October 27, 2026

NFL Week 7 AI Predictions, Graded: Stats vs. Vibes vs. Kalshi

By Geo Optimo · tuned for AI search with Daggy

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 7 recap, with every AI prediction graded against the final score.

NFL Week 7 AI Predictions vs Kalshi: Brier Scores

Kalshi won the week and the head to head: Brier 0.242 vs Daggy's 0.249, and 9 of 14 winners to Daggy's 8 of 14. This puts Kalshi up for the season by one game picking winners (67-40 to 66-41) and by 0.001 in Brier (0.238 to 0.239).

Our agents with Daggy's engine beat the ones without it 0.244 vs 0.257 (each side's four forecasts averaged per game). That's the fifth time in the last five weeks that the engine side has been ahead or level.

NFL Week 7 AI picks vs. results: every game

Probabilities are each forecaster's chance for the team it picked, made before kickoff. Brier score is the squared error of the probability (0 is perfect, 0.25 is a coin flip; lower is better).

Game	Daggy pick	Kalshi pick	Final	Daggy Brier	Kalshi Brier
Patriots vs Bears	Bears 54%	Bears 60%	Bears 30-24	0.212	0.160
Steelers vs Saints	Steelers 62%	Steelers 62%	Steelers 16-10	0.143	0.144
49ers vs Falcons	49ers 65%	49ers 59%	Falcons 29-20	0.428	0.348
Bengals vs Ravens	Ravens 55%	Ravens 53%	Bengals 35-13	0.300	0.281
Buccaneers vs Panthers	Panthers 59%	Panthers 61%	Buccaneers 43-13	0.351	0.372
Giants vs Texans	Texans 67%	Texans 80%	Giants 27-21	0.452	0.640
Colts vs Vikings	Vikings 79%	Vikings 72%	Vikings 16-10	0.046	0.078
Dolphins vs Jets	Jets 60%	Jets 67%	Dolphins 28-25	0.365	0.449
Browns vs Titans	Browns 58%	Browns 54%	Browns 27-24	0.173	0.212
Broncos vs Cardinals	Broncos 66%	Broncos 66%	Broncos 30-16	0.114	0.116
Packers vs Lions	Lions 59%	Lions 62%	Lions 20-17	0.166	0.144
Rams vs Raiders	Rams 50%	Rams 65%	Rams 35-21	0.246	0.122
Chiefs vs Seahawks	Chiefs 54%	Seahawks 57%	Seahawks 20-3	0.289	0.185
Cowboys vs Eagles	Cowboys 55%	Cowboys 64%	Cowboys 38-21	0.205	0.130
Week 7	8-6	9-5		0.249	0.242

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.
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.

Where data analysis helps, and where it hurts

Part of the goal of this agentic forecasting experiment has been to see how well we can marry quantitative data with qualitative data. Quantitative data does great to measure homefield advantage, for example, but it is less clear how we adjust our predictions when a charismatic quarterback like Jackson Dart is replaced by someone like JJ McCarthy who is now on his second team. Of course, we can use metrics like wins above replacement, but rather than get in the weeds with metrics like this we can see if we can just let the AI agents reason through the impact.

This week had several games where we had qualitative vibes differing from more quantitative data. Specifically, the 2-4 Rams taking on the 3-3 Raiders, where most perceive the Rams as being much better than their record attests; as well as the 5-0 Chiefs against the 3-3 Seahawks. On the new quarterbacks, the Eagles had a backup quarterback going against the Cowboys and the Giants started their newly acquired JJ McCarthy to go against the Texans. How did the models work through the data and the vibes to make their predictions?

Rams vs Raiders AI prediction: "better than their record" (data helped)

Kalshi gave the Rams a 65% chance to beat the Raiders. But our ensemble predicted a tossup. But most of Daggy's prediction was driven by agents without access to the statistical engine. Furthermore, the agents agreed that the Rams were the better team, but hedged their bet because of injuries. The Opus agent without the engine wrote, "The Rams look like the better team underneath their 2-4 record." But they then noted the Rams' injury list: Davante Adams, Poona Ford, Kobie Turner and Myles Garrett on IR, plus CB McDuffie and LB Landman missing Wednesday practice. The four no-engine agents ended at LV 0.48-0.57, so the Raiders were their favorite on average.

The engine agents stayed closer to the numbers:

Opus (engine), LV 0.36. It started from the 140-point Elo gap and cited a rule from its own notes: shading toward the underdog when the Elo gap is over 100 "has cost me on the Brier score 7 times out of 7." The injuries got "only a small nudge."
GPT (engine), LV 0.44. Its own model on Elo, last season's margin and rest said 41%.
Gemini (engine), LV 0.47. Its models said 38-45%, and it added injuries on top.
Kimi (engine), LV 0.55, the exception. It threw out the Elo gap as "almost entirely stale 2025 information" and called that a "tested" lesson. On its notes card, that lesson has post-mortem citations only and no data test.

