Round 19 Predictions – Last chance Saints

We are starting to end the pointy end of the season and we see probably the most important match for a contender for the top 8 in terms of keeping their season alive. A win or a loss for the Saints this week see’s a 40 percentage point swing in their top 8 chances! We also see a fairly wide gap in classes this week (both in terms of ELO ratings and ladder ratings). There is only one top 8 clash and our closest matchup in terms of ELO ratings is Collingwood (10th in our ELO ratings) v West Coast (6th). In that match we still give the away team in the Eagles a 59% chance of getting up, despite the home ground advantage against them.


While likely not a great game, from an outlier perspective an interesting matchup is the Essendon v Adelaide game on Sunday. Here we see our worst rated team in Essendon away to our (newly downgraded) 2nd rated team in Adelaide.

In terms of our match importance rating, I’ve summarised the two big matches below.

Important Games

plot of chunk unnamed-chunk-4

Geelong v Western Bulldogs, Fri 7:45pm, Skilled Stadium.

While we would see this as a big match just based on the respective ELO ratings and effect on the final ladder, the big milestones for Geelong and the way in which the Bulldogs respond to their injuries this week gives this a fair bit of spice. For both teams, this has a big change in the respective chances on the top 4, with a Bulldogs win giving them a 58% chance while a loss sees that drop to 14%. Similarly for the Cats, they can almost sow up the top 4 with a win increasing their chances to 88%, a loss dropping them to 47%. Our model thinks Geelong should be too strong, particular with HGA. Geelong by 21 points. 

North Melbourne v St Kilda, Sat 7:25pm, Etihad Stadium.

While the big story from this game is easily Brent Harvey breaking the all time games record, this game has the biggest implication on the top 8 out of all games we’ve analysed this year. Put simply, the Saints are the most likely team to make it into the 8, while North are the most likely to drop out. A win to either side tips the odds in their favour dramatically. St Kilda wins, their chances jump to 47%, while a loss sees them drop to 6%. A win for North gives them a 91% chance while a loss drops that to 45%. Unfortunately for St Kilda, our model doesn’t take any of that into account and doesn’t rate them very highly, giving the Kangaroos a fairly comfortable edge. North Melbourne by 30 points. 

Match Importance Table

We can see below that the two big standouts are the change in top 8 chances for North/St Kilda and the change in top 4 chances for Geelong/Bulldogs. A loss this week also sees a big change for Hawthorn in their question for the minor premiership, although this probably points out that I should be weighting these percentages based on the likelihood of them occurring given we only give Carlton a 19% chance of upsetting the Hawks.


The Deledio Effect – can Richmond win without him?

After the weekend, in which Richmond lost handsomely to (a very good) Hawthorn side, the narrative trotted out by the football media again seemed to surround the idea that Brett Deledio didn’t play. Every time he doesn’t play and Richmond lose, we generally get given some cherry picked data about how many games Richmond has lost while he hasn’t played. While it may be the case that he is super important to Richmond, simply stating their win/loss record without him doesn’t show this. I thought I’d use this opportunity to explore the issue slightly more in depth!

My general belief is that while the make up of a team is important (as we can clearly see from Essendon this year, after losing half of their list), the marginal gains that one player may provide are quite small compared to our overall beleif about a teams gross ability. My ELO model, which is ignorant to team make up, can do a pretty good job of predicting match outcomes. Nonetheless, data from overseas has begun to explore the idea of using Network Analysis to understand team makup. In fact, in AFL, Sargent and Bedford published research using Interactive Network Simulation Geelong’s 2011 season to understand the impact an individual player inclusing or exclusing from a lineup had on a team.

Interactive Network Simulation of the effect of Bartel on the Margin. Figure taken from Sargent and Bedford. J Sports Sci Med. 2013; 12(1): 116-121

While I don’t have the time or expertise to implement their Network analyses (yet), I can use my ELO model to gain a better understanding of the specific effect of having Deledio has on our expected outcome, rather than simply expecting them to win every game he plays in. Since I’ve got a model that does a pretty good job over a season of applying a probability to a particular matchup based on the relative rating assigned to a team (independant of the players involved), Ive got a nice tool to compare how Richmonds winning rate compares to what we would expect from a team of Richmonds level compared to how they have actually performed without Deledio.

