Seven points. That’s the margin that decided the 2026 AFL premiership, and it’s worth sitting with for a moment before we talk about prediction models. Brisbane Lions 96, Fremantle 89 at the MCG on Saturday. A one-score game, settled in the last quarter, between two teams who finished the season separated by exactly three wins on the ladder.
Our model went into that match with Fremantle at 55.1%. The bookmakers — who had real money on the line — had Brisbane as the more likely winners. The market’s read was sharper. This piece is an honest post-mortem on why.
Brisbane def. Freo
season record
season record
GF probability
Model vs Market vs Reality
Here is the Grand Final prediction in full. Three perspectives on the same game, and only one of them got it right.
The gap between those two probability estimates — 12.4 percentage points — is large. The model had Fremantle as a clear favourite. The market had Brisbane. One of them had a fundamental information advantage, and it wasn’t the model.
Why the Model Got It Wrong
Two failures, both addressable.
The first is a venue-coding bug. The Grand Final at the MCG is a neutral-venue game. Our model designated Fremantle as the nominal “home team” for scheduling purposes, which injected approximately 10 percentage points of home-ground advantage that did not exist in practice. Strip that out and model Fremantle vs Brisbane as a genuine neutral contest, and the model’s raw probability falls to roughly 45% — almost exactly where the market sat.
The second is a recent form weighting problem. Brisbane’s regular-season record (16–7) was worse than Fremantle’s (19–4) by three wins. But the model’s weighting of season-long data didn’t adequately capture the directional momentum that the bookmakers were pricing. In the five games before the Grand Final, Brisbane averaged a +24-point winning margin. Fremantle’s trajectory over the same window was flat, including a scratchy Preliminary Final against Sydney that they won by 12 after trailing at three-quarter time.
At the time of the Grand Final, Brisbane’s ELO rating was 1,710. Fremantle’s was 1,704. The model’s own opponent-adjusted metric showed Brisbane as the marginally better team — it just couldn’t express that because the venue was coded incorrectly. The model had the right information and produced the wrong output. That’s a software problem, not a fundamental modelling problem, and it will be fixed.
Brisbane’s Season in Full
It is easy to look back on a premiership and frame the winner as inevitable. Brisbane’s season was not inevitable. They finished 16–7, third on the ladder behind Fremantle (19–4) and Sydney (18–5). They lost to Geelong by 41 at home in Round 10. They got demolished by Greater Western Sydney by 78 points in Round 11 — one of the worst performances of the season by any top-eight side. Then in Round 21, they were beaten at home by Carlton by 76 points.
And yet. When Brisbane were good this season, they were dominant in a way that the win-loss record undersells. Their average winning margin across the regular season was +20.5 points. They crushed Essendon 149–59 in Round 18. They put 159 on Collingwood in Round 24. In the Semi-Final, they took Adelaide apart by 53. The Lions were a team with a wide distribution — catastrophic lows and extraordinary highs — and in September, the highs won.
How Brisbane Got There
The path to the flag is the most compelling part of this story, because it started with a disaster. Brisbane lost to Sydney by 53 in Finals Week 1 — one of the biggest hiding the Lions had taken all year, from a team they’d beaten by 43 back in Round 16. Most teams don’t bounce back from a loss like that. Brisbane did — with a 53-point win over Adelaide the following week, almost precisely mirroring what had been done to them.
| Round | Opponent | Score | Margin | Result |
|---|---|---|---|---|
| Finals Week 1 | Sydney (away) | 88–141 | –53 | Loss |
| Semi-Finals | Adelaide (home) | 144–91 | +53 | Win |
| Preliminary Finals | Hawthorn (away) | 131–122 | +9 | Win |
| Grand Final | Fremantle (MCG) | 96–89 | +7 | Win |
The Preliminary Final against Hawthorn was the pivot. Brisbane were away at the MCG — a ground they’d now play their Grand Final at — against a team that had beaten them during the regular season. The model rated it as a coin-flip. Brisbane won by 9. From that point, they had momentum, MCG familiarity, and the psychological bedrock of having survived three weeks of elimination football. Fremantle had not played a final outside Western Australia all year.
Why the Market Was Right
The bookmakers were pricing three things our model underweighted.
Neutral venue. Fremantle’s 19–4 regular-season record was built significantly at Optus Stadium, where they are genuinely exceptional. The MCG on Grand Final day removed that edge entirely. Brisbane, having won there in the Preliminary Final, were more comfortable than Fremantle on the day.
ELO vs win rate. ELO adjusts for who you beat and by how much. Fremantle’s 19-4 record included several large wins over bottom-half opposition. Brisbane’s 16-7 record included beating Geelong away (Round 17) and Collingwood away (Round 24) when the model had them as underdogs in both. Those results built ELO points that the raw win rate doesn’t capture.
September form. Brisbane averaged a +24-point margin across their last five games before the Grand Final. Fremantle’s equivalent number was near zero. Markets weight recent form more heavily than our model does. In this case, they were right to.
What We’re Fixing for 2027
The Grand Final is our biggest miss of the 2026 season. Overall model accuracy was 68.1% — respectable for a model in its first full AFL year. But a 12-percentage-point error on the Grand Final, driven by a known coding issue, is not acceptable.
Three changes before Round 1, 2027:
1. Genuine neutral-venue handling. The Grand Final at the MCG will be coded as a neutral-venue game with no home-ground advantage applied to either team. Any interstate finals held outside home grounds will receive the same treatment.
2. Finals form recency multiplier. A team’s last five games will receive heavier weighting in finals contexts. The regular season and September are different environments; the model needs to reflect that.
3. ELO-vs-win-rate balance. For high-variance teams like Brisbane — whose ceiling was consistently higher than their record implied — the model will weight ELO more heavily relative to raw win rate. Win rate can be artificially suppressed by early-season losses against quality opponents. ELO adjusts for that.
Brisbane Lions won the 2026 AFL premiership. The market had it right, and they earned it. The model saw a team that was third on the ladder and gave Fremantle a nine-point advantage it didn’t have. Next year, we’ll do better.