A simple question: how bad can a team have been in the previous season and still harbour realistic hopes of playing in the current season's Grand Final?
That's the topic we'll explore today, using teams' previous season-end MoSSBODS Ratings to quantify their ability, and footy history to contextualise and quantify their chances.
Firstly, a chart, showing the Offensive and Defensive Ratings of all 236 Premiers and Runners Up from the 1898 through 2015 seasons. (The 1897 Premiers and Runners Up, having no "previous season" are, of necessity, excluded.) Points on the chart are labelled with a short-form team name and year, coloured to reflect whether they were Premiers (red) or Runners Up (orange) in that year, and positioned based on the team's Offensive and Defensive Ratings at the end of the previous season.
What strikes immediately is the relative sparsity of points in the lower left of the chart where both the Offensive and Defensive Ratings are negative. Just 9 teams sit in that space, and only two of them are Premiers. One of those Premiers is the 1997 Adelaide team, who we looked at in the previous blog because they'd registered the 10th highest all-time season-to-season Defensive Rating increase, and the other is the 1993 Essendon team.
The Dons finished 8th in 1992, missing a spot in the six-team Finals by just a couple of wins, though their end of season percentage of 92.8 ranked only 9th-highest in that year. That percentage, and their Rating, was significantly dented by a 160-point and 40 Scoring Shot loss to the Hawks in Round 20.
In total, 16 Premiers and 29 Runners Up have finished the previous season with a negative Offensive Rating, and 13 Premiers and 21 Runners Up have finished with a negative Defensive Rating. As well, 8 Premiers and 19 Runners Up have finished the previous season with a negative Combined Rating.
And, it's Combined Rating that we'll use for the table that follows, which reveals the 25 Premiers and Runners Up whose Combined Rating was lowest in the previous season, and the Combined Rating of every other Premier and Runner Up since and including the 2000 season. The list is ordered on the basis of a team's Combined Rating, with the lowest Rating at the top and numbered 1.
The first two teams on this list are familiar to us from the previous blog post, but many of the rest are new. It interesting to note that only two of the teams in the Top 25 managed to lift their Combined Rating above +6 in the year of their success, those two exceptions being the 2007 Geelong team (also in the previous blog) and the 1980 Richmond team.
Also notable is the fact that most of the teams in the Top 25 on the Combined measure are also in the Top 80 or so in terms of both their Offensive and their Defensive Ratings (see the second- and third-last columns). The 1961 Footscray team is the most-prominent exception to this general trend, they having the lowest Offensive Rating in their previous season of all Premiers and Runners Up, but only the 121st-lowest Defensive Rating. The 1962 Essendon and 1996 Sydney teams are of the opposite sort, with exceptionally low prior-year Defensive Ratings, but unexceptional prior-year Offensive Ratings.
At the bottom of the table, where the entries relate to the Grand Finalists of the most-recent season, it's interesting to note the 2005 and 2006 Sydney teams, which have, respectively, the 2nd- and 13th-lowest end-of-previous-season Offensive Ratings of all-time.
NAIVE MODEL FOR 2016 PROBABILITIES
What then about the chances for the 18 teams in 2016?
We can take the entire history of the 1,413 teams for which we have "this season" and "last season" Offensive and Defensive Rating data and use our knowledge of whether they finished as Premier, Runner Up, or neither, to create a range of classification models.
Once fitted, those models can take as input the final MoSSBODS Ratings of the 2015 teams (see table at right) and provide probabilities for each team about their 2016 Grand Final prospects.
Though I fitted quite a few more models, ultimately I settled on the outputs of just three, selected because they were simultaneously different enough to be interesting but not so outrageously different that it seemed they'd been overfit, undertuned, or somehow otherwise rendered unhelpful.
The table below summarises the 2016 predictions for all three of the selected models, and also provides a simple weighted average of their predictions.
Looking especially at the Simple Average, it's clear that the teams fall naturally into four groups:
- West Coast and Hawthorn: each with Premiership chances of around 20%
- Sydney, North Melbourne, Adelaide, Port Adelaide, Western Bulldogs and Richmond: each with Premiership chances in the 6% to 9% range
- Geelong, Fremantle and Collingwood: each with Premiership chances in the 3% to 4% range
- GWS, St Kilda, Melbourne, Essendon, Brisbane Lions and Gold Coast: each with chances of about 1% or less
As you'd expect - unless you believed that Offensive and Defensive Ratings might carry significantly different weights in the models - the ordering of the teams in this table is virtually identical to the ordering of the teams based on their 2015 Combined Ratings. Only Fremantle and Geelong swap places. Given that, the validity of these estimates is inextricably linked to the efficacy of the MoSSBODS method, which, I'll remind you, knows nothing of list changes or schedule imbalances.
Still, it seems the results provide a reasonable first approximation. A quick comparison with the current TAB Flag market reveals that the largest differences of opinion are for Adelaide, where the TAB thinks less of them than does MoSSBODS, and for Geelong and Fremantle, where the opposite is true. Otherwise, the team orderings of the TAB and MoSSBODS are quite similar.
A QUICK TECHNICAL NOTE
For those curious about some of the technical specifics of the modelling, I should point out that I used three regressors in each model: the teams' Offensive and Defensive Ratings as at the end of 2015, and an Era variable - parameterised as a series of dummy variables to keep some of the model types happy - that grouped seasons into the buckets 1897-1919, 1920-1939, 1940-1959, 1960-1979, 1980-1999 and 2000-2015. I used the caret package in R to fit and tune all models, and had no holdout sample. For tuning purposes I used 5-fold cross-validation - so overfitting has probably been somewhat ameliorated. (My original intention was to create an ensemble model and use a holdout sample to assess its accuracy, but I subsequently discovered that the caretEnsemble package does not yet support multi-class classification problems, which is what we have here with our Premier/RunnerUp/Neither target variable.)
Also, I needed to standardise the probabilities output by the models for the 2016 season because, while they ensured that each team's probabilities summed to one across the three possible outcomes, they did not ensure that the sum of probabilities across the teams for the Premiership summed to 1, for Runner Up summed to 1, and for Neither summed to 16. The standardisation involved first dividing each team's estimated Premiership probability by the sum of those probabilities, then doing the same for the Runner Up probabilities, then calculating the Neither probability for each team as 1 - Premiership probability - Runner Up probability.