A Model for Winning Survivor

The new season of Survivor airs next week. The author enjoys debating the show's strategy, such as the best approach to playing and who the greatest player is. Survivor is a competition where groups of strangers are stranded in a remote location and compete in challenges and vote each other out every three days. The last two or three contestants, known as the jury, vote for a winner who takes home a million dollars.

The author built a machine learning model to help analyze the show and predict who will win. The model predicts who is most likely to win or go home each episode. It can be found at a website that loads the most recent season by default, but also allows exploration of previous seasons. The model predicts win and elimination probabilities across the season and shows a cumulative plot of who has been ranked highest.

A "Player Breakdown" section lets you pick a player to understand which features affect their odds. The data for the model comes from the survivoR GitHub repository, which has information on voting history, challenge results, advantages, demographics, and edit metrics. Two separate models were trained: one predicting who will go home next and one predicting the season winner.

Logistic regression was used for the models, but fancier models performed worse. There are candidate features for the models, and the ones that survived into the Win model include specific factors. The Win model outperforms chance by roughly 2x, picking the eventual winner about 20% of the time. This is better than random guessing, but still limited by the unpredictable nature of the show.

The episode-to-episode rankings have small changes early on, but become more informative as the game progresses. The model has been useful for identifying contenders and tracking their trajectories while watching the show.