DeepRace: GRU-Based Sequence Modeling Framework for Marathon Performance Prediction

Abstract

In recent years, data-driven approaches have gained prominence in sports analytics, offering new opportunities for optimizing athletic performance. This study presents DeepRace, a computational framework based on deep learning models including Recurrent Neural Networks (RNN), Long Short-Term Memory networks (LSTM), and Gated Recurrent Units (GRU) for marathon performance prediction. Using a dataset of 23,763 training activities from 549 athletes who participated in the 2017 Boston Marathon, we evaluated the effectiveness of sequence-based models in forecasting marathon finish times. The framework systematically preprocesses and analyzes sequential workout data, employing a multimodel selection algorithm to identify the most accurate predictive model. Experimental results demonstrate that the GRU model achieved superior performance, with a Mean Absolute Percentage Error (MAPE) of 5.2 percent and a Root Mean Squared Error (RMSE) of 0.11 hours, outperforming both RNN (MAPE 5.6 %) and LSTM (MAPE 6.5 percent) models. Notably, all deep learning models significantly surpassed the baseline linear regression model (MAPE 10%), confirming the advantage of deep learning for time-series performance prediction. These findings underscore the potential of GRU-based sequence modeling for enhancing marathon training strategies, offering actionable insights for athletes and coaches aiming to improve race outcomes through data-informed decisions.