Margin or Scoring Run: What Matters Most at the Final Change?
/Commentators are keen to point out how especially important they feel are the minutes just before a change - especially before the final change - so today we’re going to investigate how our in-running estimate of a team’s victory probability at three-quarter time should be influenced by any streaks of scoring leading up to the break.
THE DATA
To do this we’ll be using:
Score Progression data from AFLTables for the period 2001 to 2021
MoSHBODS pre-game Team Ratings for those same games
THE MODEL
Altogether we have data for 4,016 games, and we’ll be fitting a binary logit of the form:
P(Home Team Wins) = 1/(1+exp(-(a0 + a1 x Home Team Lead at 3QT + a2 x Pre-Game Rating Difference + a3 x Home Team Scoring Run Leading into 3QT))
where
Home Team Lead at 3QT is the difference between the home team’s and the away team’s scores at 3QT. It will be positive if the home team leads, negative if they trail
Pre-Game Rating Difference is the difference between MoSHBODS’ pre-game home team Rating and away team Rating
Home Team Scoring Run Leading into 3QT is the number of unanswered points recorded by the home team leading into the break. It will be positive if the home team has the unanswered streak, and negative if the away team does. So, for example, if the away team had kicked the last 2 goals and 2 behinds leading into 3QT, this variable would be -14
When we fit this model we obtain the following coefficients and p-values.
Intercept: 0.1929 (0.00043 ***)
Home Team Lead at 3QT: 0.1148 (tiny ***)
Pre-Game Rating Difference: 0.0291 (tiny ***)
Home Team Scoring Run Leading into 3QT: 0.0028 (0.60565)
Notwithstanding that it’s an in-sample performance metric, this model correctly predicts the winner at 3QT in 88% of the games. If we exclude the Home Team Scoring Run variable, we actually end up with a very slightly more accurate model (it gets 2 more games right).
THE INTERPRETATION
The coefficient we’re interested in is the last one, that for Home Team Scoring Run Leading into 3QT, and it is both small in absolute magnitude and statistically not significantly different from zero.
Our best estimate of it is, however, positive, which does mean that a home team scoring streak running up to 3QT boosts the home team’s estimated chances, and an away team scoring streak running up to 3QT lowers the home team’s estimated chances. But, by how much?
Here are a few scenarios:
The home team, which started as a 10 point weaker team, according to MoSHBODS, trails by 20 points at 3QT having recorded the last 12 points of the quarter. Their victory chances are estimated to be 8.6%. Had they, instead, conceded the last 12 points but still gone in trailing by 20 points, their estimated victory chances would be 8.1% or just 0.5% points lower.
The home team, which started as a 15 point stronger team, according to MoSHBODS, leads by 10 points at 3QT having recorded the last 8 points of the quarter. Their victory chances are estimated to be 85.8%. Had they, instead, conceded the last 8 points but still gone in leading by 10 points, their estimated victory chances would be 85.3% or just 0.6% points lower.
The home team, which started rated equally with the away team, according to MoSHBODS, find themselves level at 3QT having recorded the last 10 points of the quarter. Their victory chances are estimated to be 55.5%. Had they, instead, conceded the last 10 points but still gone in on level terms, their estimated victory chances would be 54.1% or just 1.4% points lower.
In short, if there is any signal in a team’s end-of-third-quarter scoring, it’s very small. Statistically, in fact, we can’t reject the hypothesis that it’s zero once we account for the far more important variables, which are the actual margin at 3QT, and the difference in the pre-game estimated abilities of the teams.
It could be that a streak of scoring before the final change hints at some on-the-day effects that will persist into the final term (dare I say, momentum), but it might also just be noise.
