National Running Club Database: Assessing Collegiate Club Athletes' Cross Country Race Results

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

Abstract

Collegiate cross-country scheduling often relies on coaching intuition rather than empirical evidence due to the scarcity of open-access performance data. This paper presents the National Running Club Database (NRCD), comprising over 23,000 race results from the 2023–2025 seasons. By applying a two-tier standardization framework and machine learning models, the research identifies race frequency as a primary driver of athlete improvement and team success, contradicting the common practice of limiting race participation to avoid burnout.

Introduction: The Data Gap in Cross-Country

In collegiate cross-country, scheduling is frequently driven by tradition and "gut feeling." Coaches often limit race frequency to "save" athletes for championship meets, fearing burnout or injury. This reliance on intuition persists because high-quality, large-scale performance data is typically locked behind proprietary paywalls that prevent bulk downloads for research.

Furthermore, cross-country is inherently "noisy." Unlike track and field, where conditions are controlled, cross-country times are affected by varying distances, extreme weather, and hilly terrain. This makes it nearly impossible to compare a 5k in high heat on a muddy trail to an 8k on a flat, cool morning. This study provides both the data and the mathematical framework to normalize these performances and understand the true drivers of athlete progression.

Methodology: Standardizing the "Noisy" Field

The authors developed a two-tier standardization framework to enable "apples-to-apples" comparisons across different races:

  1. Converted Only: Uses the Riegel Race Time Prediction Formula to adjust for non-standard race distances (e.g., converting a 5-mile race to the standard 8k for men).
  2. Standardized: A "Master Standardization Formula" that accounts for weather (temperature and dew point) and course topography (elevation gain/loss as a percentage grade).
Overview of the NRCD data preprocessing and analysis pipeline.

To analyze the data, 29 features were engineered—such as "slope" (improvement trajectory) and "experience level"—and used to train six machine learning models. Random Forest and Gradient Boosting were prioritized for their ability to handle non-linear interactions. The study utilized temporal validation, training on 2023 data to predict 2024/2025 outcomes, ensuring model generalizability.

The National Running Club Database (NRCD)

The paper introduces the National Running Club Database (NRCD), an open-access resource for sports analytics. The dataset covers 23,725 race results from 7,594 collegiate club athletes (NIRCA) and includes detailed metadata for 500 unique course locations.

Summary of the National Running Club Database (NRCD) composition.

Key Results: The Impact of Race Frequency

The machine learning models achieved significant predictive power, with the Random Forest model reaching an $R^2$ of 0.92 for the combined dataset. Several key insights emerged:

  • The "More is Better" Paradox: The strongest predictor of improvement was race frequency. Athletes who raced more often improved significantly more than those who "saved themselves."
  • Diminishing Returns for Elites: Athletes with slower initial times showed the greatest room for improvement. Elite runners (top 10%) often showed "positive slopes," meaning their times stayed flat or slightly regressed as the season progressed, likely due to physiological limits.
  • Gendered Predictors: Drivers of success differed by gender. "Slope" (previous race patterns) was a much stronger predictor for women (32.9% importance) than for men (13.2%), while "bad race count" (consistency) mattered more for men.
  • Team-Level Success: There is a statistically significant correlation ($p < 0.01$) between frequent racing and national rankings. Between 60–80% of teams finishing in the top 15 at Nationals had athletes who raced four or more times during the regular season.

Conclusion and Future Directions

While the NRCD is a landmark dataset, it currently spans only three years, limiting long-term trajectory tracking. Additionally, the models do not yet account for in-race tactics or individual physiological factors like training volume and injury history. Future research aims to integrate wearable sensor data (GPS/Heart Rate) with the NRCD to create a more holistic model of collegiate distance running performance.

Summary updated

BibTeX
@misc{jr2025national,
  author = {Jonathan A. Karr Jr. and Ben Darden and Nicholas Pell and Ryan M. Fryer and Kayla Ambrose and Evan Hall and Ramzi K. Bualuan and Nitesh V. Chawla},
  title = {{National Running Club Database: Assessing Collegiate Club Athletes' Cross Country Race Results}},
  year = {2025},
  eprint = {2509.10600},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.10600},
}
Figure