Historical Team Rating Trajectories (1970 to 2019)
/Over the past couple of blogs, we’ve been analysing historical scoring progressions to come up with archetypical game types in terms of the ebb-and-flow of the game margin.
To do that, we treated the score progressions as time series data and today we’ll do something similar with teams’ season-by-season historical MoSH2020 Team Ratings for the period 1970 to 2019, inclusive.
THE HISTORICAL DATA
As a first analysis, let’s chart the season-by-season picture for all teams. Each grey line in each panel tracks a single team (that missed the Grand Final of that year), while the blue line tracks the Premiers, and the red line tracks the Runners Up. Each team is tracked as at the start of the season (R0), the end of the first 22 rounds of the home-and-away season R1 to R22), and at the end of the Grand Final (GF). Team ratings carry through from week to week within a season, so teams that missed the Grand Final in a particular year will have a final rating equal to whatever it was at the end of whichever was their last game of that season.
One interesting aspect of this chart is how early in their respective seasons most eventual grand finalists find themselves with a large, positive rating. No Premier has finished a season with a Combined Rating below 11, and only three Runners Up have finished a season with a Combined Rating below 5. Put simply, good teams usually make the Grand Final, and very good or exceptional ones tend to win it.
SEASON COMPETITIVENESS
We can take the Combined Ratings of all the teams at the end of a particular round and measure how spread out they are by using a statistical measure of dispersion around their mean (which, because of how MoSH2020 works, will always be zero). Two candidate measures are the standard deviation and the mean absolute deviation (MAD). We’ll use the latter, which is calculated here simply by averaging the absolute team ratings since their mean is always zero. The smaller the MAD, the less dispersed are the ratings, which we might interpret as the more competitive is the competition as a whole.
In the chart below we show how the MAD has tracked across and within each of the seasons from 1970 to 2019.
We see that, in most seasons, our MAD metric on MoSH2020 ratings suggests that competitiveness lessens as the season progresses. This might, at least in part, be attributed to the mechanics of the MoSH2020 System, which drags all team start-of-season ratings towards zero relative to what they were at the end of the previous season.
To the extent that each team will tend to revert to its previous end-of-season rating, the dispersion of all ratings will tend to revert to its previous end-of-season level.
That said, bookmaker price data supports the contention that games in later rounds tend to be generally less competitive than games in earlier rounds, as we can see from the chart at right, which is based on TAB mid-week prices for home-and-away games over the 14-year period 2006 to 2019.
In terms of actual prices, it means that we would expect the average favourite in Round 1 to be priced at around $1.45, and in Round 22 to be priced at around $1.33 (assuming a roughly 5% overround, evenly levied on the favourite and on the underdog).
One other interesting aspect of the chart above is that it implies the 2019 season was one of, if not the, most competitive seasons amongst the 50 that are charted.
CLUSTERING TEAM SEASON RATING TRAJECTORIES
In the previous blog we took the scoring progressions for just under 2,000 games and used the TSrepr R package to dimension-reduce and then cluster them.
Here we’ll do the same thing, using teams’ ratings trajectories across whole seasons as the individual time series. By using the ratings at the end of Rounds 0 to 22 and at the end of the Grand Final, we ensure that all time series are of the same length for every team in every season, which is required for the methodology we’re employing.
For the technically curious, this time we’re using
repr_matrix(RatingsTSData, func = repr_paa, args = list(q = 7, func = medianC), normalise = FALSE, func_norm = norm_z, windowing = FALSE, win_size = NULL)
for the dimension reducing step, which reduces the original ratings time series dataframe from 698 x 24 to 698 x 4.
When we pam cluster the resulting matrix, there are a number of solutions with almost the same highest Davies Bouldin value, so we choose the solution with the smallest number of clusters, which is 15, the solution for which is shown below. Each grey line is the ratings trajectory for a single team in that cluster. The thick red line tracks the median rating, and the two thinner red lines define the interquartile range. The blue line tracks the rating of the team that follows the median most closely across the season (based on its mean absolute deviation from the median averaged across all 24 points).
Those blue-line producing, archetypal teams are listed below (though, like most cluster solutions, no single observation tends to completely mirror the median of all the observations in its cluster).
