Who wins Lions vs Bills according to the simulations?

It depends which engine you ask, and the gap between them is the story. Madden 27's team ratings make it a coin flip — Buffalo wins 51.0% of 10,000 simulated games and the average margin is Bills by 0.4. Our own power ratings make it Bills by 3.9, winning 60.3% of another 10,000. Kalshi's moneyline has Buffalo at 68.5¢, which implies Bills by about 5.4. Both engines sit on the Detroit side of the market, which is the only thing they agree on.

Why does Madden like Detroit more than the market does?

Because EA rates the Detroit roster higher. Detroit's 2026 Madden team composite is 84.3 overall, a z-score of +1.50 — the ratings see Jahmyr Gibbs at 98, an 88-rated quarterback and an 83.8 secondary. Buffalo grades 83.0, z +0.92, carried by a 99-rated Josh Allen. On EA's talent view alone Detroit is the better team and home field is the only reason Buffalo is favored at all. Madden knows nothing about injuries, coaching, weather or who has been playing well, which is exactly why it is worth running as a separate engine rather than folding into ours.

What do the simulations say about the over 54.5?

Both engines land well under it — Madden's scoring environment centers at 46.1 combined points and our ratings-implied read is 48.8, against a Kalshi ladder whose 19 rungs all invert to a centre near 55.0. But we do not publish a totals position on any game, and we are not publishing one here. Our totals model was tested against the market and does not beat it, so it was retired from publication on September 16, 2026. The fair total on this game is the market's number. The simulated distributions in this article are analysis — they show you the scoring environment each engine assumes, and nothing more.

What is the model's read on Jahmyr Gibbs?

Split, and interestingly so. On rushing yardage the market is more generous than we are about his floor — Kalshi has 80+ rushing yards at 60.5¢ where our 10,000 games hit it 45.5% of the time — but our distribution has a much fatter ceiling, putting 160+ at 17.6% against a 7.5¢ market. On receiving yardage we are above the market at every rung from 40 yards up, by 12 to 15 points. The simulation's read is that Gibbs is priced as a volume runner and underpriced as a pass-catcher and as a ceiling game.

How is this different from the model's published edge board?

The edge board publishes positions; this article publishes distributions. Since September 14 our game moneylines are priced off a market-anchored margin with the model weight set to zero, which means the published fair value on a game is the market's own number and the model's read is published separately as a lean in points, never as a recommendation. On this game the model's lean is 1.5 points toward Detroit and Madden's is 5.0 points, and neither becomes a call. That rule exists because Week 1 published ten game positions and lost all ten.

What did the simulation refuse to publish?

Three things. A totals position, for the reason above. The game moneyline as a recommendation, because the anchored fair value is the market's and the model's disagreement publishes as a lean. And two Keon Coleman reception contracts where our number sat more than 30 points from the market — a gap that large on a listed market is treated as a bug until proven otherwise, so it is refused and flagged rather than quietly dropped into a table.

What do these simulations not know?

A great deal. Neither engine has a drive model, so it produces score bands rather than a most-likely final scoreline. Margin and total are drawn independently because nothing we have fitted measures the relationship between them. There is no weather leg, no rest or travel beyond home field, no in-game script, and no accounting for Buffalo opening a new stadium on a short week. Madden in particular has no opinion about who touches the ball, which is why the prop section runs on our projections only.

Are these prices live?

No. Every Kalshi price here is a two-sided mid snapshotted Thursday, September 17, 2026, around 9:10 a.m. ET, roughly eleven hours before kickoff. Prediction-market prices move continuously and this board is liquid — the moneyline alone has traded more than 2.2 million contracts. Check the live ladder before doing anything with these numbers.

Lions vs Bills Simulated 10,000 Times: Madden 27 vs Our Model vs Kalshi

Two football engines played this game ten thousand times each.

One of them is a video game. Madden 27's team ratings, run through a curve we fitted on 3,903 real NFL games, know nothing about injuries, coaching, weather, or the fact that Buffalo opens a new stadium tonight. The other is our own power ratings — the same DAEPA-derived board behind every matchup page on this site.

They disagree with each other by three and a half points. They agree on exactly one thing: Detroit is better than the number.

