NFL Week 8 AI Predictions, Graded: The Agents With Data Pull Ahead
Every week this NFL season, we're pitting 8 AI agents (4 with Daggy's statistical engine and 4 without) against Kalshi's prediction market. Here's the Week 8 recap, with every AI prediction graded against the final score. Live forecasts for every game are on our NFL page.
NFL Week 8 AI Predictions vs Kalshi: Brier Scores
Kalshi won the week: a Brier score of 0.177 to Daggy's 0.188 (lower is better), and 10 of 14 winners to Daggy's 9. For the season, Kalshi leads 0.231 to 0.233 in Brier and 77-44 to 75-46 picking winners.
The bigger story was inside our panel. The four agents with Daggy's engine beat the four without it 0.181 to 0.198, their widest margin of the season. If we had shipped only the engine agents' average, we'd have scored 0.181, closer to Kalshi than our actual forecast.
NFL Week 8 AI picks vs. results: every game
Probabilities are each forecaster's chance for the team it picked, made before kickoff. "Engine-only" is the average of just the four agents with Daggy's engine.
| Game | Daggy pick | Kalshi pick | Final | Daggy Brier | Kalshi Brier | Engine-only Brier |
|---|---|---|---|---|---|---|
| Panthers vs Packers | Packers 65% | Packers 65% | Packers 27-19 | 0.124 | 0.122 | 0.122 |
| Ravens vs Bills | Bills 63% | Bills 63% | Bills 36-10 | 0.135 | 0.137 | 0.102 |
| Titans vs Bengals | Bengals 80% | Bengals 84% | Bengals 43-36 | 0.040 | 0.026 | 0.024 |
| Cardinals vs Cowboys | Cowboys 69% | Cowboys 75% | Cardinals 33-30 | 0.471 | 0.562 | 0.473 |
| Vikings vs Lions | Vikings 51% | Lions 63% | Lions 19-15 | 0.258 | 0.137 | 0.304 |
| Colts vs Jaguars | Jaguars 76% | Jaguars 81% | Jaguars 42-7 | 0.058 | 0.036 | 0.040 |
| Raiders vs Jets | Jets 52% | Jets 54% | Raiders 38-33 | 0.273 | 0.292 | 0.250 |
| Browns vs Steelers | Steelers 62% | Steelers 64% | Steelers 27-6 | 0.141 | 0.130 | 0.135 |
| Falcons vs Buccaneers | Buccaneers 52% | Buccaneers 57% | Buccaneers 16-0 | 0.226 | 0.185 | 0.221 |
| Chargers vs Rams | Rams 72% | Rams 74% | Rams 24-9 | 0.079 | 0.068 | 0.073 |
| Chiefs vs Broncos | Chiefs 51% | Chiefs 52% | Broncos 28-24 | 0.258 | 0.270 | 0.172 |
| Patriots vs Dolphins | Patriots 72% | Patriots 76% | Patriots 24-19 | 0.080 | 0.058 | 0.069 |
| Eagles vs Commanders | Commanders 60% | Commanders 70% | Commanders 27-10 | 0.156 | 0.090 | 0.176 |
| Bears vs Seahawks | Seahawks 58% | Seahawks 61% | Bears 34-17 | 0.336 | 0.372 | 0.378 |
| Week 8 | 9-5 | 10-4 | 0.188 | 0.177 | 0.181 |
Chiefs vs Broncos AI prediction: the agents split by who had the data
This was the cleanest split we've seen all season. All four agents with Daggy's engine favored Denver (55-62%). All four without it favored Kansas City (Denver 36-44%). Our blended forecast landed at a coin flip, and Kalshi had the Chiefs at 52%. Denver won 28-24.
It's a rematch of the question from Week 7's Seattle game: does a team's record and box score tell you more than its rating? Kansas City came in 5-1 with a much better point differential (about +6.7 a game to Denver's +0.4) and a 31-10 win over Denver in Week 1. Denver came in 5-2 but rated 62 Elo points higher, at home.
The agents without the engine read the box score. Each built its case from this season's stats:
- GPT: Kansas City "averages 380.5 offensive yards per game versus Denver's 270.1, and won their September meeting 31-10 while rushing for 220 yards." Denver 36%.
- Gemini: Kansas City ranks 3rd in net EPA per play and Denver 16th. Denver 40%.
