Upsets and Mismatches in VFL/AFL History: A MoSSBODS 2.0 Perspective

A few years ago, I used the MARS Rating System to identify those games from VFL/AFL history that might be considered as the greatest upsets in that the loser's MARS Rating was very much higher than the winner's. There is, of course, no objective way of identifying such games (and the MatterOfStats site is nothing if not a source of multiple answers to the same question) so today I'll be revisiting this topic using the MoSSBODS 2.0 Rating System.

In identifying upsets I'll use a mismatch metric similar to the one I used in the earlier blog, excepting that Venue Adjustments will be included in the estimate of a game's likely outcome, not just the raw Team Ratings. So, games played on interstate grounds where the home team is assessed as enjoying a substantial home ground advantage, and where the home team's underlying ability is assessed as being much greater than the away team's, will be treated as larger mismatches than would be the case were only raw Team Ratings used.

Specifically, the mismatch metric will be the Expected Margin for the game, measured in points, converting MoSSBODS' underlying Scoring Shot based assessments into points by using the average Conversion Rate from the immediately preceding season.

The table below records the 30 greatest mismatches in VFL/AFL history using this approach, as well as the 10 greatest upsets. (Note that, in another slight deviation from the previous blog, I've here allowed draws to be considered as upsets. Also note that the image is clickable.) 

According to MoSSBODS then, and based on the methodology I've described, the greatest mismatch in VFL/AFL history was the 1991 Round 9 clash between the Eagles and the Lions. This is, in fact, the only game for which MoSSBODS has ever produced an Estimated Margin of more than 100 points. As it turned out, MoSSBODS' prediction of a 105-point victory by the Eagles was exceptionally good, the final margin falling within 1 goal of that assessment.

Fourteen of the next 29 mismatches come from seasons 2011 through 2013, GWS featuring in seven of them, Gold Coast in two, and Melbourne in three more. The 2011 Brisbane and Port Adelaide sides also appear once each. Whilst GWS and Gold Coast are not the only teams to have joined the competition during its history, the fact that they've done so when seasons have been, relatively speaking, longer than average, has contributed to their prominence in this list. Compared to new teams from earlier eras they've had to play a larger number of games in their inaugural seasons before seeing their Rating reset nearer to zero, which is what MoSSBODS does at the end of each season.

Amongst the 30 games assessed as being the greatest mismatches, 20 have resulted in victory margins for the favoured team of 50 points or more, and six have resulted in margins of between 30 and 50 points. Only one of them produced the "wrong" result, and it is this game that is, therefore, deemed to be MoSSBODS' greatest upset.

That game is the 1994 Round 5 clash at Subiaco between the Eagles and the Hawks, which MoSSBODS expected to as about a 12-goal win for the home team. Instead, it finished as a 12-goal win for the visitors. MoSSBODS' enthusiasm for the Eagles' chances was spawned partly from the significant home ground advantage they were assessed as having (+3.5 Scoring Shot Nett Venue Performance and +3 Nett Travel Penalty), but also from the recent form of the two teams. West Coast entered the game on the back of a 19-point victory over the Crows, followed by 96- and 76-point wins over Richmond and Fitzroy. Hawthorn, in contrast, had lost to Melbourne by 54 points, North by 127 points, and then Carlton by 87 points.

The second greatest upset is the drawn Round 22 clash from 1980 between the Roos and the Saints. MoSSBODS had the Roos as about 11-goal favourites, they on the back end of a season in which they'd gone 14 and 7 including victories in five of their last seven games, while the Saints had gone 5 and 16 having lost eight of their last nine games, including a 222-70 loss to Richmond in Round 16, followed by a 211-107 loss to Collingwood in Round 17. It takes a while for Ratings to recover after losses of those magnitudes. The Saints lone win in that period was though against the 2nd-placed Cats in Round 19, so they did have some form for generating upsets.

Only two of the 10 greatest upsets are from this century. In eighth is the Round 7 Crows v Dees 2014 matchup at Adelaide Oval, which MoSSBODS had the Crows entering as about 9-goal favourites. Neither team went into the game in scintillating form, Adelaide with a 2 and 3 record though also with big wins over GWS and St Kilda in the two previous games, and Melbourne with a 1 and 4 record, and a loss to the Gold Coast in the previous week. Melbourne's Rating was also still bearing the scars from the team's 2 and 20 record in 2013, which saw them start the season with a Combined -8.5 Scoring Shot Rating. By comparison, the Crows' 10 and 12 record for the 2013 season, including wins in four of its last five games, had allowed it to enter the season with a Combined +2.2 Scoring Shot Rating. 

