A Model for Winning Survivor
2 days ago
- The author built machine learning models (logistic regression) to predict Survivor winner and elimination probabilities episode by episode.
- Win model features include times in danger, confessional share, age (curved), number of previous seasons, and having an advantage; elimination model uses advantages held, votes against, immunity win rate, and age.
- The model performs about 2 times better than random chance (22.5% vs 10.1% baseline) in picking the eventual winner.
- During Season 50, the model flagged runner-up Jonathan as a contender early and correctly tracked in-game events like Cirie's odds dropping after using her advantage.
- Cross-season analysis ranks winners by dominance score (e.g., Boston Rob's S22 win most dominant) and identifies biggest upsets like Russell Hantz in S19.
- The model uses leave-one-season-out training for the app and includes caveats about missing social chemistry and limited data.