Week 3 Forecasts: Florida at Auburn and Best Bets

How we forecast

We forecast each game with a language model trained to make calibrated judgments, not to write text. It takes a game's facts and its pre-game news and returns probabilities.

• What goes in: each team's ratings and form this season, the schedule spot, weather, injury and availability reports, and the pre-game coverage in plain English.
• What comes out: the chance each team wins, a spread of likely final margins, and the chance the game goes over the total.
• Two views per game: one sees the betting line and one never does. The line-blind view is our independent opinion. When it disagrees with Vegas, we look hard at why.
• No peeking: every input is published before kickoff. Every forecast is timestamped and saved before its game starts.

The hard problem: news in plain English

The news that matters most comes in late and in plain English. A starter is "questionable", a quarterback is "expected to play through" an ankle, or a coach hints at a rotation. Standard models need numbers, so they miss this or wait for someone to hand-code it.

Our model reads the report as written and updates the odds straight away. Re-running a game costs a fraction of a cent and takes seconds, so we re-forecast as news arrives:

• conference availability reports on Thursday, Friday and game day;
• weather updates;
• line moves;
• fresh previews.

Each game's final forecast uses everything published up to shortly before its kickoff.

How to read these numbers

• Win probability is our chance that a team wins outright.
• Our line is the point spread our model thinks is fair, from its forecast of the final margin. For single-number conversions we use the moneyline-to-spread relationship in about 4,100 college football games since 2013. It keeps the key numbers 3 and 7 intact.
• Cover chance is our probability that a team beats the Vegas spread. Final margins vary around our forecast with a standard deviation of about 15 points, the typical college football figure.
• Expected value is what a $100 bet at -110 would return on average if our cover chance is right. At -110 you need to win 52.4% of spread bets to break even.

Under the hood

Each forecast is a single model call. We send one structured "state" per game, up to about 30,000 tokens, and ask several narrow, typed questions at once.

flowchart LR
  A[Ratings and<br/>season form] --> S[Game state]
  B[Injury and<br/>availability reports] --> S
  C[Pre-game articles<br/>and weather] --> S
  S --> Q[Typed questions:<br/>win, margin, total, factors]
  Q --> M[Combine with<br/>the betting line]
  M --> F[Timestamped forecast]

News feeds the game state, the model answers narrow questions about it, and we combine those answers with the line before saving the forecast.

The questions. We don't ask "who wins?" and stop there.

• Margin: the model places the final margin on a ladder of buckets built around football's key numbers, 3 and 7. That gives a full distribution, which yields the win probability and the chance of covering any spread.
• Total: the same ladder approach gives the chance of going over the total.
• Factor questions: smaller questions cover quarterback availability impact, line-of-scrimmage edge, schedule spot and weather. Narrow questions answer more reliably than one big judgment.

Two views, then a blend.

• The market-aware view starts from the Vegas spread and asks how far the evidence moves it.
• The line-blind view sees the same news with every betting number scrubbed out, including odds quoted in articles.
• Where the two views share a bias, we subtract part of the blind view's move. That keeps the independent signal without double-counting.

Calibration. A forecast of 70% should win about 70% of the time. We map raw answers to calibrated probabilities with a symmetric fit, so the map can only stretch or shrink confidence. An early version quietly favored every home team; swapping home and away in test games caught it.

What testing says so far. On this season's first two weeks (185 games, each week fit on the other), our blended forecast:

• matched Vegas on picking winners (Brier 0.080 vs 0.080);
• beat a coin flip on covering the spread (0.240 vs 0.250).

On 3,925 games from 2021 to 2025, replayed week by week with only pre-game information:

• we add a little beyond the opening line, mostly after week 10;
• we add nothing beyond the closing line;
• when we leaned 10% or more off the opener, the line moved our way by about 3 points before kickoff.

That last result is the one we are building on.

Could the model just remember old results? It was trained on text from before this season, so it might recall past games. We tested that directly:

• On coin-flip games, where the spread was within 3 points, we asked with names and dates and again with teams anonymized.
• We used games played after the model's release as a control.

We found faint recall for 2021-2023 and none for 2025, so our historical tests use anonymized games.

Track record and caveats

This is our first public week, so there is no track record yet, only testing. Every forecast here was saved with a timestamp before its kickoff, and we will score all of them, misses included.

• How we keep score:
  • Brier score and calibration, not win-loss;
  • record against the spread, where 52.4% is break-even at -110;
  • closing-line value, meaning whether the line moved toward our pick before kickoff. Closing-line value shows real skill after dozens of bets; win-loss needs hundreds.
• Where we stand honestly: we have not shown we beat Vegas. The closing line is very sharp. Our tests suggest an early-week and late-season edge, not a weekly one.
• One week is noise. A great or terrible Saturday tells you little. Judge us over the season.
• Not betting advice. These are forecasts for discussion. Bet responsibly, if at all.