Methodology
Last updated: 2026-10-02
PROBST (PRobabilistic Odds for Boots, Swaps and Tribals) forecasts the airing season: who goes home at the next tribal council, who wins immunity, whether the tribes swap or merge, and each player's odds of winning. The forecasts are made before each episode airs, from the data of the episodes before it. Nothing in them comes from previews, spoilers or the edit of the episode being forecast. It is the third model on the site, after SHALLOW (strategic rating) and BEAST (challenge rating), and it uses both as inputs.
Every forecast is built from the same records that power the SHALLOW and BEAST ratings:
From the votes, the model also builds a co-voting graph: how often each pair of players has voted the same way, out of the tribal councils they attended together. Voting blocs are found on that graph the same way the Allies section on player pages finds them.
The forecast is made in stages, each fitted on past seasons and tested on seasons it did not see.
1. What kind of episode. The chance of a tribe swap, a merge, 1 tribal council or 2, and 1 elimination or 2. Swaps and merges are read from how often past seasons did the same thing at the same point, weighted toward recent seasons. Double eliminations are read from the pace the season needs to keep to finish on time.
2. Which tribe loses. Before the merge, the chance each tribe finishes last in the immunity challenge, from the tribe's average BEAST rating and how it has done in challenges this season. With 2 tribes this is close to a coin flip.
3. Who wins immunity. After the merge, each player's chance of winning individual immunity, from career BEAST rating, gender and age.
4. Who the votes land on. Among the players at tribal council who are not immune, the chance the most votes fall on each one. This is the main model. It weighs 24 features: how many votes the player has received this season and at the last tribal council, whether they voted for the player who stayed, how many tribal councils they have attended, their share of confessionals, whether they hold an idol, whether they have played before, their ratings relative to the tribe, whether their original tribe is in the majority, and 5 measures of where they sit in the co-voting graph: the share of the room they have voted with, the strength of those ties, the size of their bloc, whether it is the largest, and whether their ties connect blocs that otherwise do not vote together.
5. Idols. A player holding an idol who is the target of the vote plays it and survives about a third of the time, based on every idol on record. When that happens, the player with the second most votes goes home. Idols are also found during the season, at the rate history shows for that point in the season and the number of idols already in play, and they are sometimes played when they were not needed.
6. Who goes home. Steps 4 and 5 together give each player's chance of leaving at the next tribal council. The number shown on the page combines this with the chance of each tribe going to tribal council, so every player alive has one number for the episode.
Under each player's number the page names the features that moved it most, relative to the rest of the tribe. A feature is named only when the model is sure of its effect for that player: features are grouped into themes (voting record, alliances, challenge threat, strategic rating, airtime, idol, original tribe, experience), and a theme is named only when its total effect is at least one and a half standard errors from zero. A theme the model is not sure of still counts in the number but is shown in grey and is not given as a reason. The SHALLOW rating is the usual case: its weight in the fit is small and uncertain, so it is rarely quoted as a reason even though it is in the model.
The model does not watch the episodes. A player who is safe because of a deal made off camera looks exactly like a player who is not.
The season odds come from simulating the rest of the season 10,000 times. Each simulated episode draws the format from step 1, plays the challenges and moves the challenge ratings, finds and plays idols, and votes someone out using step 4. Fire-making at the final 4 is drawn once per simulated season. The final tribal council is decided from airtime, the only signal that separates winners from the other finalists in past seasons. The weight on airtime is fitted on the 422 jury votes cast in 50 seasons: each juror chose one finalist, and a finalist's chance of winning is the share of the jury the fit expects them to get. The voting record stays as it was when the simulation started, and its weight is halved each simulated week, because who was on the wrong side of the vote today says less about a vote 2 months out.
Every number below comes from forecasting a past season with models trained only on the seasons before it, for US seasons 11 to 50. The same forecasts are shown episode by episode on the Episode Snapshot, under "Before this episode", for the airing season and for every one of those past seasons.
Who goes home, given who is at tribal council (590 tribal councils): the model's first pick was right 29% of the time. Picking at random would be right 17% of the time. The player who left was in the model's top 3 62% of the time, against 51% at random.
Who goes home, a week ahead, with nothing given (584 boots): format, immunity and the vote all forecast. First pick right 18% of the time (11% at random); the player who left in the top 3 45% (32%) and in the top 5 64% (52%).
Which tribe loses: 50% against 44% at random. Most of that comes from seasons with 3 tribes. Who wins immunity: first pick right 26% (16%); top 3 61% (47%). When the merge comes: for the 20 seasons from Cambodia on, the most likely episode was exactly right in 15 and never off by more than one.
Who wins the season (40 seasons, from the state going into the merge): the eventual winner was in the model's top 5 60% of the time, against 43% at random. Players given a 13% chance won 11% of the time; players given a 4% chance won 2.3% of the time. Before the season starts the odds are close to even, and they stay close to even until players have voted.
PROBST knows only what the aired episodes have shown. Betting markets on Survivor are moved by people who know more than that: the season was filmed months before it airs, the cast and crew know the result, and leaks reach the markets long before the edit does. A market price will be ahead of this forecast whenever there is inside information, which is most of the time. The numbers here are for reading the game as it has been shown, not for wagering.
The full method, backtests and parameters are documented in the project wiki, and the code is in the survivor_sim package of the repository.
See also: SHALLOW strategic rating methodology, BEAST challenge rating methodology, and the FAQ.