The TAB also didn't give the Dees much hope, pricing them at $9 and offering Melbourne 44.5 points start, but the Dees prevailed by 3 points regardless, despite the Crows registering seven more Scoring Shots.

The 2012 Round 23 Richmond v Port Adelaide game is the other upset from the current century, it finishing in a draw with the Tigers assessed as about 9-goal favourites by MoSSBODS. Richmond came into the game with a 10 and 11 record for the season but also having won three of their previous four games, and enjoying a 3 Scoring Shot Nett Travel Penalty benefit and about a 1.4 Scoring Shot Nett Venue Performance benefit. The TAB didn't completely dismiss Port's chances, but did price them at $6 and have them receiving 41.5 points start.


Let's broaden our view a little now and consider entire seasons rather than just individual games from within them, our goal being to understand the mix of games and results within each. Specifically, let's consider the profile of Expected Margins in the games from each season and the prevalence of victories by the underdog teams as assessed by MoSSBODS. 

For this analysis I'll be reverting to MoSSBODS' basic unit, the Scoring Shot, as this, arguably, provides a better method for comparison across seasons separated in time where both the number of Scoring Shots and the the rate at which they were most likely to be converted could differ significantly.

In this first chart we see time series for the proportion of games in which MoSSBODS assessed Expected Margins in one of five bins. To remove some of the variability in the data, I've calculated 9-season, centered moving averages of the raw percentages for each season.

The green line, for example, tracks the proportion of games in which MoSSBODS assessed the favoured team as enjoying an advantage of less than 2 Scoring Shots (which is about 7.3 points using the most recent Conversion rates). This is then the proportion of games that MoSSBODS expected to be "close". It has fluctuated somewhat over the history of the competition, though the moving average has stayed within the 20 to 30% range. In recent seasons, it's been, as stock-market chartists might say, testing the lower bound.

We've also seen a decline in games assessed by MoSSBODS as having 2 to 4 Scoring Shot favourites (the dotted blue line), this decline being mirrored by the rise in games with favourites expected to win by 10 Scoring Shots or more. It's important to recognise, however, that the raw data for games of this latter type shows a decline for each of the last two seasons (19% in 2014 and 14% in 2015) relative to the peaks of 2012 and 2013 within which MoSSBODS assessed 26% of games to be of this type. The proportion of games expected to be won by less than 2 Scoring Shots has also increased in the past two seasons, from the lows of 16% in 2012 and 21% in 2013, to figures of 23% in 2014 and 26% in 2015.

In summary then, if MoSSBODS is an acceptably accurate assessor of likely game margins, there has been a tendency to a larger proportion of games representing significant mismatches in recent times, though this has been heavily influenced by the introduction of GWS and Gold Coast and is showing signs of abating in the last two seasons.

Mismatches might be more tolerable, however, if underdogs prevail often enough. In this final chart we'll track the success rate of the the underdogs in each of the game categories used for the previous chart. Again we'll use the 9-season, centered moving averages, and we'll maintain the same colour scheme.

The first thing to note from the chart is that, for most of the time, the ordering of the lines is as we'd expect, signifying the fact that underdogs assessed as having more forlorn hopes have tended to win at lower rates than underdogs assessed as having better chances. In modern times, for example, teams starting as less than 2 Scoring Shot underdogs have won a little less than 50% of the time, while teams starting as 10 Scoring Shot or more underdogs have won less than 10% of the time. 

It's interesting to note that the trajectory of the winning rate for the longest of underdogs, which tracks upwards from zero in the mid 1930s up to about 20% by the mid 1990s before collapsing back below 10% in recent times. (Note that the very high rates in the 1920s and 1930s are subject to high variances because of the relative rarity of games with 10 Scoring Shot or more favourites in that period). 

This trajectory might merely be an artifact of the MoSSBODS methodology, which sees it systematically overstating the differences between teams, especially at the extremes, or it might instead reflect a real phenomenon fuelled by an increase in the inherent variability of outcomes in the 1990s relative to the 1930s, which has recently plateaued.

Although it is a somewhat circular argument, I realise, the pattern of (mean absolute) errors in MoSSBODS' predictions generally, supports the contention that game outcomes varied more about their means in the mid 1990s than in the mid 1930s, and that there has been some slight reduction in variability since then.

The notion that greater variability is beneficial for underdogs is one we've explored here before on Matter Of Stats. It would be pleasing  if the history of the VFL/AFL competition turns out to have provided an 80-year case study of the theory. 

If you've evidence or opinion, supportive or otherwise, please share it in the comments section of this post, or by e-mail. In the meantime, let's hope that 2016 produces fewer mismatches and more upsets than we've seen in recent seasons, and that Investors are able to benefit from those upsets that we find mispriced.