Predicting Marathon Finishing Times Using Ensemble Learning: An Empirical Study on Boston Marathon Data

VERSION OF RECORD ONLINE: 11/09/2025

Abstract

This study proposes an ensemble machine learning model to predict marathon finishing times, using empirical data from the Boston Marathon spanning 2015–2017. After thorough preprocessing and feature engineering—including intermediate checkpoint times (5K, 10K, Half Marathon), age, gender, nationality, and year of participation—six models were implemented and evaluated: K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), Case-Based Reasoning (CBR), a prior benchmark model (FA-PP-R-ML), Long Short-Term Memory (LSTM), and a novel ensemble model combining Linear Regression, Random Forest, and MLPRegressor via a meta-learning approach. Experimental results on the test set demonstrate that the proposed ensemble model achieved the highest predictive performance, with a Mean Absolute Error (MAE) of 7.32 minutes, Root Mean Squared Error (RMSE) of 11.06 minutes, and R² score of 0.928—outperforming all baseline models in both accuracy and robustness. Visualization techniques such as scatter plots and boxplots further confirmed the model’s high agreement between predicted and actual values. Nevertheless, the study acknowledges several limitations, including a constrained dataset limited to three years of a single event, a narrow scope of model comparison, simplifications in algorithmic assumptions, and limited hyperparameter tuning. Future work should explore more diverse datasets, incorporate exogenous factors (e.g., weather, elevation), adopt advanced modeling techniques such as attention mechanisms, graph-based learning, or AutoML, and enhance model interpretability to support real-world applications in athlete coaching and performance forecasting.