Predicting biathlon shooting performance using machine learning

ABSTRACT

Shooting in biathlon competitions substantially influences final rankings, but the predictability of hits and misses is unknown. The aims of the current study were A) to explore factors influencing biathlon shooting performance and B) to predict future hits and misses. We explored data from 118,300 shots from 4 seasons and trained various machine learning models before predicting 34,340 future shots (in the subsequent season). A) Lower hit rates were discovered in the sprint and pursuit disciplines compared to individual and mass start (P < 0.01, h = 0.14), in standing compared to prone shooting (P < 0.01, h = 0.15) and in the 1^st prone and 5^th standing shot (P < 0.01, h = 0.08 and P < 0.05, h = 0.05). B) A tree-based boosting model predicted future shots with an area under the ROC curve of 0.62, 95% CI [0.60, 0.63], slightly outperforming a simple logistic regression model and an artificial neural network (P < 0.01). The dominant predictor was an athlete’s preceding mode-specific hit rate, but a high degree of randomness persisted, which complex models could not substantially reduce. Athletes should focus on overall mode-specific hit rates which epitomise shooting skill, while other influences seem minor.

KEYWORDS:

• Sport
• competition
• modelling

Acknowledgments

This study was supported by the Swiss Federal Institute of Sport Magglingen (SFISM). We would like to thank Dr Joachim Steinwendner and Dr Beat Tödtli from the Swiss Distance UAS for the training in machine learning.

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplemental online material

The complete analysis code and the raw data for this article can be accessed at Figshare (http://tandf.figshare.com/).