AI & Data
Could artificial intelligence have predicted the 2026 NRL Grand Final?
On paper, Roosters vs Knights looked like a useful test. Sydney entered the decider as the stronger statistical side. Newcastle had produced one of the season’s great finals runs, climbing from seventh and reaching the Grand Final only a year after collecting the wooden spoon. And by full-time, the match had become exactly the sort of event that exposes both the power and the limits of sports prediction.
The Sydney Roosters beat the Newcastle Knights 19–18 at Accor Stadium after Newcastle raced to a 12–0 lead. Daly Cherry-Evans’ 70th-minute field goal ultimately became the difference. The NRL’s official match report recorded 80,814 fans and a one-point finish that remained alive until the final seconds.
So how did the machines do?
Surprisingly well — and badly — at the same time.
AI models picked the Roosters before kickoff
We didn’t ask a chatbot to guess the winner after feeding it a few statistics. Instead, we looked at public machine-learning predictions that were published before the match.
Alphr’s NRL model gave the Roosters a 69% win probability. It expected Sydney to win by about 8.3 points and forecast 41 total points.
Stats Insider, which says it simulates matches 10,000 times using machine learning and data, was even more bullish by match day. Its updated model gave the Roosters a 77% chance of winning and predicted a 27–16 scoreline.
A separate Dimers simulation put Sydney at 71.4%.
| Prediction | Roosters win probability | Expected margin/score |
|---|---|---|
| Alphr | 69% | Roosters by ~8 |
| Stats Insider | 77% | Roosters 27–16 |
| Dimers | 71.4% | Roosters favoured |
| Actual result | — | Roosters 19–18 |
All three got the winner right. None captured just how close the game would become.
That doesn’t mean AI “predicted” the Grand Final
This is where sports and AI terminology often get messy.
A model assigning the Roosters a 69% probability is not saying Sydney will definitely win. It is saying that, given the patterns the model has learned and the information available before kickoff, outcomes resembling a Roosters victory occur more often than outcomes resembling a Newcastle victory.
A 31% chance of a Knights win is still enormous. If a similar event happened repeatedly, an outcome with that probability should occur roughly three times in ten.
And the final score demonstrates an even bigger limitation: predicting the most likely winner is much easier than predicting how a match will unfold.
Newcastle led 12–0. Kalyn Ponga produced a 96-metre try. Mark Nawaqanitawase scored twice. The Roosters eventually went seven points clear, only for Greg Marzhew’s late converted try to turn the final minutes into a one-point contest.
A pre-match model can estimate the likelihood of those kinds of events. It cannot know their sequence.
What does an actual NRL AI model look at?
The technology is more interesting than simply typing “Who will win Roosters vs Knights?” into ChatGPT.
Alphr publicly describes its NRL system as a four-model XGBoost ensemble trained on more than 18 years of NRL data. Its system uses roughly 160 engineered features, including team and player ELO ratings, recent form, venue, weather, referee tendencies, positional lineup strength, rest days, head-to-head history and other pre-match information.
The models perform different jobs: one estimates win probability, another predicts margin, while separate regression models estimate each team’s score.
For this Grand Final, the system identified several numerical advantages for Sydney. Its team ELO rating had the Roosters at 1646 compared with Newcastle’s 1574. It also rated the Roosters higher through the forwards and halves, while Newcastle’s strongest positional advantage was around hooker and the ruck.
That became particularly interesting because Newcastle entered the match without suspended hooker Phoenix Crossland, while the Roosters were without Connor Watson after he was not cleared following a head injury assessment.
Those are exactly the kinds of changes a useful sports model needs to incorporate. A model trained on historical results but unaware of the actual team taking the field isn’t analysing the same match.
The prediction even changed before kickoff
There is another detail that shows how these systems work.
An earlier Stats Insider match snapshot had the Roosters at around 71% and projected a 26–18 score. Its later match-day prediction moved to 77% and 27–16.
That isn’t necessarily a contradiction. Predictive systems can change as their inputs change — confirmed teams, injuries, form data and other information can alter the probability.
This is one reason a screenshot of an “AI prediction” without a timestamp can be misleading. The useful question isn’t just what did the model predict? It is what information did it have when it made that prediction?
Why ChatGPT isn’t the same thing as a sports prediction model
Generative AI and predictive machine learning are often lumped together, but they solve different problems.
A large language model can explain team statistics, compare arguments and reason about information it is given. But asking a chatbot for a percentage does not automatically make that percentage statistically calibrated.
A dedicated predictive model can be trained and tested against thousands of historical matches, with its probabilities measured against what actually happened. One common metric is the Brier score, which penalises probability forecasts for being both wrong and overconfident.
Alphr, for example, publishes its historical results and says its NRL head-to-head model achieved a 61.5% strike rate in its held-out 2025 season. Its current methodology uses a strict time-based training and validation process designed to prevent future match information leaking into historical training data.
That transparency matters more than an impressive-sounding “AI prediction”.
So, did AI predict Roosters vs Knights?
It predicted the favourite correctly. It did not predict the game.
The models saw the Roosters as the more likely champions, and Sydney ultimately won. But the difference between an expected eight-to-eleven-point advantage and the actual one-point result is the important part of the experiment.
Sport contains injuries, refereeing decisions, bouncing balls, fatigue, tactical changes, individual brilliance and plain randomness. Those aren’t proof that predictive AI is useless. They are why its output should be expressed as probability rather than certainty.
The most impressive AI sports model isn’t one that claims it knows who will win.
It’s one that knows how uncertain its prediction really is.
Sources: NRL official Grand Final coverage; Alphr model prediction and methodology; Stats Insider pre-match simulations; Dimers pre-match simulation. This article analyses prediction technology and is not betting advice. Featured image: Daniel Anthony / Unsplash.
