Squiggles are visualizations of AFL team performances, charting attack and defence over the course of a season.

Teams ⇡ high on the chart kick big scores. Teams on the ⇢ right keep their opposition to low scores. Teams in the ↗ upper-right do both. Teams in the ↙ lower-left do neither.

Squiggles are handy because:

• They visualize different game styles. Ross Lyon teams, for example, lurk on the right side of the chart about halfway up, since they keep their opposition to low scores without scoring highly themselves.

• They provide interesting season replays, showing how teams slumped or soared at different times of the year.

• They aren't affected by fixture bias, since they account for the difficulty of the opponent and venue.

• They predict how the season will play out based on current likelihoods.

Squiggle 2.0

• It is more sensitive in the early rounds, so surprising results can cause a lot of chart movement. This better captures the sometimes substantial form change that can occur over an off-season.

• It considers goalkicking accuracy, by discounting scores that resulted from unusually high accuracy (i.e. kicking many more goals than behinds), and padding scores that resulted from unusually low accuracy, since these tend to be non-reproducible.

• It generates home ground advantage from the number of times each team has previously played at the same venue and in the same state, instead of relying on a simple 12-point rule of interstate home advantage.

Squiggle is roughly as accurate as tipping the favourite every game. (Which is hard to beat!) Over a season, an informed, observant human should be able to beat it, but not by much. It will beat an average human tipper. You can review its accuracy by visiting the TIPS section of any year.

* Squiggle 2.0 numbers before 2018 are "retro-dictions"—they were made after the result. This shows how well the model fits historical data, rather than how reliable its predictions have been in practice.

Squiggle v1

Follow this link for Squiggles generated under the v1 algorithm (ISTATE-91:12).

To watch the evolution of a past season, use the top controls:

Rewind to the start of the season

Previous Round

Next Round

Reload

You can also:

• Click a team name in the legend at the top to hide/show it.

• Click a team flag to remove every other team.

For example, you might like to rewind, click Hawthorn's flag to remove all other teams, click Geelong's name in the legend to add it back in, then repeatedly step through the season to watch their dance of death.

Things work faster with fewer teams. If the chart is sluggish, try removing some!

To zoom in on an area, drag a box around it.

Weekly Tips

Click TIPS at the top of the page. These update at noon on Tuesdays. The predicted scorelines are what Squiggle expects. The more the actual scoreline differs from this, the more teams will move.

Season Predictor

This is how the ladder will look if Squiggle has correctly rated every team and nobody gets better or worse.

For the home & away season, it uses a probabilistic ladder, not a simple tally of tips. Both teams are awarded a win probability from each game, so that if Squiggle thinks Hawthorn is 68% likely to beat Collingwood, it will award the Hawks 0.68 wins and the Pies 0.32 wins, increasing both team's tally of "probable wins" by less than 1.

This is because if a team plays 10 games with 60% likelihood of winning each game, we should expect them to win about 6/10—not, as we would get if we tipped each game and tallied up the tips, 10/10. We know that upsets will happen; we just don't know when. A probabilitistic ladder accounts for the likelihood that teams will sometimes unexpectedly win or lose, even though we doesn't know when.

This can look like a bug in the predictor, if you see a team tipped to win a match that doesn't seem to be credited. For example, a team might be on "15 (14.7)" wins, which means 14.7 "probable wins" rounded off to 15. (Rounding occurs so that teams can be secondarily ranked by their percentage.) And then that team is tipped to win the following week, but it remains on 15 wins, now "15 (15.3)". What has happened is the number of probable wins hasn't risen by enough to be rounded to a higher number. It has earned 0.6 more probable wins, but this still rounds off to 15. The predictor is saying it's still most likely this team will be on 15 wins, after accounting for the likelihood that some of its tips will be wrong.

Finals matches are predicted using simple tips. However, this isn't a very reliable way of doing it, and not Squiggle's official Premier tip. For this, please see Flagpole.

Starting the season: Team starting positions are heavily influenced by their late-season performances the previous year, and the off-season is completely ignored. There is no adjustment made for recovery from injuries, or players gained or lost via the draft or trade table, or anything else. For example, Collingwood started 2015 rated very low due to their injury-plagued end to 2014, while Adelaide and West Coast started in good positions after solid late-2014 performances.

Interactive Season Predictor

Drag teams around the chart and make Squiggle predict the rest of the season based on the new positions! It's the best of both worlds: your footy insight plus Squiggle's ability to sensibly model a season.