The first thing we need to know is which games have Richmond played in during Deledios career where he did and didn’t play. I’ve used data afltables, to come up with all Deledio-less games. Knowing this, we can now take a look at Richmonds record in the time since Deledio’s debut in Round 1, 2005. In the 262 games that Richmond has played in since his debut, he has missed 19, with Richmond only winning 3 of them, a rate of 15.79%. This is on contrast to his career winning percentage of 45.27% over 243 games. It is this descrepancy that is often terrmed the Deledio Effect.

plot of chunk winsPlot

The first thing we note here is how small our sample size is for games in which Deledio missed, compared to those in which he won. We could probably just stop our analyses here and make a statement about how we need more information to better understand the true impact of having Deledio has on the Tigers. But as promised, I’ve got my ELO model that can at least give us a slightly deeper understanding about how anomalous this result is.

Below, I’ve taken the expected margin computed from the difference in ELO rating (plus a home ground advantage boost) for those games in which Deledio missed and plotted those expected Margins against the actual margins.

plot of chunk marginPlot

This plot again shows that there is really a very small dataset to work with for the missing data. It also shows that when he does play, our model does a pretty good job at predicting the margin on average, with an MAE of 31.3 for when he does play. Tentatively, this value is slightly higher when he doesn’t play (36.2), although, n = small.

In those games where he does miss, we can see that there is a relatively big proportion of games in the quadrant where we predicted a win but saw a loss (highlighted as blue), suggesting that these are the games where the Tigers underperformed against our models expectations.

To attempt to further explore this data, I’ve performed some simulations (you can read about the methodology here) on the Tigers games during Deledio’s career. Breifly, I use the difference in ELO ratings between the two teams to estimate a predicted margin. I then draw an actual simulated Margin from a normal distribution with a mean of that expected margin and some noise. I’ve then repeated this 10000 times for each game during the Deledio era. You can see the distribution of wins for each category (Missed v Played) below.

plot of chunk distributionPlot

We can see that our distribution of wins for the Missed games centers around 10 (a percentage of 52.63%), well above our actual wins of 3. In fact, we only saw 3 wins or less in 0.05% of our simulations! In games where he played, this rises to a 76.87% in games where he does play, suggesting that Richmond actually outperform expectations in those games.


So what can we conclude? I started off convinced I could debunk the Deledio myth but our data hasn’t helped. There is a lack of data points where Deledio doesn’t play, so making definitive inferences is probably not valid. We can conclude however that in games where Deledio doesn’t play, Richmond perform well below where we’d expect them. In games where he does play, Richmond perform well above where we’d expect them. My speculation for possible explanations are

  • Unexplained variance

This whole analyses is based upon my ELO model, which does a pretty good job of predicting outcomes (~70% of games tipped correctly, better than chance) and margins (a long term Mean Absolute Error of 30 points). However, at least in the Margins, it still only explains about 27.48 of the variance in the final margin. That leaves more than 2/3 of the variability unexplained (slightly worse than what Matter of Stats has shown, but in the same ballpark). Some of that variability may come from Deledio missing, but more likely (especially with our small dataset) is that a combinatino of lots of factors is bringing that variability

  • Randomness

Related somewhat to the above point, but sometimes events that we attach small probabilities to actually occur. In fact, given enough time, we fully expect that they will occur, we just don’t know when. So even though we’d only expect the Tigers to win 3 games or less based on our ELO model <1% of the time, it still did happen. Maybe we’ve just seen an extreme run of abnormal results.

  • Deledio is super important!

The final take away may be that Deledio is actually important. In the Sargent and Bedford paper I referenced earlier, they used an example of taking Jimmy Bartel out of the Geelong team and replacing him with Shannon Byrnes. This showed a net effect on the mean margin of a game to be 15 points to the score, highlighting how important Bartel was to that Geelong side. Maybe Deledio is actually that important compared to the player who does come into him.