Madden 27 makes Lions–Bills a coin flip — Buffalo by 0.4 points, winning 51.0% of 10,000 simulated games — because EA rates the Detroit roster higher (84.3 to 83.0) and home field is the only thing keeping Buffalo in front. Our own power ratings make it Buffalo by 3.9 and 60.3%. Kalshi's moneyline sits at 68.5¢, implying Buffalo by 5.4. Both engines land on the Detroit side, but neither becomes a position: since September 14 the published fair value on a game moneyline is the market's own number, so a model disagreement publishes as a lean in points. On totals we publish nothing at all — that model was retired on September 16 for failing to beat the market, and the distributions below are analysis. The live disagreement worth reading is in the props, where Gibbs' receiving ladder runs 12 to 15 points above the board.

The Two Engines

They are not two versions of the same thing. They disagree about different inputs, which is the only reason running both is worth the electricity.

	Madden 27 harness	PMP model harness
Sees	EA's team composites — offense, line, skill, front seven, secondary, special teams	DAEPA-derived power ratings (Off + Def + ST, points scale)
Detroit	84.3 overall, z +1.50	PWR +1.94 (7th)
Buffalo	83.0 overall, z +0.92	PWR +3.36 (4th)
Home field	+2.01 pts (fitted with the curve)	+2.50 pts
Margin centre	Buffalo by 0.4	Buffalo by 3.9
Margin σ	13.96	14.40
Knows about injuries	No	Through the weekly projection file only
Knows about play-by-play	No	Yes
Knows about the market	No	No

The Madden curve is not a toy. Its coefficient (2.82 points of margin per unit of rating z), its home-field term and its total spread come from OLS on 3,903 games from 2010 to 2024. Its variance is decomposed rather than pooled: a team's true strength is drawn once, then game noise is drawn on top, because a 12-win roster is not equally likely to stumble every week. Skipping that step inflates single-game spread by about 14%.

And the two priors are closer in quality than anyone wants them to be — we measured that head to head. Measured against Massey ratings solved from actual results across 480 team-seasons, a Madden composite misses a team's true season strength by 4.31 points. Our own preseason board misses by 4.48. On that test they are not separable. The video game is not the underdog here.

Where The Margin Landed

Ten thousand games each, same seed, same scoring construction.

	Madden 27	PMP model	Kalshi
Average margin	BUF +0.4	BUF +3.7	BUF +5.4
Median	BUF +0.4	BUF +3.6	—
5th percentile	DET by 22.5	DET by 20.1	—
95th percentile	BUF by 23.6	BUF by 27.5	—
Buffalo wins	51.0%	60.3%	68.5%
Average score	BUF 23.3, DET 22.9	BUF 26.3, DET 22.6	—

The market's implied margin comes from inverting its own moneyline mid at the 11.3-point sigma our board uses for moneylines — 68.5¢ is Buffalo by 5.4. That contract has traded more than 2.2 million contracts across both sides, so it is not a thin quote we are arguing with.

Against the spread ladder, both engines are on Detroit at every rung the market has posted:

Buffalo covers	Kalshi	Madden	Model
−1.5	66.0¢	46.7%	56.0%
−2.5	63.5¢	44.0%	53.0%
−3.5	55.5¢	41.3%	50.2%
−4.5 (deepest rung)	51.5¢	38.6%	47.3%
−5.5	49.5¢	35.7%	44.7%
−7.5	38.5¢	31.0%	39.5%
−10.5	32.5¢	24.0%	32.0%

Notice where the model and the market converge: from about seven points out, our curve and Kalshi's ladder are on top of each other. The disagreement is entirely about the middle of the distribution — whether this is a field-goal game or a touchdown game — and not about the tails.

And here is the part that is not a recommendation

Our model likes Detroit here. It is still not a call, and the reason is written in results, not taste.

Since September 14 the game moneyline is priced off a market-anchored margin with the model's weight set to zero. In plain terms: the published fair value on a game is the market's own number, and our disagreement is published separately as a lean in points. Madden's lean is 5.0 points toward Detroit. Our model's is 1.5 points. Neither becomes a position.

That rule cost us something to learn. Week 1 published ten game moneyline positions and lost all ten, every one a dog. The board now refuses an anchored row past 7 points of probability and withholds the recommendation entirely on an unanchored read — the Lions–Bills matchup page shows the number and says nothing about what to do with it. This article does the same.