Reposition teams to your heart's content, open up the Predictor and click RECALCULATE.

This also provides a shareable link to the generated squiggle, so you can show off your work to other people.

What causes a team to move?

Teams move when they do better or worse than Squiggle expected.

For each match, Squiggle predicts a scoreline, given their current position and the venue at which the game will be played. If this turns out to be exactly right, the teams will remain where they are, since this means Squiggle has them positioned right. If, however, the scoreline is different, the teams will move. A team that scores more points than Squiggle expected will move ⇡ up; a team that scored fewer will move ⇣ down; a team that held its opposition to a lower score will move ⇢ right; and a team that allowed its opposition to score more will move ⇠ left.

Usually two of these things happen at once! A team will score either more or less than predicted, as well as allowing their opposition to score either more or less than predicted. So they will move on a diagonal:

↗ Scored more than predicted, held opponent to less than predicted

↘ Scored less than predicted, held opponent to less than predicted

↙ Scored less than predicted, opponent scored more than predicted

↖ Scored more than predicted, opponent scored more than predicted

How far a team moves depends on how different the result was from Squiggle's prediction. If the result was close to expectation, a team may barely budge. But an unexpected thrashing will cause a lot of movement.

Do teams get more movement against easy opposition?

No, because Squiggle expects better performances against weaker opponents, and to move to a better position, the team has to beat this expectation.

For the same reason, Squiggle isn't affected by fixture bias.

Can a team lose and still move into a better squiggle position?

Yes! Squiggle believes in honourable losses and shameful victories. If a team is expected to win by 10 goals but only prevails by 5, it will slide.

How is home ground advantage determined?

As described in the Model section, home ground advantage in Squiggle 2.0 is generated from ground familiarity: How often the teams have played at the same ground and in the same state over the preceding 4 years (including the current season).

Why doesn't Squiggle consider more factors?

1. I haven't seen enough evidence that it makes a difference, i.e. significantly improves tipping accuracy.

2. I want to keep the model as simple as possible, so squiggles can be a tool that enhances your ability to understand football, rather than an algorithm that replaces your own insight.

If you want a full-on predictor, there is some great stuff at Matter of Stats, Footy Maths Institute and Footy Forecaster.

What are Squiggle's weaknesses?

Possible flaws, which you may decide to compensate for, include:

• Squiggle doesn't consider a team's level of motivation, which seems to be fairly significant. Squiggle assumes all teams are trying equally hard at all times.

• Squiggle doesn't consider injuries or suspensions. It actually seems to me that people tend to over-estimate the impact of injuries, especially to key players. But clearly it's a bad thing for a team to lose a star player, and Squiggle doesn't take this into account at all. It rates teams, not players.

• Squiggle doesn't consider the impact of weather. It may reward a team for having a good defence when in reality what was making it hard to score was torrential rain. This doesn't seem to happen often enough to throw anything too far out, but does occur from time to time.

• Squiggle doesn't place any special value on wins. That is, it doesn't see much difference between a 1-point victory and a 1-point loss. It may thus underestimate a team that does "just enough," or is especially good at holding on in tight contests. Likewise, it may overestimate a team that regularly gets itself into winning positions against good teams but lacks the ability to close the games out.

• Squiggle gets very excited about very low-scoring games. Whether this is a flaw is arguable, since the model does seem to work. But Squiggle's maths mean that if a team is expected to keep its opposition to 80 points, and it actually keeps them to 40, this is considered twice as good, while keeping them to 20 points is considered four times as good, and keeping them to 10 points is eight times as good. Those are some big multipliers: to be 8 times better than a predicted 80 points in terms of Attack, a team would need to score 640 points (80 x 8), which is a lot more than the all-time AFL record. This means it's significantly easier for a team to move rapidly rightwards on a squiggle than upwards.

• The underlying model generates some inflation over the course of a season, which causes Squiggle to rate teams about 5-10% higher by the end of the year compared to the start. This seems useful for predictive purposes, as it allows for more movement leading into finals. However, it means that a team that doesn't actually change strength at all will be shown to slightly improve its chart position over the course of the season.

Why does the model use those values?

All the numbers used by Squiggle are that way because they worked best (i.e. made the most accurate predictions) when every possible combination was tested with a simulator replaying the last few decades.

How are the year's starting values calculated?

