NFL Week 9 AI Predictions, Graded: 14-1, and the Cost of Hedging
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 9 recap, with every AI prediction graded against the final score. Live forecasts for every game are on our NFL page.
NFL Week 9 AI Predictions vs Kalshi: Brier Scores
This was the best week of the season for everyone. Daggy and Kalshi both picked 14 of 15 winners, and both missed the same game: Green Bay's 45-27 win in New England. Kalshi still won on Brier score, 0.117 to 0.125 (lower is better), because it was more confident on the games both got right.
For the season, Kalshi leads 0.219 to 0.221 in Brier and 91-45 to 89-47 picking winners.
NFL Week 9 AI picks vs. results: every game
Probabilities are each forecaster's chance for the team it picked, made before kickoff. "Extremized" pushes Daggy's forecast away from 50-50 (explained below).
| Game | Daggy pick | Kalshi pick | Final | Daggy Brier | Kalshi Brier | Extremized Brier |
|---|---|---|---|---|---|---|
| Jaguars vs Ravens | Jaguars 74% | Jaguars 81% | Jaguars 33-10 | 0.066 | 0.036 | 0.019 |
| Bengals vs Falcons | Bengals 66% | Bengals 61% | Bengals 34-17 | 0.119 | 0.152 | 0.061 |
| Broncos vs Panthers | Broncos 68% | Broncos 57% | Broncos 26-21 | 0.104 | 0.185 | 0.047 |
| Cowboys vs Colts | Cowboys 62% | Cowboys 69% | Cowboys 27-23 | 0.148 | 0.096 | 0.095 |
| Jets vs Chiefs | Chiefs 78% | Chiefs 81% | Chiefs 35-12 | 0.049 | 0.036 | 0.010 |
| Lions vs Dolphins | Lions 73% | Lions 74% | Lions 19-9 | 0.075 | 0.068 | 0.024 |
| Browns vs Saints | Saints 60% | Saints 61% | Saints 23-22 | 0.162 | 0.152 | 0.112 |
| Giants vs Eagles | Giants 64% | Giants 62% | Giants 19-11 | 0.126 | 0.144 | 0.069 |
| Rams vs Commanders | Rams 64% | Rams 67% | Rams 20-13 | 0.132 | 0.109 | 0.076 |
| Texans vs Chargers | Texans 67% | Texans 67% | Texans 44-22 | 0.111 | 0.109 | 0.053 |
| Raiders vs 49ers | 49ers 76% | 49ers 81% | 49ers 38-17 | 0.059 | 0.036 | 0.015 |
| Packers vs Patriots | Patriots 59% | Patriots 62% | Packers 45-27 | 0.343 | 0.384 | 0.417 |
| Cardinals vs Seahawks | Seahawks 69% | Seahawks 79% | Seahawks 17-3 | 0.098 | 0.044 | 0.041 |
| Buccaneers vs Bears | Bears 65% | Bears 73% | Bears 27-24 | 0.122 | 0.073 | 0.065 |
| Bills vs Vikings | Bills 61% | Bills 63% | Bills 23-6 | 0.155 | 0.137 | 0.103 |
| Week 9 | 14-1 | 14-1 | 0.125 | 0.117 | 0.081 |
The extremized forecast would have won the week
Extremizing takes the average of our agents' forecasts and pushes it away from 50-50. We use a standard method (Platt scaling, with the slope from the forecasting literature): a 65% forecast becomes 74%, and a 78% forecast becomes 90%. Averaging eight forecasts tends to pull the result toward 50%, and extremizing undoes some of that.
In Week 9 it would have scored 0.081, far ahead of Kalshi's 0.117. It was better than our actual forecast on 14 of 15 games, and worse only on the one we got wrong, Packers vs Patriots.
We've been here before. In Week 1, extremizing would have helped, so we switched it on for Week 2, and it cost us in a week of upsets. The same thing would have happened this season: extremizing would have lost badly in Week 6 (0.303 vs 0.267) and Week 7 (0.266 vs 0.249). Over Weeks 4-9 it comes out even: 0.214 extremized against 0.213 for the plain average.
Extremizing is a bet that the favorites keep winning. In a chalky week like this one it wins big, and in an upset week it loses. So far the wins and losses have cancelled out.
Why were our agents less confident than Kalshi?
In every week since Week 4 except Week 6, our forecast has sat closer to 50-50 than Kalshi's. In Week 9 our average forecast was 17 points from 50%, against Kalshi's 19. Part of that is averaging, and part of it is the agents themselves.
The agents keep notes between weeks, including a record of how their own forecasts have done. Earlier this season, in a run of upset weeks, those notes said their confident picks had come in too often for the favorite. In about 20 of 120 Week 9 forecasts, an agent cited its own calibration record to pull toward 50%:
- Gemini with the engine, on Bucs at Bears: "Earlier season calibration tracking … highlights that forecasters historically showed overconfidence on moderate favorites." It held Chicago at 68%. Chicago won.
- Opus with the engine, on Packers at Patriots: "My own record shows forecasts in the 0.6-0.8 range have run high." It trimmed New England to 58%.
The lesson was right for the upset weeks it came from and wrong for this one. One agent was starting to notice. Kimi without the engine, on Jets at Chiefs: "My calibration ledger warns my mid-range overfavors home but that pull-downs on strong home leans have cost five right-side games." It stayed at 75% on Kansas City.
Packers vs Patriots AI prediction: the one game everyone missed
All eight agents picked New England, at 53-62%, and so did Kalshi (62%). Green Bay won 45-27.
The agents' case was reasonable. New England rated 71 Elo points higher, was 3-0 at home, and Micah Parsons was expected to miss the game for Green Bay. Kimi without the engine leaned on the weather, with ~30 mph gusts and rain forecast in Foxborough, writing that it would punish Green Bay as "pass-dependent." Green Bay scored 45.
One thing in Green Bay's favor did get noticed. It had 10 days of rest to New England's 7, and Opus with the engine tested it on 2010-25 games: "a home team with 3 or more fewer rest days did about 2.5 points worse." That cut its forecast from about 64% to 61%.
This is the cost of extremizing in one game. Turning our 59% into 65% would have made the miss worse (0.417 vs 0.343 Brier).
Did the agents with the engine win again?
Barely. The four agents with Daggy's engine beat the four without it 0.124 to 0.126, and won 52% of head-to-head pairs. After Week 8's record gap of 0.018, it was the smallest edge since Week 6. When almost every favorite wins, there's little room for one side to separate.
| Week | Pairs | Engine minus no-engine (Brier, lower = engine better) | Engine won the pair |
|---|---|---|---|
| 3 | 64 | −0.010 | 53% |
| 4 | 64 | −0.011 | 53% |
| 5 | 60 | −0.012 | 55% |
| 6 | 56 | −0.000 | 50% |
| 7 | 56 | −0.013 | 59% |
| 8 | 56 | −0.018 | 59% |
| 9 | 60 | −0.003 | 52% |
The engine's best game was Cardinals at Seahawks, where Opus showed the biggest split. With the engine it had Seattle at 72%. Without the engine it had a coin flip, 52%. Seattle won 17-3.
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, including its own calibration record.
- 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 9? Daggy's agents picked 14 of 15 winners with a Brier score of 0.125. Kalshi's market also picked 14 of 15, with 0.117.
What does extremizing a forecast mean? Pushing an averaged probability away from 50%, because averages of many forecasters tend to be too cautious. It helps when favorites win and hurts in upset weeks.
Previous: Week 8: The agents with data pull ahead · All roundups