NFL

Analyst · Week 10

The stats model is the foundation. This layer reads the week's news and forecasts, reports only what the model is not already counting, and moves the number by a capped amount: every claim linked to its source, every claim graded after the final. · predictions are published to show the model works. nothing here is betting advice.Last updated: predictions Oct 1, 11:02 AM ET · results graded Sep 28, 11:22 PM ET

W1 W2 W3 W4 W5 mode: api · 6,185 news items · latest Oct 1, 10:53 AM ET · cap ±10 pts per game

Is the news layer helping?

Since the layer went live, over 93 graded claims:

The adjusted forecast has been worse than the stats model alone: net change in squared error (Brier) of +0.334, where negative is better.
42 of 93 claims pushed the number toward the actual result.
It changed the final pick in 1 game, and those flips went 1–0.

This is the whole test. News can't be backtested the way the base model was (2010–25), so the layer earns its place only forward, and only if it lowers the error over many games, not on any one result. Its job is to move the number toward the truth, not to rescue a pick: when it cut a heavy home pick by 10 points and the team still lost, the move was correct (it lowered the error) even though the pick was wrong: the base model had simply started too far above the market. Picks where we disagree with the market's final price are our weaker spot (about 44% in the backtest, versus 68% when we agree): they're flagged “we differ” on the picks table so you can watch them.

What we do with the news, step by step

This page is the news layer's public notebook. In plain terms:

Before kickoff, every run pulls the feeds and attaches new items to the games they concern. Items published after a game has started are never read: in-game news is not a prediction input.
The reader turns new items into claims: a type (QB change, key starter out, resting starters…), a team, a player, a direction, and a confidence.
The filters decide, in the same run, whether a claim is applied or declined. Declined if it is below the confidence floor for its type, if the model already counts it (the player is on the injury report, the QB is already the named starter), or if it is the kind of story that has no place (offseason hires, turf, "limited in practice").
Applied claims move the probability right then, by a fixed number of points for the type, scaled by confidence, capped. That adjusted number is the pick. "Model 74% → analyst 76%" on a game below means exactly that.
After the final, every claim is graded, applied or declined: did it point the right way, and did it lower the error? Declined claims are graded as if they had been applied, so we learn whether the filter was right to say no.
Once a month, with enough graded cases, the rules change for future weeks: a type that keeps hurting loses weight; a rejection reason that keeps "would have helped" gets loosened. Past picks are never touched. Every rule change is dated in the changelog.

A claim only affects the game it is attached to. It does not carry into next week's model: a starter who misses Week 2 is caught by the injury report by then, and the model counts him there.

How it works

Sources: ProFootballTalk, CBS and ESPN feeds; the National Weather Service forecast at kickoff for outdoor stadiums; the nflverse injury report and named starters are what the model already counts. The reader (one language-model call per game, only when there is new news) returns typed claims about things not already counted. A fixed table turns each claim into points, scaled by the reader's confidence, capped per type and ±10 in total. Weather is logged at zero weight until a season of grading says otherwise.

Claim type	Max points
starting QB change not yet in the feed	6.0
key starter out, not on the injury report	2.0
key starter returning	1.5
resting starters / playing backups	8.0
suspension	2.0
coaching change	2.0
motivation / eliminated / nothing to play for	1.5
locker-room or organizational turmoil	1.0
unusual travel or schedule disruption	1.0
weather at kickoff (logged, zero weight)	0.0
other (in-week, concrete, high confidence only)	0.5

Scorecard by claim type

After each final: did the move go the right way, and did it lower the squared error (Brier)? Negative Brier change is good. Types that don't earn their weight get zeroed at season's end.

Type	Claims	Graded	Avg pts	Helped	Brier Δ
key starter out, not on the injury report	96	58	1.9	29/58	+0.021
other (in-week, concrete, high confidence only)	21	12	0.4	5/12	+0.008
starting QB change not yet in the feed	12	9	5.4	2/9	+0.294
motivation / eliminated / nothing to play for	10	7	1.1	4/7	+0.007
unusual travel or schedule disruption	3	3	0.8	1/3	+0.005
weather at kickoff (logged, zero weight)	2	2	0.0	0/2	+0.000
locker-room or organizational turmoil	1	1	0.8	0/1	+0.010
coaching change	1	1	1.5	1/1	-0.010

Shadow scorecard · what we declined

Rejected claims are kept at zero weight and graded as if applied. If a rejection reason keeps "helping", the rule is too strict; if "other" keeps hurting, it stays out. This is how the layer learns.

Rejected because	Type	Claims	Graded	Would have helped	Brier Δ if applied
no ruled-out language (practice status is not 'out')	key_player_out	111	81	38/81	+0.043
retired: now on the injury report	key_player_out	61	59	31/59	-0.015
confidence 0.75 below floor for other	other	33	27	10/27	+0.018
no ruled-out language (practice status is not 'out')	qb_change	30	22	7/22	+0.470
return of a player the model was not counting out	key_player_return	25	18	11/18	-0.032
confidence 0.65 below floor for other	other	14	13	7/13	-0.002
confidence 0.70 below floor for other	other	8	4	2/4	+0.001
confidence 0.60 below floor for other	other	5	3	2/3	-0.002
confidence 0.45 below floor for key_player_out	key_player_out	4	2	2/2	-0.009
confidence 0.65 below floor for motivation	motivation	4	4	2/4	-0.007
confidence 0.75 below floor for turmoil	turmoil	4	4	3/4	-0.005
no ruled-out language (practice status is not 'out')	suspension	4	3	1/3	+0.033
offseason / logistics story, already priced in	motivation	4	4	0/4	+0.047
soft status (limited/questionable/cleared)	key_player_return	4	3	2/3	-0.002
confidence 0.35 below floor for key_player_out	key_player_out	3	2	0/2	+0.012
confidence 0.70 below floor for turmoil	turmoil	3	3	2/3	+0.001
not an in-week coaching change	coaching_change	3	3	1/3	+0.011
offseason / logistics story, already priced in	other	3	3	0/3	+0.010
already on the injury report	key_player_out	2	2	1/2	+0.005
confidence 0.30 below floor for key_player_out	key_player_out	2	2	1/2	-0.001
confidence 0.40 below floor for key_player_out	key_player_out	2	2	0/2	+0.013
retired: schedule feed now names this QB	qb_change	2	2	0/2	+0.104
QB is already the named starter	qb_change	1	1	0/1	+0.063
confidence 0.00 below floor for coaching_change	coaching_change	1	1	0/1	+0.000
confidence 0.00 below floor for qb_change	qb_change	1	0	–	–
confidence 0.30 below floor for other	other	1	1	0/1	+0.002
confidence 0.50 below floor for other	other	1	1	1/1	-0.002
confidence 0.55 below floor for other	other	1	0	–	–
confidence 0.60 below floor for coaching_change	coaching_change	1	1	1/1	-0.007
confidence 0.60 below floor for motivation	motivation	1	1	0/1	+0.006
confidence 0.60 below floor for turmoil	turmoil	1	0	–	–
confidence 0.65 below floor for coaching_change	coaching_change	1	0	–	–
confidence 0.65 below floor for travel	travel	1	1	1/1	-0.005
offseason / logistics story, already priced in	coaching_change	1	1	1/1	-0.010