2015 starting positions are very similar to their end 2014 positions—the only difference is that 2013 data is no longer considered, so teams are modeled from the start of 2014 with each beginning on 50 ATTACK and 50 DEFENCE.

This means late-season 2014 results weigh quite heavily. For example, Collingwood had an injury-plagued end to 2014, and so is rated very low. Adelaide and West Coast, by contrast, finished the year with several solid performances, and so begin the year higher than you might expect.

What's with those crazy charts for the 1900s!?

Football scores were a lot lower a century ago, especially in the very early years, when single-digit scorelines abounded. Squiggle is calibrated for modern football, and thinks a game in which one team is held to a single goal (or no goals!) signifies an unbelievably good defensive effort. This causes teams to go shooting off to the right quite often in charts from the 1890s, 1900s and 1910s. So it's not a particularly good visualization of the strength of any particular team in that era. But it is interesting in terms of how different the whole league looks: how low and flat it is compared to today.

Similary, it can be interesting to look at where the mass of teams tends to sit in different decades; for example, how attacking the late 1980s was, with plenty of teams sitting high & centre/left compared to today.

The foundation of the Squiggle model is the OFFDEF engine, which rates teams separately in terms of attack and defense. Each team is initially assigned a starting value of 50 for each.

Scores are predicted for each match using the formula:

PREDICTED SCORE = 85 * (TEAM ATTACK ÷ OPPOSITION DEFENCE) + HOME ADVANTAGE

For example, in a match at a neutral venue between a team with ATTACK 56 and an opposition with DEFENCE 50, the team is predicted to score: 85 * (56 ÷ 50) + 0 = 95 points.

HOME ADVANTAGE is 6 when a team plays an interstate side at home, -6 when it plays an interstate side away, otherwise 0.

Predicted scores are compared to the actual scores, and ATTACK and DEFENCE adjusted accordingly. For example, if a team scored more highly than predicted, its ATTACK score needs to be increased, since Squiggle underrated it. Likewise, the opposition's DEFENCE score should decrease, since they failed to restrict the team as well as predicted. This is done by calculating what these scores would have to have been to predict the result perfectly, then constructing a weighted average of this along with all other results for the year, in which the most recent round comprises 9% of the final value.

At the start of a season, team starting points are calculated by doing the above for the previous season. For example, to calculate starting points for 2014, each team is assigned 50 to ATTACK and DEFENCE, then the 2013 season is played through.

The constants in this model (weighting the most recent round at 9%, considering 85 points a "typical" score, awarding a 12-point relative advantage to home sides playing interstate sides, and using exactly one previous season's results as a starting point each year) were chosen simply by simulating all possible values and seeing which produced the most accurate predictions over the course of the last 20 seasons or so.

The axis units are completely arbitrary, and entirely due to the choice of 50 as a starting value for each team's ATTACK and DEFENCE. They have no meaning except when comparing teams to each other.

Home Ground Advantage

In Squiggle v1, 12 points were awarded for HOME ADVANTAGE in games between interstate teams. Squiggle 2.0 uses a more nuanced model based on ground familiarity, where teams are compared on the number of times they have played at the same venue and in the same state. From this, up to 11 points of home ground advantage can be awarded (in the extreme situation where one team has played at the venue many times while the other hasn't even visited the same state), plus 2 points for hosting an interstate opponent at home.

Tip Probability

When determining "probable wins" in the Season Predictor, an algorithm is used that reflects the actual accuracy of Squiggle tips vs real-life results. Three factors affect the likelihood of a tip being correct:

• Margin: The greater the predicted margin, the more likely the tip is to be correct.

• Round Number: Games that occur later in the season are a little more likely to be tipped correctly.

• Weeks Until Game: Games that are weeks or months in the future are a little less likely to be tipped correctly.

Flagpole is tuned to predict premiers. In particular, it's good at:

• Rating the eventual premier highly throughout the year

• Picking the winner of the Grand Final

It's similar to eyeballing how close teams are to the ideal cluster of premiership cups on the Squiggle chart.

Record

Flagpole has a 15-5 record (75%) at picking Grand Final winners over the last 20 years (1997-2016). The years it got wrong are 1997, 2004, 2008, 2012, and 2016.

On average, it has rated the eventual premier within 1.55 rungs of top spot through the season. This is equivalent to rating the eventual premier 4th for half the season and 1st for the other half.

Why Flagpole?