Hopefully with some more knowledge, and data, I can further explore this idea using a similar methodolgy. For now, lets just say that Richmond underperform from what we expect when Deledio doesn’t play and leave it there.

Round 18 Results

Last weekend saw our  ELO model have its worst week for the season, tipping at less than 50% to record 4 correct tips out on 9. It was also the 2nd worst round of the year for the Margin component of the predictions, with a mean absolute error (MAE) of 36.9 points.

In 4 matches last week, our had difficulty separating teams after the Home Ground Advantage was taken into account, giving the tipped winners less than a 55% chance of getting over the line. Unfortunately for our model, all of teams ended up losing (North Melbourne at 50.1%, Freo at 54%, Adelaide at 52% and Port Adelaide at 53%), with two other upsets also hurting (Essendon at 58% losing to the Lions and the Bulldogs at 70% losing to the Saints). Luckily the Swans scraped over the line against the Blues while West Coast managed to hold on against the Demons despite losing the Inside 50’s handsomely.

Onto our rating system and, with a poor round of tipping, that gives my model more ammo for changing team ratings to try and make sense of it all. Adelaide loses its 8 week long rein as our top team, handing the title back to Hawthorn after their disappointing loss to Geelong, who also benefit, gaining 13 points and jumping back into our top 4 in ratings. The biggest mover in our ratings was North Melbourne, who jumped 17 points and 2 positions to be tied for the 7th best team in the league.

Our simulated season sees Hawthorn as the first team to clinch a finals birth for the year. They also strengthened their grip on securing home field advantage for the whole finals series and avoiding any travel, which could be important given some of the teams just below them. In fact, Hawthorn now finish on top in a neat 2/3 of our simulations.



The biggest loser from the weekends matches in terms of finals positioning appears to be Adelaide, losing 12%, 21% and 18% in their top 4, 2, and minor premiership chances, respectively. The main beneficiary of those chances has been Hawthorn, with only Geelong’s top 4 chances (up 15%) changing significantly outside of Hawthorn.

North bumped out their chances for holding onto top 8 by winning and Port losing, although the Saints have jumped up to be the most likely team to sneak into the 8 if North falter (currently North are at 73%, with Saints at 22%). Port has dropped to an 11% shot to make the finals, with no other team higher than 1%.

Round 18 Predictions

We’ve talked for a while about how tight the top 8 is, meaning every game the top 8 play in is important. Interestingly this round, we actually only see two top 8 teams play each other – our likely match of the round in Geelong v Adelaide. Given that Adelaide are our top rated team according to our ELO model and that match has some big implications as measured by our match importance rating, it is setup to be a cracker.


Two of our other higher rating implication games relate to likely the only two teams who may swap places between inside and outside the top 8 in Port Adelaide and North Melbourne. Port need to keep winning to stay in touch, while North can’t afford to keep dropping games.

I’ve outlined some of the bigger games below.

Collingwood v North Melbourne, Fri 7:50pm Etihad Stadium

As mentioned, North need to start winning again, going 1-5 after their very good 9 match winning streak to start the season. They still have a buffer of 2 games over 9th placed Port Adelaide but have dropped below them in our ELO ratings and the difference between a win and a loss is 34 percentage points for their top 8 chances. Given Collingwood has steadily improved to be rated just below North, our ELO model is predicting basically a coin flip. North Melbourne by 1 point. 

Western Bulldogs v St Kilda, Sat 7:20pm Etihad Stadium

The Bulldogs probably haven’t impressed our ELO model as much as their ladder position suggests, sitting in 8th spot. They do however sit at 3rd on our simulated season table and need to win these games to maintain their top 4 chances. St Kilda improved their top 8 chances to a non-trivial 14% last week but can’t afford to drop too many more games given they are 2 games and percentage behind North Melbourne. Our model favours the Bulldogs as 70% chances but it has a high importance rating nonetheless. Bulldogs by 34 points. 

Geelong v Adelaide, Sat 7:25pm Skilled Stadium

Adelaide has been sitting at the top of our ratings for a few weeks now and sit fairly comfortably in 2nd spot of our simulated season. Geelong on the other hand have been disappointing in recent works to our model and have seen their top 4 chances slipping as the season goes by. This week sees them face a 41 percentage point swing in their top 4 chances based on a win or loss. Our model slightly favours Adelaide with Geeling getting over Geelongs HGA . Crows by 3 points. 



