Predicting Olympic Medal Counts by Bootstrap-Enhanced Gradient Boosting Machine with Log-Transformed K-Means Clustering

Abstract

Olympic medal counts serve as a critical indicator of a nation's comprehensive sports strength. Accurate predictions offer distinctive perspectives for strategic resource allocation. To handle the strongly skewed distribution featuring long tails in medal data and capture complex nonlinear relationships among variables, this study proposes a hybrid methodology combining log-transformed K-means clustering and Bootstrap-Enhanced Gradient Boosting Machine (BEGBM). Countries are stratified into five strength tiers (e.g., elite to bottom) via K-means clustering on log-transformed medal features, reducing skewness bias. A robust BEGBM model trained on bootstrap samples (B=150) incorporates fluid sports variables and clustered tiers, with missing values imputed by random forests. The work presented here is the first to integrate tiered and mindful of uncertainty GBM into medal prediction. The model achieves RMSE=1.16 (gold)/1.91 (total), MAE=0.33/0.81, and =0.50/0.61. It further provides 95% CIs for the 2028 Los Angeles Olympics (e.g., USA: [35.8, 42.7] gold medals) and identifies nations with significantly changing medal potential. This data-based method offers sports authorities with a dependable tool for optimizing resource investment.