Chiefs vs Seahawks AI prediction: the record vs. the rating (data helped most)

Kalshi picked Seattle (3-3) over the Chiefs (5-0) at 57% while Daggy predicted Kansas City to win with 54%. Seattle won 20-3. The records, however, didn't play much of a role in the agents' reasoning. Two of the engine agents had considered the impact of a record during a review of their notes:

Opus (engine), SEA 0.61. Its test: "a much worse home record added nothing beyond Elo (n=178)," over weeks 5-9 since 2010.
GPT (engine), SEA 0.61. Its test: adding recent scoring margin "did not improve leave-one-out fit" once Elo, last season's margin and rest were in the model (3,958 games).

The cautionary tale is Gemini (engine). Its own four models said Seattle 61-66%. It then reasoned its way down to 0.46 with a story: KC's Elo was low only because of Mahomes' 2025 knee injury, and Kenneth Walker had moved from Seattle to KC. That's a reasonable story, and it was wrong this time. Its own notes warn: "Subjective narrative adjustments overriding models inflate tails." Kimi (engine) also re-rated off current form ("2025-inflated Elos persist (SEA)") and landed at 0.42.

This was the engine side's biggest win of the season's blind era on one game: −0.128 Brier per pair. In this case, the qualitative information seemed to hurt the models and those who had access to Daggy's statistical engine were able to not let the qualitative information override the statistics.

Cowboys vs Eagles AI prediction: when the data can't see the injury (data hurt)

Jalen Hurts is out for the season (deep-vein thrombosis, clots in both lungs; on IR 10-17). Tanner McKee made his second start. Kalshi had Philadelphia 36%, and Dallas won 38-21.

Every agent knew. Hurts' status was already in all 8 agents' memory notes from last week's reflections, so none of them had to search for it. Knowing wasn't the problem. Sizing it was, and this time the data was the anchor that held them back:

Opus (engine), PHI 0.55. It started from an Elo-implied 61% and trimmed only 6 points. Its notes say cutting again once the backup is already priced into Elo "has gone 1-2." That's an untested rule from 3 games, and Elo had seen only one McKee start.
GPT (engine), PHI 0.49. It said its own model "cannot directly measure" the QB downgrade, then landed near the model anyway.
The no-engine agents (PHI 0.36-0.44) were closer to the market. GPT without the engine had PHI at 0.36, almost exactly Kalshi.

Elo knows nothing about who plays quarterback. Our table has no QB column. So an agent that trusts the table has to apply the injury itself, and the engine agents under-applied it. This was the engine side's second-worst game of the week (+0.073 per pair).

Giants vs Texans AI prediction: a backup is not a backup is not a backup

The Giants (3-2-1) started J.J. McCarthy at 2-4 Houston. Houston was a 10-point favorite (Kalshi: Houston 80%), and the Giants won 27-21. McCarthy went 20/28 for 252 yards, no TD, no INT, +5.6 passing EPA.

All 8 agents treated McCarthy as a QB downgrade, and every one of them moved toward Houston:

Opus (engine) called him "effectively their third-string QB."
Gemini (engine) cited "severe execution headwinds" for a first start on the road.
Kimi (no engine) filed him under its "healthy QB1 vs backup behind a depleted line" pattern, noting "McCarthy is no wily veteran."

Only two looked past the label:

GPT (no engine) noted he "sparked a 17-point rally against New Orleans."
Kimi (no engine) noted he "was competent in relief (243 yds)."

That relief game was real evidence. In week 6 McCarthy had +5.0 passing EPA in relief, against Jameis Winston's −12.0 in the same game.

This is the same mechanism as Philadelphia, in the other direction. Our data has nothing on McCarthy. It doesn't know who's at quarterback, and a new starter has one relief appearance at most. So the agents fell back on a category ("backup QB") instead of the player. For McKee the category fit. McCarthy was a former first-round pick and Minnesota's starter last season, and it didn't. The agents (0.67) were still less wrong than the market (0.80): Kalshi's "Houston is better than 2-4" was the bigger miss.

Do AI agents with a statistical engine make better NFL picks?

Yes, as far as five weeks can tell. It's small, and it has been steady since the agents stopped seeing betting lines:

Week	Agents	Pairs	Engine minus no-engine (Brier, lower = engine better)	Engine won the pair
1	see lines	64	+0.002	—
2	see lines	64	+0.001	—
3	see lines	64	−0.000	—
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%

When the agents could see the market, the engine added nothing: everyone just stayed near the line. Since they've been blind, the engine side has been about 0.01 better per forecast, about the same for all four models. Week 7 is typical, at the high end.

The pattern inside week 7 is the interesting part. The engine's two biggest wins were the two "data over vibes" games, Seattle (−0.128) and the Rams (−0.078). Two of its four losing games were Philadelphia and Houston (+0.073 each), the QB-change games, where the data had nothing to say.