Training load responses modelling and model generalisation in elite sports

Area of Science:

• Sport Science
• Performance Analytics
• Biostatistics

Background:

• Accurate sport performance modelling requires effective generalisation of predictive models.
• Understanding training load accumulation is crucial for optimising athlete performance.
• Existing models may not adequately generalise across individual athletes or training periods.

Purpose of the Study:

• To develop and evaluate a transferable methodology for sport performance modelling.
• To compare the generalisation and predictive accuracy of different modelling approaches.
• To identify optimal modelling strategies for elite short track speed skaters.

Main Methods:

• Collected training data from seven elite short track speed skaters over three months.
• Modeled cumulative training responses using impulse, serial, and bi-exponential functions.
• Compared variable dose-response (DR) models with Elastic Net (ENET), Principal Component Regression (PCR), and Random Forest (RF) using time-series cross-validation.

Main Results:

• Elastic Net (ENET) and Principal Component Regression (PCR) showed significantly better generalisation ability than the dose-response (DR) model.
• ENET and PCR models exhibited improved predictive accuracy compared to the DR model.
• ENET achieved the highest generalisation and predictive accuracy among all tested models.

Conclusions:

• Elastic Net (ENET) offers superior generalisation and predictive performance in sport performance modelling.
• A generalisation-enhancing procedure is essential for building and evaluating predictive models in sports science.
• The proposed methodology provides a transferable framework for analysing athlete training data.