warwick.ac.uk/dfirth  @firthstat          alt-3.uk  @footymath
The Special One!
Talking in March 2014, when Chelsea topped the table on points but had played more matches than all of their nearest rivals…
As at 17 March 2014, the top 4:
(Source: footstats.co.uk)
How to make a league table that is ‘less fake’ than the standard one?
Can we solve ‘the pandemic problem’??
Football league administrators everywhere in a panic.
How to finish the 2019–2020 football season?
Final positions in French Ligue 1 and Ligue 2 awarded on the basis of points per match.
rules agreed in advance!
fairness
transparency
(as was used in France in April 2020, as an ad hoc solution)
\[ \textrm{rate} = {\textrm{points won} \over \textrm{matches played}} \]
This at least would help to address the ‘games in hand’ problem.
But we can do better!
The key is to recognise that some matches are harder than others. Not all matches are equal, when ‘weighed’ as opportunities to gain league points.
An example: Premier League after 5 matches
September 2017
Everton had won one match, and drawn one (of the five played).
The official table had Everton in 18th place.
Facts neglected in the official table:
The alt-3 method
Adjusted league table, fully respecting:
3 points for a win and 1 for a draw
double round-robin balance, home/away
The principles of a good solution are best considered in a simpler setting.
Let’s simplify, for now, by considering a single round-robin tournament with only binary match outcomes (one point for a win).
Consider a generalized version of the points-per-match rate:
\[ \textrm{rate} = {\textrm{points won} \over \textit{effective } \textrm{matches played}} \]
with
\[ \textrm{effective matches played by team } i = \sum_{j \textrm{ played}} e_{ij} / \bar{e}_{i.}, \] with the notion that \(e_{ij}\) is a probability that \(i\) beats \(j\) (or equivalently expected points for \(i\) when playing \(j\)).
The simple ‘points per match’ measure is the special case with all of the \(e_{ij}\) equal to \(1/2\).
Simple points per match has a very appealing aspect: probabilities \(e_{ij} = 1/2\) imply no distinction between teams.
But the assumption of \(e_{ij} = 1/2\) for all matches is (almost always) in conflict with match results already known.
This is resolved by seeking probabilities \(e_{ij}\) which are as uninformative as possible, while preserving agreement with match results already seen.
Consider first the situation at the end of the season.
Given only the points totals \(p_i\) for all teams, what is the most uninformative (or ‘least prejudiced’) set of probabilities \(e_{ij}\) that matches those points totals exactly, i.e., such that
\[ \sum_j e_{ij} = p_i, \textrm{ for all } i \] ?
A natural mathematical formulation of ‘most uninformative’ is to maximize the total entropy of probabilities \(e_{ij}\), \[ \textrm{total entropy} = - \sum_{i, j} [e_{ij}\log e_{ij} + (1 - e_{ij})\log(1 - e_{ij})] \]
Under the (essential!) constraint that end-of-season points totals are matched exactly, entropy is maximized by
\[ e_{ij} = {s_i \over s_i + s_j} \] with the \(s_1,s_2,\ldots\) determined uniquely to match the team-by-team points totals.
This solution is well known to statisticians as the Bradley-Terry model (with its parameters determined by maximum likelihood).
And the connection with maximum entropy is also well known (e.g., Joe, 1988, Annals of Statistics).
Hence, during the season: use probabilities \(e_{ij} = t_i / (t_i + t_j)\), with the ‘team strengths’ \(t_i\) determined uniquely by points totals to date.
(sometimes called a ‘retrodictive’ Bradley-Terry model)
Not so well known?
The second of these is an essential fairness property in round-robin leagues.
An evolution:
Bradley-Terry (1952; and Zermelo much earlier!) — \(s_i\) : \(s_j\) : \(0\)
Davidson (1970): 2 for a win, 1 for a draw — \(s_i\) : \(s_j\) : \(\delta (s_i s_j)^{1/2}\)
Davidson and Beaver (1977): incorporate home advantage — \(\gamma s_i\) : \(s_j\) : \(\delta (\gamma s_i s_j)^{1/2}\)
alt-3 (2017–2021): 3 for a win, and separate home/away strengths — \(h_i\) : \(a_j\) : \(\delta (h_i a_j)^{1/3}\)
The alt-3 method comes straightforwardly from the same maximum-entropy argument as before, but matching 3-for-a-win points totals (home and away).
Sep/Oct 2017
After 5 matches played: official table 18th; alt-3 table 11th
After 9 matches played: Ronald Koeman sacked
(Everton finished the season in 8th place, under Sam Allardyce. Koeman went on to be the Netherlands national coach, and — until very recently — manager of Barcelona.)
6 Oct 2019
Straight eight (wins from start of season):
Premier League (England)
Bundesliga 1 (Germany)