## Round 15 Predictions

I’ve been caught out by the Thursday game so I won’t have a chance to write a preview until tomorrow. Here are the predictions anyway!

UPDATE: So I’ve got around to taking a bit of a closer look at this rounds matches.

Finally we reach our last bye round – I know I’m looking forward to getting to a full round of matches. There are a few tight matches according to our ELO model, however the tight matches all appear pretty meaningless according to our Match Importance metric.

By far and away the most important match this round is the Syd v WB match – having big implications for both teams top 4 chances. I had a nice discussion this week with @Matterofstats on twitter about that measure – he suggested trying to weight the value of importance based on the likelihood of it occurring. I hope to implement that this week but I encourage anyone to jump on and have a read of his methodology.

## Round 14 Results

Continuing through the bye rounds, this week saw my ELO model managed to another 4 of the 6 games tipped correctly, with an MAE of 24.8. The two games we got wrong were the Friday night clash between Collingwood and Freo (we gave Freo a slight edge at 53%), and the enthralling upset win of St Kilda over Geelong.

We didn’t see a whole lot of movement in our ratings this week in the top 8 teams, apart from Adelaide extending their lead at the top of the ratings system. We are perhaps seeing a little bit of order appear in our top 8 after the logjam that existed a few rounds ago as well.

Last week our three biggest games on our match importance rating were Adelaide v North, St Kilda v Geelong and Hawthorn v GC. Given the relative strength of the teams in those last 2 games, I was a little disappointed that my new metric saw them as important, but after a big upset win by St Kilda, I can see why. With that loss, Geelong has dropped from 2nd to 5th on our simulated season ladder, with their top 4, top 2 and top 1 chances falling by between 12% and 18%. North also felt the impact of that loss, falling in all of our ladder position measures.

Our simulated season table below shows that the biggest impact on the weekend was probably felt by

We now see our top 2 rated teams also heading the top of our simulated season, with Hawthorn and Adelaide opening up pretty big gaps on the rest of the field in the top 4, top 2 and top 1 races.

Port Adelaides remain a roughly 1 in 5 chance to jump into the 8 after their bye, with St Kilda now making a big leap as a potential candidate to join them. They jump into finals roughly 17% of simulations, up from 9% last week.

Fremantle also becomes the 3rd team to not feature in finals in any of our 10000 simulated seasons after their big loss to a relatively poorly rated Collingwood.

## Round 14 Predictions

Our next bye round, subjectively, doesn’t look great on paper. Only the Collingwood v Freo match is considered tight by our ELO model, however it is fairly insignificant as measured by our Match Importance metric. The remaining matches, despite having some significance on the relevant team finishing positions, look fairly one sided.

Probably the most interesting match of the round, and certainly with the biggest Match Importance rating, we see the top rated Crows favoured pretty heavily over the 7th ranked Kangaroos, particularly with the Home Ground Advantage. A loss for either side affects their top 4 chances by roughly 30%. Adelaide by 23 points.

###### Collingwood v Fremantle, Fri 7:30pm MCG

As indicated, this match doesn’t really have an effect on our final finishing positions for either team, giving a combined change in probability of the top 8 of 2.8% for both teams. Despite that, the game looks like it could be tight, with our 9th ranked Freo team on the up after a tough start. Freo by 5 points.

## Round 13 Results

The first bye round of the year saw my ELO model managed to tiptoe through the horror round 12 correctly tip 4 of the 6 games with an MAE of 25.5. Like last week – the model struggled to separate the teams in both matches that it got wrong (Bulldogs v Geelong and Freo v Port Adelaide).

The main mover in our ELO ratings were Geelong, after their dominating win over the Bulldogs, lifting them by 28 rating points and jumping into a clear 3rd spot on the ELO ladder. Adelaide holds their spot on top during the bye, while interestingly Fremantle has almost got themselves back to the 1500 point mark that is our league average.

