Do sports problems require tailored methods or direct applications? On a swimming-oriented journey with a Bayesian roadmap

Abstract

When looking at a new problem for the first time, the initial question that comes to mind can generally be summarised as: has this problem, or a closely-related one, already been tackled before? Although being a rather recent scientific field of study, sports sciences often face the same pattern, despite the number of recognised ‘resolved problems’ remaining relatively low compared to more established disciplines. In particular, when it comes to statistical tools, there exists in the literature an abundance of methods to handle a variety of problems, such as image classification, missing data reconstruction, time series forecasting, dimensionality reduction, and data visualisation, among others. In this presentation, I will try to illustrate how, in the past 5-years, determining whether my current sport-related problem required developing a novel tailored statistical tool or simply applying an established method on this particular dataset, was generally the most important step of the project to result in an adequate answer. From a few articles using direct applications of Bayesian mixed models on morphological swimming datasets to the 3 years-long development of a novel machine learning framework specifically dedicated to tackling the problem of forecasting irregular time series of swimmers’ performances follow-up, let us explore through these examples the variety of challenges, with their intrinsic complexity, coming from sports sciences applications. In conclusion, I will try to emphasise how those new statistical models, originally tailored to handle sports-related problems, can often find natural applications in other fields like medicine, biology, or robotics (among others), thus contributing to the common methodological toolbox shared by many applicative sciences.