Bayesian modelling of elite sporting performance with large databases

Jim E. Griffin, Laurentiu Hinoveanu, James HopkerJournal of Quantitative Analysis in Sports 2022 · DOI

TL;DR

Researchers have developed a flexible Bayesian hierarchical model to track how elite athletic performance evolves with age, accounting for individual variability and environmental "confounders" like wind speed. Applying this to massive sprinting and weightlifting databases, they found that men typically peak two years earlier than women and that performance distributions are significantly skewed—athletes are far more likely to have an "off day" than a "superhuman" one.

What Problem Does This Paper Address?

Predicting an athlete's trajectory is the "Holy Grail" for coaches and funding bodies. However, statistical modeling of sports data is notoriously difficult. Simple averages (cross-sectional analysis) suffer from selection bias: only exceptional teenagers compete at the elite level, making the "average" youngster look better than they are. Conversely, older athletes who stay in the sport are usually those who haven't declined yet.

Furthermore, performance isn't just about age; it’s affected by "confounders" like wind speed in sprinting or the time of year. Previous models often assumed a "one-size-fits-all" aging curve or used rigid distributions that couldn't account for the fact that sports results are rarely "normal" (Gaussian).

Methods

The authors propose a Bayesian longitudinal model that decomposes an athlete’s performance into three parts:

  1. The Population Trajectory: A 4th-order polynomial representing the "average" aging effect.
  2. Individual Excess Performance: A flexible, individual-specific curve modeled using piecewise linear splines. To prevent overfitting (since some athletes have many more data points than others), they use Bayesian variable selection to decide which "knots" in the spline are actually necessary for a specific athlete.
  3. The Error Term: Instead of a standard Bell curve, they use a skew-t distribution. This allows the model to handle "heavy tails" (outliers) and skewness—essential because an athlete can easily run a second slower than their best, but they almost never run a second faster.

The model is fitted using Markov chain Monte Carlo (MCMC), specifically a Gibbs sampler with an adaptive Metropolis-Hastings proposal to handle the high-dimensional variable selection efficiently.

Datasets

The study utilizes two massive longitudinal datasets:

  • 100m Sprinting: Data from the World Athletics website (2001–2021), including 2,834 males (95,376 performances) and 1,297 females (48,999 performances). All athletes had a personal best below 10.5s (men) or 11.6s (women).
  • Weightlifting: Data from the International Weightlifting Federation (1998–2021), featuring 1,609 males and 1,212 females. To compare across weight classes, the authors used Sinclair totals, a standard formula that normalizes lifted weight relative to body mass.

Key Insights and Ideas

  • The "Off-Day" Factor: The use of the skew-t distribution was vindicated. The data shows a clear negative skew, confirming that elite performance is fragile; external factors are much more likely to drag a performance down than push it significantly above an athlete's physical limit.
  • The Gender Peak Gap: In both sprinting and weightlifting, men peak significantly earlier than women. The population peak for men is roughly 24 (sprinting) and 25 (weightlifting), while for women, it is 26 and 27, respectively.
  • Sprinting vs. Strength: Weightlifting performance plateaus for much longer than sprinting. While sprinters see a relatively sharp peak followed by a decline, weightlifters often maintain near-peak performance from their mid-20s until age 30.
  • Individual Plasticity: The model successfully identified "double peaks." For example, one athlete showed a peak at 24, a decline due to injury at 26-27, and a second peak at 29. Standard models would have smoothed this into a single, inaccurate curve.

Key Results

  • Environmental Impact: In sprinting, a 1 m/s tailwind improves male times by an average of 0.05s and female times by 0.06s.
  • Seasonality: Athletes perform slightly better in June and July (the heart of the outdoor season) compared to January, though the effect is smaller than that of wind (approx. 0.04s improvement).
  • Peak Age Distributions: The Bayesian approach provides a probability distribution for peak age. For men's 100m, the peak is tightly clustered around 24.2 years, whereas for women, it is broader, centered around 26.2 years.

Limitations and Future Work

The authors acknowledge several constraints:

  • Selection Bias: The data only includes athletes who reached a certain elite threshold. It doesn't account for the thousands who dropped out before reaching the "top 100" level.
  • Strength of Field: The model treats every competition equally. It doesn't currently account for the "big stage" effect (e.g., performing better in an Olympic final vs. a regional meet).
  • Doping: The authors note that the history of anti-doping violations in these sports is a "latent variable" that is difficult to model but likely impacts the data.
  • Future Directions: The researchers suggest extending this model to rank-based sports (like Tennis) or team sports where player metrics (distance covered, pass accuracy) could be modeled longitudinally.

Summary updated

Abstract

Abstract The availability of large databases of athletic performances offers the opportunity to understand age-related performance progression and to benchmark individual performance against the World’s best. We build a flexible Bayesian model of individual performance progression whilst allowing for confounders, such as atmospheric conditions, and can be fitted using Markov chain Monte Carlo. We show how the model can be used to understand performance progression and the age of peak performance in both individuals and the population. We apply the model to both women and men in 100 m sprinting and weightlifting. In both disciplines, we find that age-related performance is skewed, that the average population performance trajectories of women and men are quite different, and that age of peak performance is substantially different between women and men. We also find that there is substantial variability in individual performance trajectories and the age of peak performance.
BibTeX
@article{griffin2022bayesian,
  author = {Jim E. Griffin and Laurentiu Hinoveanu and James Hopker},
  title = {{Bayesian modelling of elite sporting performance with large databases}},
  journal = {Journal of Quantitative Analysis in Sports},
  year = {2022},
  doi = {10.1515/jqas-2021-0112},
}
Figure