Rob Claessens - Predicting the Spring Classics of cycling with my first neural network

Built a neural network to predict cycling race rankings and optimize team selection for fantasy cycling game Scarito

Model used 16 input features including rider characteristics (age, weight, skills) and race characteristics (distance, vertical meters, difficulty) to predict continuous rank values

Form (recent race results) was found to be the most influential predictor, followed by vertical meters and hills skill rating

Initial model showed too much dependence on age (H) parameter, which was corrected through data augmentation and fine-tuning

Used SHAP values to analyze feature importance and interactions between different predictors

Benchmarked predictions against Cycling Oracle platform - fine-tuned model performed better than initial version but still trailed expert predictions by ~5%

Automated workflow using PyTorch datasets with ~2200 historical samples from 120 riders over 5 years of races

Model validation using rank correlation coefficient rather than R² due to ordinal nature of predictions

Tested model in real competition by running parallel accounts - one with human captain selection, one with AI predictions

Opportunities identified for improvement including better form modeling, feature engineering, and exploring alternative architectures like random forests