Arbs › Guides › How AI sports models price games
How AI sports models price games
Arbs research desk · Updated 2026-11-01 · Reviewed 2026-11-01 · 2 min read
The short answer: A sports model turns information into a probability for each side. The strongest inputs are usually the betting market itself, team strength and recent form. AI language models help summarize news, but the probability should come from a calibrated quant model, and the proof is a public record.
Inputs
- Market consensus: no-vig prices from several books
- Team and player strength: records, ratings, home field, rest
- News: injuries and lineups, which move prices fastest
Pooling in log-odds
Probabilities from different sources are usually combined in log-odds rather than averaged directly, so a confident source near 90% is not dragged toward the middle as hard:
logit(p) = ln(p ÷ (1 − p)); the pooled logit is the weighted average of the sources’ logits
Worked example: Two books say 90% and 70%
- Logits: 2.197 and 0.847
- Average logit: 1.522
- Back to a probability: 1 ÷ (1 + e^−1.522)
82.1%, against 80.0% for a plain average.
Calibration
A good model checks calibration: games it calls 60% should win about 60% of the time. Without it, a model can sound confident and be wrong in a consistent direction. A common fix is Platt scaling, which refits the model’s log-odds on graded games:
calibrated p = 1 ÷ (1 + e^−(a × logit(p) + b))
Scoring a model
Two standard scores reward honest probabilities. Lower is better for both:
Brier = average of (p − outcome)²; log loss = −average of [o × ln p + (1 − o) × ln(1 − p)]
Worked example: Ten games called at 70%, seven won
- Winners: 7 × (0.7 − 1)² = 0.63
- Losers: 3 × (0.7 − 0)² = 1.47
- Brier: (0.63 + 1.47) ÷ 10
0.21, against 0.25 for always saying 50%.
The bar that matters for betting is not a coin flip but the market. A model is only useful where it scores better than the price you would bet against.
Where language models fit
Large language models are useful for reading injury reports and writing up a game. They are not reliable probability engines on their own. On Arbs an AI judge may move the quant probability by at most 6 points after reading the news, and the AI read explains the number without choosing the pick. The methodology page has the details.
How to judge any model
- Picks locked before kickoff, with a timestamp you can verify
- Every loss kept on the record
- Positive closing line value, not just a hot streak
- A Brier score or log loss compared with the market’s, not only with a coin flip
The Arbs track record is built that way.
Frequently asked questions
Can AI predict sports games?
Models can estimate probabilities, sometimes better than the market on specific games. Nobody predicts outcomes reliably; the question is whether the probabilities beat the price over many bets.
How accurate is the Arbs model?
Every pick is locked before kickoff and graded from Polymarket. The win–loss, return per unit and closing line value are on the public track record.
What makes a sports model trustworthy?
A long public record of picks made before the game, including the losses, and evidence that its prices beat the closing line.