The Total — Read This Before You Use It

Both engines produce fewer points than the market. Neither number is a position, and we are not going to dress one up as one.

	Centre	Over 48.5	Over 54.5	Over 60.5
Madden harness	46.1	43.1%	26.2%	13.4%
PMP model	48.8	50.9%	33.7%	19.1%
Kalshi	~55.0	68.5¢	51.5¢	34.5¢

Two honesty notes, both of which matter more than the numbers:

Madden does not have a totals opinion. EA's ratings carry no scoring-environment signal we have ever fitted, so that harness draws the league's distribution — 2025's realized combined scoring, 46.03 points — and lets only the margin come from the game. The "46.1" above is the NFL's number, not Madden's. Anywhere you see it quoted, it should carry that sentence.

Our totals model was retired from publication on September 16. Walk-forward against the market it carries no information the closing line does not already have, so pricing totals off it manufactures edges out of noise. The switch is off in the code. The only totals signal we price at all is an outdoor-wind adjustment to the market's centre, and tonight's forecast does not clear the threshold that turns on.

So what is the 48.8 doing in a table? Showing you the scoring environment the margin distribution was built on. That is all it is for. Coach Gene took the over 54.5 at 45¢ on Monday — that piece prices his exact words across the ladder, and it carries the same rule.

One thing worth admiring on the other side: Kalshi's total ladder on this game is internally immaculate. Invert all nineteen rungs individually — from over 30.5 at 95.5¢ to over 72.5 at 10.5¢ — and they imply centres of 54.9 on average with a standard deviation of 0.54 points. There is no shape to trade against. That is a well-made book, and it is a useful thing to be able to say about a market you disagree with.

Score Bands, Not Scorelines

There is no drive model in either harness, so neither one knows that football scores come in threes and sevens. A most-likely-exact-scoreline out of a Gaussian would be a fabrication. Bands are what the simulation can honestly give you:

Points scored	0–13	14–20	21–27	28–34	35+
Buffalo (Madden)	15.7%	23.5%	27.9%	20.7%	12.2%
Buffalo (model)	9.6%	18.3%	27.3%	24.3%	20.5%
Detroit (Madden)	16.7%	23.7%	28.5%	19.8%	11.4%
Detroit (model)	17.8%	24.3%	26.9%	19.7%	11.3%

Detroit's band profile is nearly identical across both engines. The whole difference between the two harnesses lives on Buffalo's side of the ledger — our model gives the Bills a 20.5% chance of a 35-point night where Madden gives them 12.2%.

The Props

Here the two engines stop being two engines.

Madden's composites are team aggregates. EA has no opinion about how many carries a running back gets, and inventing one by scaling our projections against a points ratio would be a number with no measurement behind it. So the prop table below is our model only: 10,000 draws per player from the same fitted distribution our prop board prices in closed form — a zero-inflated lognormal around each projection for yardage and receptions, Poisson for touchdown counts, with the dispersion read per projection level from a fit on four seasons.

We checked the sim against that closed form on all 221 priced strikes. Maximum disagreement: 0.98 points of probability, against a Monte Carlo standard error of 0.50 at a coin flip. The simulation and the shipped board are the same model.

Jahmyr Gibbs

The most interesting player on the board, and the reason the prop section leads with him.

Contract	Kalshi	Model	Gap
50+ rush yds	79.5¢	67.3%	−12.2
80+ rush yds	60.5¢	45.5%	−15.1
100+ rush yds	37.5¢	35.5%	−2.0
140+ rush yds	13.5¢	22.0%	+8.5
160+ rush yds	7.5¢	17.6%	+10.1
180+ rush yds	3.5¢	14.3%	+10.8
40+ rec yds	32.0¢	45.5%	+13.5
50+ rec yds	20.5¢	35.0%	+14.5
60+ rec yds	12.0¢	27.1%	+15.1
70+ rec yds	7.0¢	20.8%	+13.8

Two separate reads in one player. On the ground, the market is paying up for the floor — Gibbs clearing 50, 60, 80 yards — and our distribution is less sure of it, while being far more willing to pay for the ceiling game from 140 yards up. Through the air, we are above the board at every rung from 40 yards, by 12 to 15 points. The simulation is pricing Gibbs as a pass-catcher the ladder has not fully accounted for, on a night when the model has Detroit trailing in most of the games it plays.