I introduced the match importance rating last week to try and quantify the relative effect a win or loss in a match could have on the end of season finishing position of a team. Last week, the Bulldogs v Geelong game was the most important and we can see the effect this has had on the Bulldogs in our simulated season – they have dropped to 7th on our simulated ladder, with their Top 4 and Top 2 probabilities dropping by 19% and 16%, respectively.

Geelong and Hawthorn are firming as our favourites for the top 2, both having almost a 50% of finishing top 2 and roughly a 25% chance of finishing as minor premiers. We are perhaps finally starting to see a big of order amongst our teams as we get further into the season.

Port Adelaides loss also saw their Top 8 probability drop from 36% to 22% clinging as the only team currently outside of the top 8 with a greater than 10% chance of jumping into it.

## Round 13 Predictions

The first bye round of the year presents us with two groups of games – coin flip’s in the North/Hawks, Freo/Port and Bulldogs/Geelong games, while the others the ELO model expects to be relative blowouts. You’ll also notice that I’ve added ‘match importance’ to the table below, which you can read about in my recent blog post.

I’ll summarise some of the important games below.

My model, as I keep harping on about, is still unimpressed by North Melbourne, rating them as the 8th best team and having them finish 7th on a (tightly packed) simulated season ladder. The Hawks on the other hand sit in 2nd in our ELO ratings. The system gives North a bit of a bump for being the home team. Our model is also naive to player selection so it doesn’t care about the Kangas injuries! Hawks by 5 points.

Freo, despite their awful start to the season, has been steadily improving their ELO rating and currently sit just behind Port Adelaide on the ELO ladder. They get a decent bump for their home ground but a relatively impressive Port team just gets the nod. Port by 7 points.

Possibly the game of the round – certainly the most important by our new importance rating! Two teams with almost identical ELO ratings – the Bulldogs get a bump with the notional Home ground. Has some big implications for each teams final ladder positions, roughly reducing the losers chance of a top 4 spot by 35%! Bulldogs by 7

## Beyond the 8 point game – estimating match importance in the AFL

As I was watching an enthralling match between Western Bulldogs and Port Adelaide over the weekend, there was a lot of discussion about how important the match was – the Bulldogs needed to win to cement their spot in the top 8, while Port needed to win to have any chance of jumping up. It’s also not uncommon for commentators to discuss the notion of an “8 point game“, typically when two teams close on the ladder play each other.

It got me to thinking about whether ‘match importance’ was an empirical measure that we could define? I stumbled upon an article entitled “The Importance of a match in a tournament“, which tried to answer this question in the context of the Premier League. While I confess I don’t understand a lot of the maths in the paper, a lot of which actually dealt with their predictions of matches, the general idea of measuring ‘importance’ was pretty simple. By looking at the change in the probability of a team winning the premier league based upon the result of the match, we gain a relative measure of importance for that particular team, in that particular match in the context of their season. In slightly more technical terms, we can use the difference in conditional probability of winning the league between the conditions of winning or losing a particular match to estimate the relative importance of that match.

$Importance(X_{team,match}) = Pr(X_{team}|Win_{team,match}) - Pr(X_{team}|Loss_{team,match})$

The above can read as the Importance of a particular match, for a particular team achieving outcome X (e.g. winning the league) is given by the probability of that team achieving X given (the “|” symbol) they win the particular match minus the probability that they achieve X given they loss the particular match. So if winning a particular match gives Team A a probability of winning the league of 40%, while losing the match gives them a probability of winning the league of 10%, the relative importance of that match is 0.3.

While estimating these specific probabilities over such a long time period is difficult (e.g. what is the probability that Hawthron will finish 1st as of round 12, 2016?) – our ELO simulations that we run each week give us pretty good starting point. In fact, that specific question yields an estimate of 23%, based upon running a Monte Carlo simulation of the 10 000 seasons and using the proportion of times that result was achieved as our probability estimate. We can use these simulations with the above formula by simply splitting them into two groups for each team – ones where they win the match in question and ones where they lose it, and then use the proportion of times our measure of interest (notionally X in our formula) occurred as our probabilities.

