Sprint performance prediction in a national-level database of 58,000 athletes: developmental patterns, population composition effects, and the role of anthropometric features

Abstract

Abstract Background This study evaluated 100-m sprint performance prediction in one of the largest national-level databases of registered 100-m sprinters analyzed to date, quantified how the target-population definition and the outcome definition shape reported accuracy, and tested the incremental value of anthropometric features. Methods Competition records (129,351 100-m results from 55,351 athletes, 2002-2025). Ridge regression, random forests and gradient boosting predicted lifetime personal best at successive age cutoffs, with gradient boosting used for the main series so that one algorithm is applied across all cutoffs and both sexes. A two-layer design separated full-sample analysis from a continuing-career subsample (Group A). Province-disjoint cross-validation tested generalization across province groups, the number of which varies by cutoff and sex.Competition records (129,351 100-m. Result Full-sample R² was 0.870 (95% CI 0.836 to 0.899) at cutoff age 16, but after restricting to continuing-career athletes R² decreased to 0.604, and within the elite stratum (PB < 11.0 s) to 0.446 (0.251 to 0.609). Male R² increased from 0.604 (age 16) to 0.799 (age 18) with province-disjoint ΔR² = −0.006, although this gain attenuated under longer follow-up requirements (0.799 to 0.377 at age 18 with at least two years of post-cutoff observation). Female R² was 0.726, 0.737 and 0.738 at the same cutoffs and showed no comparable increase. Against a genuinely prospective outcome, the best performance recorded after the cutoff, male R² was 0.303, 0.321 and 0.396 at cutoffs 16 to 18, and improvement within a standardised two-year window was not predictable in either sex. Differences between the three algorithms were negligible, and a single recalibrated predictor matched the full feature set to within 0.014 in males and 0.010 in females. Height, weight and BMI added nothing beyond trajectory features (paired ΔR² −0.002 to +0.002 in the largest samples). Peak dropout occurred at age 18 in males (83.9%) and age 17 in females (74.4%). Conclusions Reported R² depended strongly on the target population and the outcome definition, differing by 0.15 to 0.27 units between the full sample and the continuing-career subgroup and by roughly a factor of two between retrospective and prospective outcomes. Within a fixed target population, trajectory shape was the strongest available predictor, but the apparent increase in predictability with age is substantially attenuated under longer follow-up requirements and no useful prediction of future improvement was achieved.

The authors' abstract, as published at the source. BMC Sports Science Medicine and Rehabilitation, 2026 · DOI ↗