If you eyeball the Squiggle chart, you can see that premiership cups aren't evenly distributed: there are more in the attacking area of the chart. So there's a slight difference between how likely a team is to win any random match (which is what the regular squiggle does), and how likely they are to win the flag. To predict the premier, we need to account for this.

Also, tipping finals one game at a time, like the Season Predictor does, is very dependent on knife-edge results, so that small changes can lead to very different precitions. Flagpole is more stable, has better historical accuracy, and is more consistent with what you see on the chart, i.e. that a team surrounded by premiership cups is in a better place than a very defensive one out on its own.

What Flagpole Rates

Compared to the regular squiggle algorithm:

• It ignores home ground advantage, looking only at final scores.
• It rates attack more highly.
• It is more sensitive to late-season results
• It is more sensitive to good results than bad ones—that is, it rates erratic teams that can unexpectedly kick big scores more highly than consistent ones that never do.

Why Flagpole Rates That Stuff

Because that's what premiers have done over the last twenty years. Explosive attack specialists such as Hawthorn and Geelong have won plenty of flags; defence merchants like Fremantle and St. Kilda have not.

Defensive teams have won premierships, of course, but far less frequently, even after accounting for how many of them there are. For example, in 2015, the Final Eight had five teams that were either balanced or attack specialists, and three defence-heavy teams in Fremantle (1st), Sydney (4th), and Richmond (5th). Those three lost every single game except when they played each other.

Flaws

Flagpole ignores the path a team must take to the flag. It ignores the difficulty of the games it must win, whether the team gets to play finals at home or away, how many interstate trips are required, whether it has a week off, and even how likely it is to make finals at all.

This can look pretty strange, e.g. it may rank a team 5th when it is going to miss finals. This will happen if a team displays the kind of form that would give it a decent shot at the flag if only it did make finals (such as a team making a run late in the year). For this reason, Flagpole can be occasionally definitely wrong about some mid-ranked teams compared to others, by ranking a team that won't make finals higher than one that will.

I haven't attempted to adjust for this because we don't have enough objective data on how beneficial it is to play finals at home, or travel less, or have a week off, or anything else. So Flagpole is purely a form rating.

Design

Like all my squiggle stuff, this came about by leaving the computer alone to brute-force hundreds of thousands of combinations of variables, algorithms, and filters, and coming back later to see what did well. The advantage of this method is that I don't need to be smart enough to design clever algorithms; I just find them. In fact, usually I don't even know why they work at first; I have to figure that out afterwards.

The disadvantage is it's easy to find algorithms that work well from pure chance, which fit the historical data well but don't actually have any predictive or descriptive power. To reduce the risk of this, I prefer simple algorithms with very few inputs (often only scorelines), and I use as much data as possible, like 20 years worth of results. But that still isn't much data when it comes to finals and Grand Finals.

How To Make Your Own Flagpole

Flagpole is an algorithm named OFFDEF4G-95:79.

1. Assign every team an ATTACK score of 50 and a DEFENCE score of 50
2. Look up the results of the first game of last year
3. Predict scores using the formula: TEAM A PREDICTED SCORE = 79 * TEAM A ATTACK / TEAM B DEFENCE
4. Compare predicted scores to actual scores by using the same relationship described above, so that TEAM A HYPOTHETICAL ATTACK = TEAM A ACTUAL SCORE * TEAM B DEFENCE / 79 and TEAM A HYPOTHETICAL DEFENCE = 79 * TEAM A ATTACK / TEAM B ACTUAL SCORE.
5. Here is the unique thing about OFFDEF4: square and double the HYPOTHETICAL ATTACK score. That is: NEW HYPOTHETICAL ATTACK = HYPOTHETICAL ATTACK * HYPOTHETICAL ATTACK * 2.
6. Assign new ATTACK and DEFENCE scores to each team, where 95% of the new score is the old score and 5% is the new score. That is: TEAM A NEW ATTACK = (0.95 * TEAM A OLD ATTACK) + (0.05 * TEAM A NEW HYPOTHETICAL ATTACK). Ditto for DEFENCE.
7. Repeat steps #2-#6 for every game.

That's it! Just one other thing: because this algorithm generates a lot of inflation over a season (which is good, I believe, because it means more more weighting for finals performances), its tips are only good as margins. If you want it to predict absolute scores, you need to increasingly scale back its predicted numbers, or else you get Grand Final tips where both teams are tipped to score well over 100.