified the research as not human subject research.

Generative AI Usage Disclosure … B Cross Country Performance Band and NCAA Context

The appendix documents validation and robustness analyses for standardized collegiate running results, showing that distance conversion drives most XC variance reduction while environmental effects and cross-sport adjustments are generally smaller. It also contextualizes standardized XC career performance against NCAA team depth and highlights limitations of finisher-only, heterogeneous, and non-independent analyses.

  • Generative AI Usage Disclosure / Appendix: Moderate AI use supported code generation and paper editing, while all content was validated by the authors and no AI handled result upload or validation.The appendix extends the CIKM paper submission.
  • A.1 Stepwise variance decomposition (XC): 96.5% of women’s and 88.7% of men’s raw-to-full median-SD reduction came from distance conversion, while environmental adjustment contributed 3.5% and 11.3%.Converted-only to full standardization further reduced median cross-meet SD by 3.5% for women and 5.6% for men.
  • A.2 Environmental factor distributions (comprehensive XC): 1.0 was the elevation-factor median with P95 1.002 after converting gain/loss from feet to grade percent, while weather produced the only material active-factor tail.Population-level environmental corrections were small near the median, but tails mattered for individual meets.
  • A.3 Within-season improvement direction: 70.2% / 68.2% of women’s / men’s athlete-seasons were classified as improving under converted-only times, versus 67.8% / 65.9% under full standardization.Converted-only times also inflated median improvement magnitude by 12% / 21% for women / men, consistent with environmental drift mimicking fitness gains.
  • A.4 Illustrative ML R^2 and bootstrap uncertainty: The illustrative gender-stratified ML analysis trained on 2023 and tested on 2024, using same-result-time features and bootstrap 95% CIs; higher converted-only R^2 reflects retained environmental structure, not external validity.Gradient-boosting gaps were larger and significant for both genders, whereas random-forest gaps were small and not significant; neither model used weather columns directly.
  • A.5 Extended robustness checks / A.5.1 Riegel exponent sensitivity / A.5.2 Finisher-only export and minimum-race sensitivity: 0.3 pp was the maximum XC variance-validation shift from a unified b=1.06, while forcing b=1.08 for both genders shifted men’s reduction from 34.4% to 35.0%.Gender-specific default exponents were b=1.055 for men and b=1.08 for women; requiring ≥2, ≥3, or ≥4 races yielded men’s inflation of 18.6 to 20.7% and women’s of 8.7 to 13.4%.
  • A.5.3 Outdoor track wind and venue factors / A.5.4 Cross-sport standardization magnitude / A.5.5 Road race exploratory transfer checks: 6.4% of applicable outdoor-track sprint/hurdle rows recorded wind, and the full track pipeline had median factor 0.998; XC had 51.5% of comprehensive-era factors outside [0.98, 1.02].Indoor track had 0.10% outside the band, while outdoor track and road were intermediate; road environmental adjustment reduced median within-athlete SD by only 1.5–2.7% across heterogeneous distances.
  • B Cross Country Performance Band and NCAA Context: 7.2% of men’s career PRs occupied the NIRCA–D1 gap, while only 1.3% of women’s PRs reached D1 5th-runner pace versus 10.8% at NIRCA/D2.Career PRs used fully standardized XC results, included nationals, and about 25% occurred at NIRCA nationals; NCAA D1 5th-runner pace remained uncommon in longitudinal subsets (≤12% men, ≤4% women).

C Metadata Missingness: Results vs. Meets … H Factors Not Covered by NRCD Standardization

NRCD provides a versioned, community-maintained collegiate running dataset and standardization package, while documenting uneven metadata coverage, linkage and selection limitations, and factors that remain unmodeled. Its longitudinal depth varies by sport, and users should interpret metadata as meet- or course-level rather than athlete- or route-level measurements.