This measure is something I’ll report each week but to explore it in this post, I’ll use Round 12, 2016 as an example. This round was a round that saw some very close and difficult to predict matches, each of which felt quite important in the scheme of the league. In particular, close matches between top 8 teams occurred such as North Melbourne v Geelong, Adelaide v West Coast, GWS v Sydney and even Western Bulldogs v Port Adelaide (the latter of which is trying to beat the odds and jump into the 8 this year).

The table below shows the Importance measure for each team ahead of round 12 for the range of outcomes I typically report in my simulations.

Taking the Bulldogs v Port game as an example last week – from the Bulldogs perspective, in simulations where they won that match, they finished top four 64.8% of the time, while in simulations where they lost that match, they finished top four 37.4% of the time. Thus, the relative importance (for top four) for the Bulldogs is 27.5. For Port Adelaide, the match didn’t have a big bearing on their top 4 chances (influence score of 8.6), mostly because their chances of making the top 4 are so low anyway, but did have a big bearing on their top 8 chances (influence score of 29).

You’ll see from that table that In order to pick an ‘influence’ score for an actual game, I need to decide which X I choose in my calculation – essentially my output measure for the season. I could choose Top 1, as in the Scarf et al paper, but we don’t have as high of an importance on finishing on top of the ladder as the Premier League since we have a finals series. However, choosing just a Top 8 for instance may also not be that interesting, particularly if we expect the Top 8 to not change from here.

At least for me, I feel that subjectively, matches that can also have an influence of the make up of the 8 are quite important. As such, I’ve decided 1 to take the maximum importance measure for each team across the values in the table above (Top 8, Top 4, Top 2 and Top 1) to give an idea of the maximum effect this result of this match might have on that teams change in position. I’ll then add these two values together for the teams taking place to give an idea of the overall match importance.

Not surprisingly, the importance of matches is much higher for the better teams in the competition, so it will be interesting to explore how this measure changes over the course of the season. I may also revisit round 12 as, at least the general feeling I had while watching and reading about it, it seemed like quite an important one!

Notes:

1. at least for now, I may end up using ‘sum’ as we progress through the season

## Round 12 Results

My ELO model managed to tiptoe through the horror round 12 with a respectable 7 out of 9 tips and an MAE of 31.2. The two games it incorrectly tipped (Port v Bulldogs and WCE v Adelaide) were both within the margin of HGA in terms of tightness, so not unexpected results. The season total percentage is now creeping up with 80 correct tips (74.1%) and an MAE of 30.6.

There were some big movers in the ELO ratings this week. Adelaide extended its lead out in front, with an impressive win against West Coast in Perth while GWS cemented itself in the top 3 of our ratings with a big win over the previously 3rd best rated team in Sydney. Hawthorn received a big bump for its blowout win against the Bombers, although I suspect that Essendon this season might cause some headaches with our rating system as we go on. The other big movers were Fremantle and Melbourne, with bigger than expected wins, albeit against very low rated sides.

I may explore the horrible rating of the Lions and Bombers but the relative ratings difference between the Bombers and the 3rd worst team in the league (Gold Coast Suns) is 65 rating points (worth approximately 15 points in a game) is the same as the difference between Adelaide in 1st spot and North Melbourne in 8th spot! Pretty remarkable.

With Sydney losing big this week, they’ve dropped from 1st to 6th on our simulated ladder! North have also dropped from 2nd to 7th after a poor performance. Our model is beginning to become more certain on the top 8 as we go along, with Port Adelaide’s loss dropping their chance of a top 8 spot from 45% to 36%. The fact that the wins percentages are so close however suggests that, as per our ratings, there is little separating the top 9 teams this year.

## Round 12 Predictions

Last week was a great week for our model, tipping all 9 games correctly! Onto this weeks set of games and it looks like a tough round to tip! There are a bunch of games that are difficult to seperate by bookmakers and also by my ELO model.

