Can Garmin Data Predict Ultra-Trail Race Outcomes? A Statistical Modeling Approach Case Study

Using fundamental Garmin endurance training data, including metrics such as heart rate (HR), moving time, and elevation changes, it is possible to identify and predict factors that contributed to successful race outcomes. PURPOSE: We aimed to evaluate the effectiveness of advanced statistical techniques in predicting race performance for a high-performing, male ultra-trail runner (42 yrs old) using data collected from a Garmin running watch. METHODS: A total of 1,400 Garmin files, collected between 8/21/2018 and 9/15/2024, provided metrics such as HR, pace, cadence, elevation change, and distance. These files were summarized, and the aggregated results compiled along with race outcomes, including placement and podium finish (i.e., top “3” overall). RStudio (version 2024 build 375) was used to develop R scripts for model implementation. Several statistical models were applied, including K Nearest Neighbor (KNN), k-means clustering, multiple forms of linear regression, LASSO, Ridge Regression, Elastic Net, and Random Forest. All models, except for KNN and k-means clustering, were subjected to 10-fold cross-validation to calculate Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE). KNN and k-means clustering models were used to predict podium position. Other models predicted race placement. RESULTS: The pruned decision tree (PDT) demonstrated the best performance (MAE = 2.83, RMSE = 6.28), indicating its ability to predict race outcomes accurately with fewer predictors. The Random Forest (MAE = 3.31, RMSE = 6.99) KNN and k-means clustering model showed the worst results. The optimized PDT consisted of four terminal nodes based on decisions nodes for moving time (< 1948 sec), maximum HR (< 168.5 bpm), and total ascent (< 108.5 ft) that helped determine race outcomes. Additionally, calories and average HR were found to be consistently influential in several models. CONCLUSIONS: While achieving an MAE of approximately three places may seem promising, these results must be interpreted cautiously, particularly for practical coaching applications. Despite the model's apparent accuracy, the findings indicate that training sessions exceeding 33 minutes predicted an average placement of around 2.8, whereas shorter, less intense training (Max HR < 168.5 bpm) predicted a placement of around 1.1.