Traditional exercise science relies on single-dimensional physiological metrics such as VO2Max to predict athletic performance on, e.g., running events such as a marathon. These metrics are often unreliable and inconvenient for amateur athletes because they require laboratory tests. Given the commodification of "smart" activity trackers that can measure speed and heart rate, a computational approach to performance prediction is proposed, i.e., one that does not require the athlete to perform a lab test.

Compared to exercise data from elite athletes, there is more inherent noise and uncertainty in data from amateurs for various reasons such as cheap devices, the sparsity of activities, or variance in "form of the day." Therefore, a probabilistic model was used to predict the performance of amateur athletes for running events of several distances.

This talk will provide details about the modeling process using the probabilistic framework PyMC3. In particular, it will be shown how potential models and priors are specified and checked before seeing any data, how (Bayesian) model fitting is diagnosed, and how such a probabilistic model is used to make predictions (since the latter is not as straightforward as, e.g., Scikit-learn's estimator.predict(X)).