This game is setup to be super close – Port Adelaide have been surging in the last couple of weeks, bridging the originally wide chasm between 8th and 9th place. After gaining a whopping 28 ELO points last week, they enter this game as slight favourites due to the Home Ground Advantage. Port by 3 points.

My model continues to be unimpressed by North Melbourne, despite having the Xth best start to a season in the history of the AFL, currently rating them as 8th best side. Geelong’s recent mishaps against Collingwood and Carlton however were treated rather harshly by the ELO model, seeing them slip from the 2nd best to 7th best team. The two teams are almost identical in ratings, with Geelong getting a slight bump due to HGA. Geelong by 9 points.

Adelaide this week became the first team to topple Hawthorn as the top ranked team since 2014, however go into this one as slight underdogs due to playing away against the 4th best team by our ELO ratings. Could be the game of the round. West Coast by 5 points.

###### GWS v Sydney, Sun 4:40pm Spotless Stadium

Even taking away the subplot of big brother versus little brother, the ‘battle of the bridge‘ promises to be a tight matchup between the 5th (GWS) and 3rd (Sydney) best teams in our ratings. HGA gives GWS the slight edge according to our model. GWS by 4 points.

Even games like Melbourne v Collingwood and St Kilda v Carlton promise to be close – my model is perhaps operating the home team chances here due to applying a flat home ground advantage, despite one probably not existing in these Melbourne based games 1. The only real confidence that my model has appears to be losses for the 3 worst sides in my ratings system (Essendon 20%, Brisbane 37% and Gold Coast 31%).

Notes:

1. something I’d like to fix

## Round 11 Results

After a few weeks struggling to juggle and balance various commitments, I think I’ve finally got into enough of a routine and automated enough of my R scripts that I can release my ELO ratings update early in the week and put my predictions out later in the week! I’ll also hope to write a bit more interesting post mid-week such as my one on the Tackle Machine, or the Round 7 rule. Here’s hoping anyway!

Anyway, after 4 weeks in a row tipping 6 out of 9 winners, my ELO model has had its first official perfect round, correctly tipping 9 winners over the weekend! It helped that the favourite got up in 8 of the 9 rounds, making it a relatively easy week to tip, but like a proud parent, that doesn’t bother me in the slightest! The MAE for this week was 29 points, bring the season total to 73 correct tips (73.7%) with an MAE of 30.53.

The updated ELO ratings are below and, for the first time since Hawthorn lost the 2014 Grand Final, they are no longer the top team in our ratings! With Adelaide’s trashing of the Saints, and Hawthorn not putting Melbourne away by as much as expected, the Crows leapfrogged the Hawks into top spot! Port Adaide and North Melbourne also impressed our model, bridging the gap to the other top teams, with a relatively close group of 9 teams now clear of the rest of the pack.

Our simulated ladder still favours Sydney, obviously taking into account their rating of 3rd, their current record and remaining draw, with North Melbourne also jumping up into 2nd spot by average. The model still has trouble splitting the make up of the top 8, but is getting pretty certain that no-one apart from Port Adelaide can jump up into the 8. For Port, their win on the weekend gave them a nice bump in ratings and lifted their percentage of top 8 finishes to 45% in our rating system.

We are also starting to see some different shapes emerge in our distribution plots for seasons, essentially signalling that the model is more certain about where some teams will finish than others. Hawthorn and North, for example, despite winning on the same amount of wins, show different distribution shapes. It will be interesting to see how this changes during the season.

## Round 11 Predictions

I’ve separated out the simulations and ratings update into another post and I’ll just focus on this weeks games here.

Last week gave my ELO model a 4-peat of 6 out 9 tipping weeks. There were quite a few close tips this week, with an MAE of 23, bringing our season total to 64 (71%) with an MAE of 30.6.

To this weeks games, we see some tight matchups. Concretely, Geelong v GWS gives Geelong a slight edge with Home Ground Advantage. Similarly, the Bulldogs are tipped to slightly get up over WCE despite being rating below them due to being the home team. In our closest matchup, the away Port are expected to just get over Collingwood in a coin flip.