A Model for Winning Survivor

This builds a machine-learning system that predicts, episode by episode, who will be eliminated next and who will ultimately win Survivor. It uses structured Survivor data (voting history, challenges, advantages, edit metrics) and trains two logistic-regression models: one for next-episode elimination and one for season winners. Key features in the Win model are times in danger (tribal councils with votes), a control for how many players remain, recent confessional share, age modeled quadratically (peaking around 30), number of previous seasons, and whether the player holds an advantage. The Elimination model uses remaining players, number of advantages held, votes against in the last three episodes, individual immunity win rate, and age. The models show winning and surviving are different outcomes, reflecting jury appeal versus simple vote avoidance.

Performance and case studies show practical value and limits. The Win model selects the ultimate winner roughly twice as often as random (about 20% vs ~10%), and its top-ranked player tends to win more often than their raw probability implies. Applied to Season 50, the model flagged Jonathan early, kept Aubry as a steady contender, and accurately adjusted odds after in-game moves (for example, Cirie’s played advantage raised her elimination risk). Feature analysis highlights that confessional share (the “winner’s edit”) strongly associates with winning, while being frequently in danger hurts prospects. Small sample size, editorial bias, and unquantified strategic factors constrain causal interpretation.