How it works

The scale

Every player has a hidden strength. The chance one wins a single game against another is their strength divided by the sum of both - the Bradley-Terry model. Ratings are the logarithm of that strength, scaled so that 100 rating points means 2:1 odds in a game. A 100-point favourite wins two games in three; a 200-point favourite wins four in five.

Where the numbers come from

Ghost reads finished tournament results - 77,370 matches across 2,497 events, 547,550 individual games. A match score is game-level data: a 9–5 win is nine games won and five lost. Fitting means finding the set of ratings that makes everything that actually happened as likely as possible.

Recent games count more

Games lose weight on a three-year half-life. A game from three years ago counts half, six years ago a quarter, and so on. Nothing ever fully expires, because an old result may be the only thing anchoring a rarely-seen opponent.

What is left out

Handicap events are excluded. The model assumes every recorded game is an honest trial of who wins a rack, and handicap formats spot games rather than playing them. Pyramid, snooker, blackball and straight pool are also excluded - a 100-point straight pool run is not a rack of nine-ball.

How much to trust a rating

Games is the time-weighted count behind a rating. Under 200 it is provisional and moves quickly.

Reach counts how many countries a player's opponents come from. It is useful but easily misread, and we learned that the hard way: a weekly tournament in a city like Amsterdam draws forty nationalities without anyone leaving the room, so players who had beaten nobody outside their own club were scoring higher on reach than world champions.

Depth is the share of a player's games in races to 7 or longer. Serious competition is played over long races; club nights are races to three, where one fluke rack is a third of the match. It is a blunter signal than reach and a far harder one to fake, because it describes the event rather than who happened to turn up.

The board therefore asks for a third of a player's games to come from long races. That is measured rather than picked: club records top out near 18% and professionals start near 47%, so the line goes in the gap. That is a judgement, not a law of nature - a genuinely excellent player who only plays short formats is excluded, and the full list is one click away. But the claim "these are the best" needs them to have been tested, and this is the most honest test the data supports.

Does it actually predict anything

Fitted on matches up to a cutoff date and tested on matches played after it, Ghost picks the winner about 68% of the time among players with a real record, and its stated probabilities land close to the truth: when it says 70%, the favourite wins about 70%. The per-game edge over a coin flip is small, and honestly so - one game of nine-ball between two tournament players really is close to a coin flip. The signal is in the race, not the rack.

Not FargoRate

Ghost is an independent implementation built from public tournament results. It takes no data from FargoRate and is not affiliated with them. The 100-points-is-2:1 convention is shared deliberately, so the numbers mean something to people who already know the sport.