Clubs and agents can use it for scouting and player valuation, managers can use it for team selection, and anyone can use it to follow a current player's effectiveness on the field (paying customers get access to ~250,000 players globally).
As for "influence", our model measures the amount a given player impacts their team's odds of winning. Players with higher scores are more likely to tilt the odds in their team's favor when they're on the field.
We created this metric because most soccer models focus on discrete event statistics (passes, shots, expected goals, saves, blocks, etc) that paint an incomplete picture of a player's absolute quality.
There are some players (like the players currently in positions 3-6 in kxrank's top 100 list) who don't stand out statistically, but whose teams nevertheless are substantially more likely to win when they're playing versus when they're on the bench.
Does this answer your question(s)? Do you think we can document this better on the site?
On HN people often like to see the math. It's very understandable that you would keep the secret ingredient, well, secret. But readers are just looking at a opaque magic number now.
Of course, I can't just link the model code (even though I'd love to!), but it's not just for reasons of secrecy that we use an "opaque magic number". All semi-sophisticated ranking systems wind up being opaque magic numbers.
We can't, for example, express a player's impact in terms of raw odds, because those odds are also determined by the level of their teammates and the opposition players. For the same reason, we can't express this number in terms of goals or points or any other easily-mapped statistic.
So we need to use an abstract points system -- just as Elo and Fargo and other player rating systems use -- as a means of tracking and ranking player impact over time.
However, right before kickoff -- when all the players are known from both teams -- we can distill those ratings back down to odds/goal/point advantages. This is, in fact, how we wagered with the model.
4 comments
[ 3.6 ms ] story [ 26.2 ms ] threadAs for "influence", our model measures the amount a given player impacts their team's odds of winning. Players with higher scores are more likely to tilt the odds in their team's favor when they're on the field.
We created this metric because most soccer models focus on discrete event statistics (passes, shots, expected goals, saves, blocks, etc) that paint an incomplete picture of a player's absolute quality.
There are some players (like the players currently in positions 3-6 in kxrank's top 100 list) who don't stand out statistically, but whose teams nevertheless are substantially more likely to win when they're playing versus when they're on the bench.
Does this answer your question(s)? Do you think we can document this better on the site?
We can't, for example, express a player's impact in terms of raw odds, because those odds are also determined by the level of their teammates and the opposition players. For the same reason, we can't express this number in terms of goals or points or any other easily-mapped statistic.
So we need to use an abstract points system -- just as Elo and Fargo and other player rating systems use -- as a means of tracking and ranking player impact over time.
However, right before kickoff -- when all the players are known from both teams -- we can distill those ratings back down to odds/goal/point advantages. This is, in fact, how we wagered with the model.