A carregar...
Publicação
Pacing strategies in 800m and 1500m freestyle : a data-driven analysis from the 2024 olympic games
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.92 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Compreender o ritmo dos nadadores de elite durante as competições é essencial para melhorar o seu desempenho em provas de longa distância. Este estudo explora as estratégias de ritmo adotadas pelos atletas nas provas de 800m e 1500m livres nos Jogos Olímpicos de 2024, com um foco específico na identificação dos fatores que explicam as variações da velocidade de natação e na avaliação de qual o modelo de machine learning que melhor as prevê. Inicialmente, foi considerada uma abordagem baseada na classificação, com o objetivo de prever perfis de ritmo a partir de características da prova. No entanto, devido ao tamanho e à natureza limitada do conjunto de dados, esta abordagem foi descartada. Como alternativa, adotou-se uma metodologia em duas etapas: (i) foram exploradas estratégias de ritmo através de agrupamento hierárquico aglomerativo; e (ii) foram utilizados modelos baseados em regressão para explicar e prever a velocidade do nadador ao longo da prova. A análise de agrupamentos revelou três perfis distintos de ritmo nos 800m, duas estratégias em forma de U (um mais rápido e outro mais lento) e uma estratégia de ritmo positivo, enquanto dois perfis em forma de U foram identificados nos 1500m. Os testes estatísticos confirmaram que estes agrupamentos estavam associados ao sexo, ao tempo de entrada e à variabilidade do ritmo (CV%), mas não à classificação final da prova. Para estudar os determinantes da velocidade, foram geradas novas variáveis, incluindo a aceleração, a distância até à chegada e o tempo do parcial anterior. A análise de importância das características identificou o sexo, a aceleração e o tempo de entrada como os preditores mais fortes. Entre os modelos testados, o Gradient Boosting apresentou o melhor desempenho preditivo, superando o Random Forests, as Redes Neurais e a regressão OLS. A análise dos resíduos, incluindo o teste de Durbin-Watson, confirmou a robustez estatística dos modelos.
Understanding how elite swimmers pace themselves during competition is essential for improving performance in long-distance events. This study explores the pacing strategies adopted by athletes in the 800m and 1500m freestyle races at the 2024 Olympic Games, with a particular focus on identifying the factors that explain variations in swimming velocity and evaluating which machine learning model best predicts it. Initially, a classification-based approach was considered, aiming to predict pacing profiles from race features. However, due to the limited size and nature of the dataset, this approach was discarded. As an alternative, a two-step methodology was adopted: (i) pacing strategies were explored through agglomerative hierarchical clustering; and (ii) regression-based models were used to explain and predict swimmer velocity throughout the race. The clustering analysis revealed three distinct pacing profiles in the 800m, two U-shaped patterns (one faster and one slower) and one positive-split strategy, while two U-shaped profiles were identified in the 1500m. Statistical tests confirmed that these clusters were associated with sex, entry time, and pacing variability (CV%), but not with final race ranking. To study the determinants of velocity, new variables, including acceleration, distance to the finish line, and previous split, were computed. Feature importance analysis identified sex, acceleration, and entry time as the strongest predictors. Among the models tested, Gradient Boosting revealed the best predictive performance, outperforming Random Forests, Neural Networks, and traditional OLS regression. Residual analysis, including the Durbin-Watson test, confirmed the statistical robustness of the models
Understanding how elite swimmers pace themselves during competition is essential for improving performance in long-distance events. This study explores the pacing strategies adopted by athletes in the 800m and 1500m freestyle races at the 2024 Olympic Games, with a particular focus on identifying the factors that explain variations in swimming velocity and evaluating which machine learning model best predicts it. Initially, a classification-based approach was considered, aiming to predict pacing profiles from race features. However, due to the limited size and nature of the dataset, this approach was discarded. As an alternative, a two-step methodology was adopted: (i) pacing strategies were explored through agglomerative hierarchical clustering; and (ii) regression-based models were used to explain and predict swimmer velocity throughout the race. The clustering analysis revealed three distinct pacing profiles in the 800m, two U-shaped patterns (one faster and one slower) and one positive-split strategy, while two U-shaped profiles were identified in the 1500m. Statistical tests confirmed that these clusters were associated with sex, entry time, and pacing variability (CV%), but not with final race ranking. To study the determinants of velocity, new variables, including acceleration, distance to the finish line, and previous split, were computed. Feature importance analysis identified sex, acceleration, and entry time as the strongest predictors. Among the models tested, Gradient Boosting revealed the best predictive performance, outperforming Random Forests, Neural Networks, and traditional OLS regression. Residual analysis, including the Durbin-Watson test, confirmed the statistical robustness of the models
Descrição
Trabalho Final de Mestrado, Métodos Quantitativos para a Decisão Económica e Empresarial, ISEG, 2025
Palavras-chave
Pacing strategy Swimming Velocity prediction Machine learning Olympic Games Regression models Estratégia de ritmo Natação Previsão de velocidade Machine Learning Jogos olímpicos Modelos de regressão