The rest of the key names

Widest disagreements on the players anyone is actually looking at, every price a two-sided mid:

Player	Contract	Kalshi	Model	Gap
DJ Moore	50+ rec yds	63.0¢	34.4%	−28.6
DJ Moore	4+ receptions	66.0¢	43.5%	−22.5
Amon-Ra St. Brown	60+ rec yds	69.0¢	44.8%	−24.2
Amon-Ra St. Brown	6+ receptions	73.0¢	51.3%	−21.6
Josh Allen	25+ rush yds	62.5¢	43.7%	−18.8
Jared Goff	225+ pass yds	70.5¢	50.4%	−20.1
Jameson Williams	50+ rec yds	60.5¢	43.5%	−17.0
Khalil Shakir	6+ receptions	26.5¢	38.4%	+11.9
Khalil Shakir	7+ receptions	16.0¢	27.9%	+11.9
Josh Allen	325+ pass yds	14.5¢	26.2%	+11.7
Josh Allen	350+ pass yds	8.0¢	19.5%	+11.5

Shakir is the cleanest single read on the board: his receptions ladder sits above the market at every rung from five up, and his receiving-yardage ladder is the one player on this game where our number and Kalshi's agree almost exactly at the bottom and diverge only in the tail. That is the shape of a market that has the right median and a thin ceiling.

The pattern underneath all of it

Sort all 221 priced strikes by how far the strike sits from the projection and something systematic falls out:

Strike ÷ projection	Strikes	Model minus market
Under 0.7	31	+0.8
0.7 – 1.0	33	−9.2
1.0 – 1.5	54	−2.4
1.5 – 2.5	79	+2.3
2.5 and up	16	0.0

Across the whole board the median gap is +0.6 points — the two are well calibrated on average. But the shape differs in a consistent direction: our distribution is flatter than the market's. Just below a player's projection, where the most liquid contracts sit, we are nine points under the board. Out in the ceiling, we are above it.

That is not automatically an edge. It is equally consistent with our fitted dispersion being too generous in the tails, which is a live question rather than a settled one — the fit is four seasons deep and beats a pooled alternative walk-forward, but it has never been graded against this market specifically. Read the individual rungs above knowing that a systematic component sits underneath them.

The Three Things We Refused To Publish

A model is only as good as the numbers it declines to print.

A totals position. Both engines are six to nine points under the market. The totals model does not beat the market and was retired from publication on September 16. The fair total on this game is Kalshi's.
The moneyline as a recommendation. The anchored fair value on a game is the market's own price. Our lean is 1.5 points toward Detroit, Madden's is 5.0, and both publish as reads.
Two Keon Coleman reception contracts. Our simulation put them 35.2 and 32.6 points above the market. A gap that large on a listed market is a bug until proven otherwise, so they are refused and flagged rather than dropped silently into a table where they would have been the two biggest "edges" on the page.

What The Simulation Does Not Know

Both harnesses, stated plainly, so nobody has to guess:

No drive model. Scores are bands, not scorelines.
Margin and total are drawn independently. Real games have a weak relationship between how lopsided they are and how many points they hold. Nothing we have fitted measures it, so nothing here pretends to.
No weather, rest or travel beyond the home-field term.
No game script. A team trailing late throws more. Neither engine adjusts anyone's volume for the game it is simulating.
No new stadium, no short week, no Thursday. Buffalo opens Highmark tonight on four days' rest. That is real and it is not in either model.
Madden knows nothing about 2026. The composites are launch ratings. Two weeks of football have happened since.

The distribution is what the engine can tell you. The rest is why it might be wrong.

Game: Detroit at Buffalo · Thursday, September 17, 8:15 p.m. ET · Highmark Stadium, Orchard Park, N.Y. · matchup page

Board: Kalshi KXNFLGAME-26SEP17DETBUF, two-sided mids snapshotted 9:10 a.m. ET, September 17, 2026.

Method: 10,000 simulations per harness, seed 20260917. Sim probabilities verified against the shipped closed-form pricer on all 221 priced strikes, maximum disagreement 0.98 points.

The live board moves. Check the ladder before you do anything with a number on this page.

We Ran Lions-Bills 10,000 Times Through Madden 27 — Then 10,000 More Through Our Own Model