Round 17 Results

As the season starts to heat up, we are getting some interesting games in the context of the season, as noted last week by the super important Sydney v Hawthorn game. In combination with a really tight top 8, whenever two top 8 teams are matched up, my ELO model has difficulty separating them. Nonetheless, we continued on with an OK 6 out of 9 games tipped correctly, with a respectable MAE of 22. Our 3 misses were all ones that our model predicted to be close, with Sydney (54% chance), North Melbourne (56% chance) and Melbourne (54% chance) all losing. This gives us a season total of 105 tips from 144 games (73%) and an MAE 28.6.


One note of interest is I’ve had to adjust my scale for this plot because of Brisbane doing so badly. I might write a bit of a historical piece comparing them to other badly performing teams.

Given the tightness of the MAE this week, we didn’t see many big jumps in our rating system. The two most notable movers are Port Adelaide and North, who swapped places this week (7th and 9th, respectively). North has now almost dropped to the level of an average team (1500) and face a big matchup against a similarly rated Collingwood this weekend. I haven’t released my full match importance ratings but the Friday night clash between North and Collingwood is big for North’s chances of holding onto their top 8 position.

I briefly noted last week that my match importance rating was very high on the Sydney v Hawthorn game last week, which can be seen the Swans’ chances of top 1 and 2 taking about a 15 point hit. Their top 4 chances interestingly didn’t change too much (dropping from 65% to 58%), probably as North and WCE chances of going on a run and getting into the top 4 are getting smaller.


The biggest drop from the weekend was easily North Melbourne, whose chances for top 8 dropped from 76% to 62%. The main beneficiaries for that were Port Adelaide – jumping from 20% to 31% and St Kilda, who are now into double figures. We’ve also seen Freo now mathematically 1at least in my 10000 simulations they don’t make it once[\ref] falling out the race for the top 8.


    Round 17 Predictions

    Stupid Thursday game caught me out again. Will update when I get a chance.


    UPDATE I’ve now added in the match importance ratings for each game. By far the most important game was the Thursday night epic between Hawthorn and Sydney. I’ll write a bit more about that in the review but that match had approximately a 40 percentage point change in top 4 chances for each team.

    The next most important involves Port Adelaide and North Melbourne, the two sides sitting in 8th and 9th. While North are much better chances to stay in the 8, a loss to Port this week sees those chances shift further into Port Adelaide favour, with a ~25 percentage point swing either way based on the match.



    Round 16 Results

    For our first week out of the bye rounds and back to a full 9 games, my ELO model had a pretty good week, tipping 7 out 9 with an MAE of 24. The incorrect tips were GWS losing to Collingwood (a 29% chance)e and Geelong losing to Sydney (a 23% chance). The seas total is up to 99 tips from 135 games (73%) and an MAE 29.0.

    Those two upset losses saw the biggest change in ELO ratings for the round with Geelong losing 21 ratings points (giving them to Sydney) and GWS losing 26 rating points (giving them to Collingwood). That saw both Sydney and West Coast leap above those two teams into 3rd and 4th, respectively. Collingwood has also jumped up to be an almost ‘average’ team.


    I mentioned last week that the Geelong v Sydney game had our biggest match importance rating since I’ve started reporting it. That can be seen by the new simulated ladder, where Geelong has dropped down in top 4, top 2 and top 1 contention by between 8 and 16 percentage points. The model now only gives them a 5% chance of finishing on top. In contrast, Sydney has jumped up by 20 percentage points in top 4 calculations, with some smaller increases in top 2 and top 1 chance.

    was certainly the Sydney v Bulldogs game, is would be expected by a matchup between two top 4 sides. The last minute goal to Jason Johannisen saw the Bulldogs leapfrog the Swans on our simulated season ladder. The Swans top 4 aspirations took a hit, dropping from 54% down to 44%, while the Bulldogs top 4 chances jumped from 34% to 46%. Given the tightness on the actual AFL ladder this season, there should be a few more of these big impact games coming up!