  • C Metadata Missingness: Results vs. Meets: Result-level metadata coverage can overstate calendar-wide coverage because large invitationals contribute many results from relatively few weather-enriched rows; meet-level rates are often more informative.Among 280 comprehensive-era XC meets, the largest 10 contain 28.0% of results and the largest 25 contain 45.5%; median meet size is 25, with a maximum of 1,361.
  • C Metadata Missingness: Results vs. Meets: Historical rows remain sparse by design, with planned backfilling focused on course gain/loss, measured length, outdoor-track weather, and per-meet completeness.Roughly half of export rows predate the comprehensive metadata era, when course weather and measured-distance fields were rarely available.
  • C.1 Geolocation, altitude, and weather provenance: Weather and altitude represent meet- or course-level API and geocoded locations, using one hourly weather snapshot that may differ from conditions along the race route.Altitude comes from USGS EPQS at geocoded host coordinates, while weather uses OpenWeather One Call 3.0 timemachine at the same coordinates; air-quality history begins 27 November 2020.
  • C.2 Athlete deduplication: Duplicate athlete identities are reviewed administratively rather than auto-merged, prioritizing avoidance of false merges but leaving possible split identities in the public export.Suggestions use same-team initial matching or normalized full-name similarity >90%, and researchers should treat multi-ID athletes as residual linkage error.
  • C.4 Community submission error rates and selection: Every released row is admin-approved, but transcription errors, non-random submission, and under-representation of inactive programs and small invitationals remain possible.NRCD does not claim a measured residual error rate because no independent census exists for precision/recall estimation.
  • D Longitudinal Depth by Sport: XC offers the strongest longitudinal structure, with a median of 2 meets per season, P90 of 4, and 35.7% of athletes spanning at least 2 school years; road racing is sparse.Track athletes accumulate more results per season but fewer meets, while road racing has a median of one result per athlete-season.
  • F Sustainability, Versioning, and Intended Users: The live site accepts meets continuously, exports are planned yearly or after major schema changes, and the MIT-licensed nrcd package enables users to apply standardized formulas to private or new results.Approved-meet volume changed from 68 to 142 to 340 to 281 across 2022–2025; standardization adjusts distance, course length, grade, heat, altitude, and track venue and wind when metadata exist.

H.1 Training load, aerobic fitness, and recovery markers … H.19 Officiating, timing, and wind on non-sprint track events

NRCD standardizes observed race outcomes but omits many training, physiological, behavioral, environmental, tactical, selection, and measurement factors that influence performance or race availability. These latent factors require gender-stratified interpretation and explicit modeling or bounding in causal and predictive analyses.

  • H.1 Training load, aerobic fitness, and recovery markers; H.2 Within-season race frequency and cumulative fatigue: NRCD records race dates and finish times, but not training volume, aerobic capacity, recovery markers, body composition, or surrounding meet density, leaving readiness and cumulative fatigue latent.Unobserved factors include VO2 max, HRV, sleep-stage data, training logs, and races started during the prior fortnight.
  • H.3 Injury and illness; H.4 Sleep; H.5 Nutrition and energy availability; H.6 Caffeine and legal ergogenic aids; H.7 Hydration, dehydration, and exercise-associated cramping; H.8 Travel and jet lag; H.9 Academic load and student-athlete time demands; H.10 Circadian rhythm and time of day; H.11 Precipitation, footing, and solar load: Injury, illness, sleep, nutrition, supplementation, hydration, travel, academic demands, circadian timing, and footing can affect performance or availability but are incompletely captured or unmodeled.NRCD lacks injury and illness status, sleep fields, fueling and supplement timing, travel duration, academic stress, circadian adjustment, and footing factors.
  • H.12 Altitude acclimatization; H.13 Footwear and surface conditions; H.18 Course geometry beyond aggregate grade: Meet elevation is recorded, but acclimatization history, footwear, spike choice, surface moisture, terrain, turn frequency, and choke points are not, leaving additional course-related variation in the residual.About 1% higher metabolic cost per 100 g added mass is associated with proportionally slower time trials, while uneven micro-terrain can raise energy cost by several percent.
  • H.14 Gender-specific factors in women; H.14.1 Menstrual-cycle phase and endogenous hormones; H.14.2 Menstrual symptoms and race availability; H.14.3 Iron deficiency and anemia; H.14.4 Low energy availability and menstrual dysfunction (RED-S); H.14.5 Oral contraceptives and hormonal contraception; H.14.6 Implications for women’s XC in NRCD: Women comprise 39.4% of unique athletes and 36.3% of result rows, while menstrual symptoms, iron status, RED-S, contraceptive use, and race skipping can affect observed performance and availability.84 to 91% of surveyed exercising women reported cycle-related symptoms, and symptom burden predicted missing or altering training and missing competitions with odds ratios ≈1.07 to 1.09 per symptom-index unit.
  • H.15 Latent factors in men’s physiology: Men constitute 60.6% of unique athletes and 63.7% of result rows, yet NRCD also omits male-specific physiology, iron status, and low-energy-availability information.The passage notes that male distance runners can experience iron deficiency without anemia and RED-S-related bone stress and hormonal suppression.
  • H.16 Psychology, pacing, and race tactics; H.17 Selection and roster composition: Expected duration, stakes, pack dynamics, pacing, field depth, course familiarity, home support, and teammate-oriented effort affect racing, but NRCD stores one aggregate finisher time without tactical or motivational context.The dataset does not include splits, pack position, within-race tactical profiles, or whether athletes raced for personal outcomes or teammates.
  • H.19 Officiating, timing, and wind on non-sprint track events: Photo-finish latency, course-cutting penalties, starter variability, and missing non-sprint wind metadata introduce measurement error unrelated to fitness, so standardized times remain environment-adjusted outcomes rather than complete fitness proxies.Wind corrections are not applied to non-sprint track events, and researchers are advised to explicitly model or bound latent factors in causal or predictive analyses.
Loading 2608.14776v1…