How AI predicts AFL games

There's no crystal ball in a footy model — just bookkeeping done relentlessly. Most public AFL models (the Squiggle family of models, ELO systems, the bookmakers' own) share the same skeleton. Here's what's inside it, in plain language.

Step 1: rate every team

The core is a rating per team — one number (or a small set) summarising how good they are right now. After every game, ratings update: beat a good side by more than expected and yours rises; scrape past a battler and it barely moves. ELO systems do this with a simple exchange formula; fancier models rate attack and defence separately or work from expected scores rather than actual ones (to strip out goal-kicking luck).

The elegant part: ratings are opponent-adjusted by construction. A four-game winning streak against bottom sides moves a rating far less than one win over the flag favourite — which is exactly the correction human "form" reading misses.

Step 2: turn ratings into a margin

For an upcoming game, the model takes the rating gap and adds situational adjustments:

• Home-ground advantage — mostly a travel effect. Hosting an interstate side is worth real points; sharing the MCG with your opponent is worth almost none.
• Rest and travel — six-day breaks, back-to-back interstate trips, post-bye bumps.
• Personnel — some models adjust for key outs; simpler ones let the next few results do the adjusting.

The output is an expected margin: "Cats by 16".

Step 3: margins become probabilities

This is why models speak in percentages, not verdicts. The number is the prediction — the winner is just the side with more than half the curve.

Why a losing 70% favourite doesn't break the model

A 70% chance fails three times in ten — that's what the number means. Judging a model on one game is like judging a coin on one flip. The real tests are run across seasons: calibration (do its 70% calls win about 70% of the time?) and Brier score / log-loss (how sharp and honest the probabilities are). Public AFL models publish these; the good ones tip around 68–72% of games and, more importantly, their percentages mean what they say.

Where the large-language-model AI fits

FootyAI's analyst is an LLM — brilliant at reading, reasoning and explaining, unreliable at remembering numbers. So it isn't asked to remember: before answering, it queries live data — fixtures, ladder, form, head-to-head, statistical-model probabilities, bookmaker odds, player stats — and writes its analysis from what comes back. The statistical model supplies the probabilities; the AI supplies the reading of them. Neither is asked to do the other's job.

What models still can't see

A ratings system doesn't know the coach sprayed the group on Tuesday, that the young ruck is managing soreness, or that it's someone's 300th in front of 90,000. Sometimes those things matter; mostly they're noise humans overweight. The honest position — and the one FootyAI takes — is that the model's number is the best starting point, and everything else is commentary to weigh, not certainty to bank. For how the numbers become betting decisions, see value betting.