- Kimi: a "~6-point neutral-site gap" in point differential, with "home field at altitude" clawing back about 2.5. Denver 40%.
The agents with the engine started from history: 26 years of games in their Daggy table, before the box score.
- Gemini cited the four models it had trained during an earlier weekly review, which put Denver between 57% and 65%. Its own query found that divisional home teams with a similar Elo edge, but a worse point differential, won 53.6% of the time (59 of 110). Denver 55%.
- Kimi pulled divisional home teams with a similar Elo edge: they won 60% of the time (n=420). Denver 62%.
- GPT's historical model on Elo, scoring margins and rest said 62%. It settled at 60%, writing that "the single head-to-head result and unconfirmed injuries do not justify a larger adjustment."
- Opus made a single data call. It started from Elo (about 62% for Denver) and trimmed for Kansas City's point differential using a test from its own notes. Denver 57%.
The engine agents saw the same 31-10 game and the same point differential. They gave them less weight because they had looked up how often those things beat a rating. This was the engine side's best game of the season: −0.188 Brier per pair.
Where the engine lost: Vikings vs Lions and Bears vs Seahawks
The same habit cost the engine side two games.
Minnesota came into Detroit 6-0, with a 68-point Elo edge and a defense allowing 12.3 points a game. The engine agents trusted that: all four had Detroit between 42% and 47%. GPT's two historical models "both put Detroit near 46%." Kimi without the engine took Detroit to 61%, arguing that Minnesota's defensive numbers "were built against a weak slate" and that Detroit "is the best offense they've faced." Detroit won 19-15, and the engine side gave back +0.085 per pair.
In Seattle, the engine agents had the Seahawks at 58-65%; the agents without it were at 50-61%. Chicago won 34-17 (+0.080).
What if we only listened to the engine agents?
We'd have scored 0.181 instead of 0.188. And if we had also extremized that average (pushed it away from 50-50, as we did in Week 2), we'd have scored 0.172 and beaten Kalshi's 0.177.
That's hindsight, so we're not switching yet. One week isn't enough to drop half the panel, and the panel exists so we can measure this gap. But the gap keeps showing up.
Do AI agents with a statistical engine make better NFL picks?
So far, yes, by a small and steady margin since the agents stopped seeing betting lines in Week 3. Each pair is the same model on the same game, once with the engine and once without:
| Week | Agents | Pairs | Engine minus no-engine (Brier, lower = engine better) | Engine won the pair |
|---|---|---|---|---|
| 3 | blind | 64 | −0.010 | 53% |
| 4 | blind | 64 | −0.011 | 53% |
| 5 | blind | 60 | −0.012 | 55% |
| 6 | blind | 56 | −0.000 | 50% |
| 7 | blind | 56 | −0.013 | 59% |
| 8 | blind | 56 | −0.018 | 59% |
Week 8 was the biggest gap so far, and all four models were ahead with the engine: GPT −0.028, Kimi −0.024, Opus −0.010, Gemini −0.009. Over Weeks 3-8, the engine side is about 0.01 better per forecast.
One reason may be what the agents now do between weeks. Since Week 6, the engine agents use Daggy during their weekly review to test the rules in their notes against historical games. In Week 8 some of those tested rules showed up at forecast time. Opus's point-differential adjustment and Gemini's four models both came from those reviews.
How our AI NFL predictions work
- Eight AI agents per game. Four models (Kimi K3, Claude Opus 5.5, GPT-6 Sol and Gemini 3.8 Flash), each run twice: once with Daggy's statistical engine and once without.
- Blind to the betting market. The agents search the web for news and injuries, but betting lines and odds are withheld so they can't copy the market (why).
- They learn between weeks. Every agent writes a post-mortem on each game and keeps notes on every team.
- One data-driven forecast. The eight probabilities are averaged into Daggy's prediction, made hours before kickoff and never changed.
- Graded every week against Kalshi's prediction market with the Brier score.
Top questions, answered
How accurate were the AI NFL predictions in Week 8? Daggy's agents picked 9 of 14 winners with a Brier score of 0.188. Kalshi's market picked 10 of 14 with 0.177.
Do AI agents with data analysis tools make better NFL picks? In our experiment, slightly. Through Week 8, agents with Daggy's statistical engine beat the same models without it by about 0.01 Brier per forecast, and by 0.018 in Week 8.
Previous: Week 7: Stats vs. Vibes · Next: Week 9: Kalshi and Daggy both go 14-1