Predicting the Spring Classics of cycling with my first neural network
summary
A three‑layer neural network implemented in PyTorch trains on about 2200 samples (16 Classics, 5 years) to predict a continuous rank and reaches roughly Spearman 0.5 after tuning. SHAP analysis shows form and vertical meters matter most, an age bias inflates predictions until corrected by shuffling age and post‑training, improving Spearman by ~5%. Benchmarking against Cycling Oracle shows a remaining ~5% gap, and the human captains beat the robot by about 5% overall.
AI-written summary · about
chapters (9)
- 0:00Predicting cycling Classics
- 3:34Neural network setup
- 8:12Data collection & preprocessing
- 13:33Training & fine-tuning
- 16:23Model performance with SHAP
- 24:13Benchmarking vs cycling Oracle
- 26:40Rob versus Robot results
- 28:03Conclusions & future work
- 29:36Q&A
related talks
32mHacking Fitness
27mPredicting F1 Race Strategies using ML.NET
30mApplying The Magic Of Neural Networks
39mForecasting With Classical and Machine Learning Methods
38mDeep Learning plays Handball
42mHow Deep Learning is Revolutionizing Sports Analytics
30mShift your ML model into overdrive with AWS DeepRacer analysis tools
34mPredicting the Moment of Birth using Sensor Data in Dairy Cows