    Hawthorn has firmed considerably in Minor Premier race, finishing on top in just over 1/3 of simulations, with Adelaide remaining the next best chance. Both of those teams also firmed slightly for top 4 and top 2 chances after the weekend.

    Port’s loss saw its chances of top 8 fall again – they are now 3 wins outside of the 8 – now down to 20%. By far and way the most likely to drop down is now North Melbourne, missing the top 8 in 24% of simulations.

    Round 16 Predictions

    As our ELO model suggests, the top 8 is super tight. In fact, our simulations have a really tough time separating out clear favourites in our top 8 race. Data from The Arc, suggests that this is the tightest race we’ve ever had at this stage of the year. Considering that, anytime two top 8 teams play each other, its going to be important.


    We’ve got 2 top 8 clashes this week in Sydney v Geelong and WCE v North, which have big implications for the respective finishing positions of those teams. In fact, in the very short life of our Match Importance rating, the Sydney v Geelong clash is the most important game of the season.

    importance_R15The three most important matches are below:

    Port Adelaide v Hawthorn, Thurs 7:20pm Adelaide Oval

    The other super important match is this clash between 1st and 9th. Hawthorn is sitting pretty at the top of a very congested ladder while Port remains our best chance to jump into the top 8 according to our match simulations. The difference in a win or a loss for their top 8 chances is ~28%, so a win goes a long way to them making the leap. Hawks superior rating gives them the upper hand but HGA brings Port into it. Hawks by 5 points. 

    Geelong v Sydney, Fri 7:50pm Skilled Stadium

    The clash between 2 of the premiership favourites according the bookies, our model has this as the most important match of the year so far (at least since I’ve been doing importance rankings). The percentage change in top 4 chances from a win or a loss is 36.5 for Sydney and 38.4 for Geelong! Our model gives Geelong the slight edge in ranking points, as the 3rd best team, while Sydney dropped slightly back to 5th after last weeks loss. The HGA gets Geelong a bit more of a bump.  Geelong by 12 points. 

    WCE v NM, Sun 1:20pm Domain Stadium

    Given how tight the top 8 is, this has big implications on the makeup of the bottom part of the 8. Both these teams are at most risk of dropping out of the 8. A loss here puts them behind the pack for top 4 and at risk of a rampaging finish by Port Adelaide. WCE are rated about 30 rating points better, with the home ground advantage. Eagles by 15 points. 

    Round 15 Results

    My ELO model has completed a clean sweep of 4 out 6 tips for the bye rounds on the weekend, incorrectly tipping against Gold Coast and the Western Bulldogs. It was far and away the best round of the year for margin tipping, with an MAE of 16.5! That gives us a season total of 92 tips from 120 games (73%) and an MAE that has tipped under 30, sitting at 29.4.

    With such a tight MAE, there isn’t a lot of movement in our ratings this week since each team performed relatively close to where our model expected. The obvious big movers are Gold Coast and St Kilda, after the big win by the relatively weaker rated Suns. Adelaide remains relatively clear on top of the ratings after the 2nd, 3rd and 4th teams all had byes.


    The biggest impact on our finals simulations last week, as measured by the match importance rating was certainly the Sydney v Bulldogs game, is would be expected by a matchup between two top 4 sides. The last minute goal to Jason Johannisen saw the Bulldogs leapfrog the Swans on our simulated season ladder. The Swans top 4 aspirations took a hit, dropping from 54% down to 44%, while the Bulldogs top 4 chances jumped from 34% to 46%. Given the tightness on the actual AFL ladder this season, there should be a few more of these big impact games coming up!


    The race for the top 2 didn’t change too much after the weekends matches, again, likely due the bye for 3 of our top 4 sides. The Crows and Hawks remain almost 50% chances get finish inside the top 2.

    Port has improved its chances of dislodging one of the top 8 after its win on the weekend, jumping up to a 28% chance. The most likely of our top 8 to drop out is North Melbourne, who miss the 8 in 18% of simulations, despite their fantastic start to the season.