Using Case-Based Reasoning to Predict Marathon Performance and Recommend Tailored Training Plans

Marathon training advice is often one-size-fits-all, even though runners’ fitness, effort patterns, and goals vary widely. A case-based reasoning system built from Strava training data aims to fix that by predicting marathon finish time and recommending the next week of training using patterns from similar past runners—without requiring lab tests or prior marathon performances. The approach matters because it can give novice and recreational runners actionable guidance throughout the 12–16 weeks leading to race day, including when training is already underway.

The system is trained on more than 1.5 million training sessions from over 21,000 runners, covering the 16-week lead-up to a marathon. Each runner’s activity is converted into 100-meter pacing profiles (minutes per kilometer) with elevation information, then aggregated into weekly features such as number of sessions, total distance, mean pace, longest run, fastest/slowest 1.5 km and 10 km paces, and mean elevation gain. Cases are defined by a runner’s weekly training features plus the marathon time achieved a fixed number of weeks later, with cases separated by training week and sex to reflect physiological differences and to avoid recommending training for the wrong point in the cycle.

For race time prediction, the method retrieves the k most similar historical cases (using Euclidean distance over the weekly feature space) and averages their known marathon times. Three model variants are compared: a baseline using all features, a feature-selected model using stepwise forward selection to keep only the most predictive variables for each week, and a multi-week model that improves on earlier ensemble strategies by incorporating past predicted times as additional inputs. Across men and women, prediction error generally drops as race day approaches, but the baseline and feature-selected models show a sharp error spike around weeks one to two before the marathon—consistent with the “marathon taper,” when runners reduce or stop training and variability rises. The multi-week model largely removes that problem, delivering the best overall performance and achieving statistically significant gains (p < 0.01) about 12 weeks out compared with the other approaches.

Training plan recommendation uses a different mechanism: given a runner’s current training week and an adjusted goal time (marathon time star, represented via a delta), the system filters historical runners who hit similar goal times and match the current training stage. It then selects a next-week training profile by targeting the “middle” of the retrieved set’s next-week loads—avoiding recommendations that are unrealistically hard or too easy that could result from relying on a single closest match. Evaluation uses 10-fold cross-validation on a subset of roughly 20,000 runners (focusing on marathon times under five hours for men and between three and five hours for women across Dublin, London, and New York City marathons from 2014–2017). The recommended plans show encouraging proof-of-concept behavior: aiming for faster goals (delta < 1) yields recommended weeks with faster mean weekly pace than the runner’s actual plan, while slower goals (delta > 1) produce easier pace targets.

Beyond accuracy, the system’s practical value is explainability. By showing how a runner’s recommended plan compares to the distribution of the 15 retrieved similar cases (sessions, distances, and pace metrics), it provides a traceable rationale that many prediction-only tools lack. The work concludes with plans for a live user study and future extensions—such as adding heart-rate data or using time-series methods to classify training sessions—and notes that the framework could transfer to other endurance sports if comparable training data exists.