Hacking Fitness
summary
A probabilistic pipeline predicts amateur race speed from noisy watch GPS and heart‑rate data by fitting per-activity heart‑rate vs speed curves, bootstrapping samples, and feeding three intensity features into a hierarchical Bayesian model that outputs full predictive distributions rather than points. Classification of races uses a random forest on Strava labels, model fitting uses PyMC3, and partial pooling avoids massive overfitting seen when modeling 10,000 parameters for 230 users on ~1,600 observations while producing useful probability intervals that match held‑out races.
AI-written summary · about
chapters (11)
- 0:00Performance prediction in sports
- 3:14VO2 max limitations
- 7:57Training data reliability
- 11:10Labeling races in Strava
- 14:04Fitting curves to training data
- 17:21Probabilistic models
- 21:00Pooled model for prediction
- 25:24Hierarchical models
- 28:16Balancing complexity vs overfitting
- 30:31Personalized predictions
- 31:39Conclusion & Q&A
related talks
30mShift your ML model into overdrive with AWS DeepRacer analysis tools
28mShift your ML model into overdrive with AWS DeepRacer analysis tools
28mShift your ML model into overdrive with AWS DeepRacer analysis tools
34mPredicting the Spring Classics of cycling with my first neural network
22mRunning, walking, sitting or biking? - Motion prediction with acceleration and rotation
33mAnalytics in the age of the Internet of Things #bbuzz
27mPredicting F1 Race Strategies using ML.NET
37mCreating correct and capable classifiers