Cluster 1: Above-Average Team Drifts Slightly Upwards
Geelong 2018: soundly-beaten Preliminary Finalists in 2017, and beaten Elimination Finalists in 2018 after back-to-back 100-point wins in their last two home and away gamesCluster 2: Above-Average Team Has Very Good Season
Carlton 2011: finished 8th of 16 in 2010, and 5th of 16 in 2011, losing by just 3 points in a Semi FinalCluster 3: Average Team Becomes Poor
Sydney 2009: finished 6th of 16 in 2008, losing in a Semi Final, and 12th of 16 in 2009Cluster 4: Below-Average Team Remains Below-Average
Fitzroy 1970: finished 10th of 12 in 1969, and 9th of 12 in 1970Cluster 5: Above-Average Team Drifts Slightly Downwards
West Coast 2017: beaten Elimination Finalist in 2016 after finishing 6th, and beaten Semi Finalist in 2017 after finishing 8thCluster 6: Below-Average Team Consistently Improves
Fitzroy 1981: finished last in 1980 but won 1 and lost 5 of the last six games all by less than 3 goals, eventually scoring 36 more points than the Minor Premiers (but conceding 900 more). Finished 5th in 1981, eventually going out by 1 point in a Semi Final.Cluster 7: Below-Average Team Remains Below-Average
Collingwood 1999: finished 14th of 16 in 1998, and last of 16 in 1999Cluster 8: Average Team Becomes Poor
Fitzroy 1994: finished 11th of 16 in 1993, with 10 wins and 10 losses, and 14th of 15 in 1994Cluster 9: Marginally Above-Average Team Drifts Towards Being Average
North Melbourne 2016: finished 8th of 18 in 2015, going out in a Preliminary Final, and 8th of 18 in 2016, going out in an Elimination FinalCluster 10: Average Team Remains Average
Geelong 1979: finished 5th of 12 with a 12 and 10 record in 1978, going out in an Elimination Final, and finished 6th of 12 with a 12 and 10 record in 1979Cluster 11: Above-Average Team Falls Away in Second Half
Geelong 2019: beaten Elimination Finalists in 2018 after back-to-back 100-point wins in their last two home and away games, and Minor Premiers in 2019, despite winning only 5 of their last 10 home and away games, eventually going out in a Preliminary Final after losing a Qualifying FinalCluster 12: Below-Average Team Gets Worse in First Half
Essendon 2016: finished 15th of 18 in 2015, and went 1 and 18 to the end of Round 20 in 2016 before winning 2 of their last 3 games.Cluster 13: Slightly Below-Average Team Becomes Average
Footscray 1984: finished 7th of 12 in 1983 with a 10 and 12 record, finished 7th of 12 in 1984 with an 11 and 11 recordCluster 14: Marginally Above-Average Team Becomes Very Good
Sydney 2012: finished 7th of 18 in 2011 with a 12, 1 and 9 record, going out in a Semi Final, and finished 3rd of 18 in 2012 with a 16 and 6 record, and going on to finish PremiersCluster 15: Well Below-Average Team Gets Worse in First Half
Gold Coast 2018: finished 17th of 18 in 2017, with a 6 and 16 record, and started 2018 with a 3 and 13 record before finishing 17th of 18 with a 4 and 18 record
Premiers and Runners Up come from only 6 of the 15 clusters:
Cluster 1: 10 Premiers and 8 Runners Up
Cluster 2: 19 Premiers and 20 Runners Up
Cluster 6: 3 Premiers and 4 Runners Up
Cluster 9: 1 Premier and 4 Runners Up
Cluster 11: 7 Premiers and 7 Runners Up
Cluster 14: 10 Premiers and 7 Runners Up
Finally, let’s have a team-by-team look (remember that you can click on any image to access a larger view).
Here we depict Premiership seasons with red lines, and Runner Up seasons with blue lines, with each row pertaining to a team and each column a cluster.
We see that:
North Melbourne are the only team to win a premiership with a Cluster 9 season. They did this in 1977 when they finished with a Combined Rating of just +13.7.
Adelaide (1997; +11.0), Sydney (2005; +14.0), and Richmond (2017; +20) are the only teams to win a premiership with a Cluster 6 season.
Only 11 teams have had a Cluster 2 season but missed the Grand Final. They are:
Hawthorn (1977; +18.7, 1982; +24.2, and 1992; +26.4)
Geelong (2010; +26.3, 2013; +24.2, and 2016; +23.7)
Carlton (2000; +27.6, and 2011; +26.2)
Sydney (1987; +12.8)
Adelaide (2006; +20.2)
Richmond (2018; +25.7)
